construction technology | Construction Metrics

Mach9 adds drone and aerial LiDAR to its AI survey drafting, and Langan says it cut delivery time about 40%

Fri, 06 Nov 2026 06:34:15 -0500

Mach9, a San Francisco company that sells web-based AI CAD software for civil survey work, said on November 4, 2026, that its platform now takes aerial LiDAR and photogrammetry data from drones, helicopters and planes. Until now it worked from mobile and terrestrial scans. The release affects the step between a site flight and a usable topo: classifying the point cloud, drafting the linework and building the ground surface that goes into Civil 3D or OpenRoads.

What it does

On a site job today, a drone pilot can fly the parcel in a morning. Then a survey technician spends hours in the office classifying points, tracing curbs and pavement edges, and building a surface in several different programs. Mach9’s release says one day of flying on a detailed electric utility project can turn into months of office work.

Mach9 puts those steps in one browser application. The survey team uploads LAS or LAZ point clouds from LiDAR or photogrammetry, along with an orthomosaic, which is a stitched aerial photo corrected to map scale. Mach9 says its AI produces a first draft in hours. The draft includes 3D points and lines for curbs, building footprints, pavement edges and vegetation outlines, plus a classified point cloud, meaning each point is tagged as ground, building, vegetation or another class.

The surveyor then checks and edits the draft in the same tool. Surface tools build ground surfaces, breaklines and contours, and a deviation heatmap shows where the surface departs from the points. Exports include DXF, DWG, SHP, GDB and LandXML. For utility work, the release says the software also drafts utility poles and the curved 3D lines of overhead wires for clearance analysis and make-ready design.

On a contractor’s side, the users are an in-house survey or VDC group on a sitework, heavy civil or utility job. The output feeds existing-conditions surfaces for earthwork quantities on land development, solar and data center sites, along with site and ALTA surveys and corridor basemaps.

Mach9’s own guidance on drone data sets limits. Its density guide says a flight needs at least 100 points per square meter for automated linework, and it recommends 400 or more. The same guide lists vegetation edges and overhead wires as beta on drone data. It says signs, poles and manholes are extracted from mobile mapping data but still need an operator on drone data.

Drone LiDAR density and what Mach9 says its automation drafts
Drone LiDAR density (points per sq m)	Approximate point spacing	What Mach9's guide says to expect
Below 100	More than 10 cm (4 in)	Ground classification still runs; most linework is drafted by hand
100	About 10 cm (4 in)	Minimum for automated linework; a 6-inch curb gets one or two points
400 or more	About 5 cm (2 in)	Recommended; a curb gets about three points
About 650 and up in a cross-grid	Not stated	Separate curb lines plus crosshatch and guardrail came through in Mach9's drone projects

Source: mach9.ai

Evidence so far

The evidence so far is Mach9’s release and blog posts, with comments from survey managers at three engineering firms that use the software. AEC Magazine’s coverage repeats the same customer figures.

These claims come from Mach9 and its customers:

Russ Hall, senior survey project manager at Langan, said Mach9 cut the firm’s delivery time by weeks, about 40%, on several multi-thousand-acre projects in Texas that it could not otherwise have finished on time.
Thomas Tucker, survey project manager at Welch Comer, said he saves an average of four to five hours per project. “[Every] two projects, we’re saving a day,” he said.
Welch Comer flies a GeoCue TrueView 540 at 150 to 200 feet above ground and about 5 meters per second. Mach9 calculates from GeoCue’s published figures that those settings give roughly 900 to 1,200 points per square meter in one pass.
Mach9 says it has been used on more than 2,500 projects in all 50 states and that teams work up to 10 times faster at one-fifth the cost.

All three quoted firms are engineering and survey consultants. Mach9 describes its accuracy as centimeter-level at infrastructure scale. A breakdown of Langan’s 40% by step and an accuracy report against surveyed checkpoints are not yet public.

How to measure it on your projects

Track office hours per acre of finished topo. Start the clock when the flight data reaches the office and stop it when the LandXML or DWG goes to the estimator or the engineer. Count classification, drafting, surface building and QA. Log any return trip to the field for missed features as a separate count.

Capture the baseline before the pilot:

Pull three recent drone topos of similar site type, with acreage and the flight’s point density.
Get the office hours on each from timesheets or from the technician.
Note how many features were added after the first delivery, and whether anyone went back to the site.

On the pilot, log the same figures, plus the hours spent fixing the AI draft. Check the finished surface against your own control shots, and record the vertical difference at each check point. That gives you an accuracy figure for your own sites.

A first step on one job is to check the point density of a flight you have already processed. If it falls below 100 points per square meter, Mach9’s guide says most linework will still be drawn by hand, so plan the next flight lower and slower before the pilot.

Availability and cost

The aerial support is available now. Mach9 prices by quote through a demo request. The company is holding a webinar on the aerial workflow on November 17, 2026.

Shackleton AI launches agents that draft RFIs and chase overdue submittals, with $1 billion of projects on its platform

Thu, 05 Nov 2026 06:34:37 -0500

Shackleton AI, a New York software company, came out of stealth on November 4, 2026, with a project management platform for general contractors and owners that runs AI agents on the paperwork. The company says the agents draft RFIs, follow up on overdue submittals, flag budget variances before a payment request goes for approval, and turn field notes into daily reports. It also announced a $5 million seed round led by Battery Ventures. Shackleton says the platform launched in June 2026 and now manages more than $1 billion in project value across more than 15 states.

What it does

On most jobs today, a project engineer logs RFIs and submittals in one system, the superintendent writes the daily report in another, and the project manager checks the budget in a spreadsheet or the accounting package. Shackleton’s announcement says a company commonly runs 7 to 10 such systems, and that teams spend 10 to 15 hours a week on data entry and administrative work.

Shackleton puts those records in one system. Its website lists projects, bids, budgets, RFIs, submittals, change orders and field reports in a single data model. An AI agent here is software that reads those records and prepares the next step, such as a draft RFI or a reminder on an overdue submittal, for a person to approve.

The work moves through the system in the order a job runs:

Field staff mark up sheets and record walkthroughs on the mobile app. Sheets stay readable and can be annotated offline, then sync when the crew has signal again.
A field agent turns field notes into daily reports. The homepage demo shows it turning a floor walk’s photos into punch items.
An office agent drafts RFIs and follows up on submittals that are past due.
A finance agent flags budget variances before the payment request is submitted for approval.
The project manager reviews the agents’ work. The release says review and approval workflows keep people in charge of it.

The website names seven agents in all: field, growth, planning, development, procurement, legal, and office and finance. The browser side holds the Gantt schedule and look-aheads.

Shackleton sells to midsize general contractors in multifamily, industrial and commercial work, and to owners that run rollouts across many locations. The release names Red Lobster, Dog Haus and Pepper Lunch as customers.

Evidence so far

The evidence is Shackleton’s launch announcement and its website. The three founders previously grew Ernest, a commercial general contractor in the Southeast, to $35 million in revenue. “I was spending more time moving data between systems than running my jobs at Ernest,” CEO Michael Ragheb said in the release.

These claims come from Shackleton:

Ragheb said multi-location customers have finished buildouts up to 35% ahead of schedule and as much as 18% under budget.
The website credits the growth agent with bid turnaround three times faster, and the development agent with a 29% shorter time from concept to construction start.

The 35% and 18% figures describe the best results, as the words “up to” and “as much as” indicate. The number of projects behind them, the baseline they were measured against, and any RFI or submittal turnaround times are not yet public.

How to measure it on your projects

Track RFI turnaround in calendar days, from the date a field question is first raised to the date the answer reaches the crew. Shackleton’s agents draft the RFI and chase the reply, so log three dates for each one: when the question came up, when the RFI went to the architect, and when the answer was distributed.

Metrics to capture before piloting an agent on RFIs, submittals and daily reports
Workflow	Metric	Definition	Baseline to capture
RFIs	RFI turnaround (calendar days)	Date the question is raised in the field to the date the answer reaches the crew	Median and slowest 10% from two or three recent jobs
RFIs	Draft acceptance rate	Share of agent-drafted RFIs sent with no edits	None before the pilot; count from the first week
Submittals	Overdue submittals	Count of submittals past their required-by date on the log each Monday	Weekly count from a current job
Daily reports	Same-day filing rate	Share of daily reports filed before midnight on the day of the work	Last 30 working days on one job
Pay applications	Variances caught before approval	Budget variances found before the pay request goes out for approval	Count from the last three pay apps

Capture a baseline from the RFI and submittal logs of two or three recent jobs of similar size and delivery method. Our post on measuring RFI turnaround before an AI pilot covers setting up that baseline. During the pilot, also count how many agent drafts the project engineer sent unchanged, edited, or discarded, since edited drafts still cost review time.

For daily reports, record the time between the end of the shift and the report being filed, and the share of reports filed the same day. Compare it with the voice-first approach in our post on Punch’s daily reports for superintendents.

A first step on one job: export the RFI log for your most recent closed project and add a column for the date each question was first raised in the field.

Availability and cost

Shackleton takes demo requests at shackleton.ai/demo. Prices are not listed on its website. The agents work on records stored in Shackleton’s own data model, so a contractor would move its RFI, submittal and daily report logs onto the platform to use them. The company says it will spend the seed money on more agent workflows and on sales.

AI cross-check of structural calculations against rebar schedules found 253 errors, 29 of them critical, on a 19-story Seoul apartment building

Wed, 04 Nov 2026 06:34:44 -0500

A research team led by Jaewook Jeong, an associate professor at Seoul National University of Science and Technology, has published an AI pipeline that reads a building’s structural calculation package and its structural drawings, matches every beam and column across the two, and flags where they disagree. The paper, by Minsang Gu, Jaewook Jeong and Louis Kumi, went online in Engineering Applications of Artificial Intelligence on October 13, 2026, and the university announced it on November 3. On a 19-story reinforced concrete apartment building in Seoul, the system found 253 discrepancies, and the authors class 29 of them as critical to structural stability. The workflow it targets is structural document review, the check of rebar sizes, bar counts, stirrup spacing and member dimensions between what the engineer calculated and what the drawings tell the field to build.

What it does

Two documents go in. The calculation package comes out of analysis and design software (MIDAS ADS and MIDAS SDS on the case project) and lists each member’s loads, reinforcement and layout. The structural drawings are the schedule-type sheets a drafter builds from that package in CAD. The paper’s premise is that this hand transfer between two formats, often by different people, is where omissions and typos get in.

The pipeline then runs in three stages:

It collects the documents and drops low-quality pages.
It pulls member data out of each document. On the calculation pages, a layout model called DocLayNet-YOLO finds the regions that hold member information, and Surya-OCR, an optical character recognition model that turns a page image into text, reads them. On the drawings, the software finds the member codes first, then uses OCR with edge and line detection to locate the schedule tables, whose rows and columns vary in width.
It matches each member across the two documents by member code and direction, compares reinforcement and dimensions, and sorts every disagreement by type.

Results come out in a web application the team built. The method works from the documents alone, with no BIM model, and the authors aim it at small and mid-size projects that do not build one. They name structural engineers, design consultants, contractors and construction supervision teams as the people who would use it. The paper also notes that US practice splits structural documents among architects, structural engineers and CAD technicians, which is the kind of divided authorship the check is built for.

Evidence so far

The evidence is one case study the authors ran on the Seoul project, reported in the paper’s abstract and the preview ScienceDirect shows of the full text. The authors declare no competing interests, and a Korean government grant for commercializing research outcomes funded the work.

Extraction accuracy, the share of fields the software read correctly, was 95.56% on the calculation package and 90.47% on the drawings. The 253 discrepancies were all in beams and columns.

Discrepancies between the calculation package and the structural drawings, by type
Discrepancy type	Count
Missing entries	224
Main reinforcement mismatches	15
Dimension mismatches	8
Stirrup mismatches	3
Duplicate entries	3
Total	253

Source: techxplore.com

So the 29 critical errors are the 253 minus the 224 missing entries. Every mismatch in bar size, bar count, stirrup spacing or member height, plus the three duplicates, landed in the critical group.

The 29 errors the authors class as critical to structural stability
Critical error	Count
Dimension height	8
Bottom rebar diameter	6
Top rebar diameter	3
Top rebar count	3
Bottom rebar count	3
Stirrup spacing	3
Duplication	3
Total	29

Source: sciencedirect.com

The two accuracy figures matter for reading the larger number. If the software misread about one drawing field in ten, some of the 224 missing entries could be fields it failed to read and not gaps in the drawings, and the preview does not show how the authors confirmed each flag. The time the check took, and a manual review of the same documents for comparison, are not in the abstract or the preview. (The paper cites conflicting design documents as a contributing factor in the New World Hotel collapse in Singapore and an underground parking structure collapse in Incheon.)

How to measure it on your projects

The metric to borrow is confirmed discrepancies per structural sheet, counted before rebar detailing starts and sorted into the paper’s five types. Count a flag only after the engineer of record or your detailer confirms it, so OCR misreads do not inflate the total. Dividing confirmed flags by all flags gives you the tool’s precision, the share of its flags that turned out to be real errors.

You need the calculation package for this, and the engineer of record does not always issue it to the contractor. Ask for it at the start of the pilot.

Capture two baselines on a recent concrete job: the RFIs that traced back to a conflict between the calculations and the beam or column schedules, and the hours your detailer or project engineer spent checking schedules by hand. A first step on one job is to take a single floor’s beam schedule and the matching calculation pages, check them by hand, and log the hours and each discrepancy by type. Run the tool on the same floor afterward and compare both counts and both times.

Availability and cost

This is a research system. The paper describes a web application, and the abstract and preview do not mention a public code release, a commercial version or a price. The full text sits behind Elsevier’s paywall.

Earlier posts covered commercial drawing review tools, including Buildcheck, Document Crunch’s drawing review and Bluebeam Max.

Bluebeam survey of 750 US AEC firms finds 28% run AI in production and 17% credit it with less rework

Tue, 03 Nov 2026 06:34:15 -0500

Bluebeam released its 2027 AEC Technology Outlook Report on October 13, 2026. Hanover Research ran the survey online for Bluebeam in July 2026, and the US edition covers 750 decision-makers at architecture, engineering and construction firms. About 28% of those firms say AI is in production, and 7% say it runs across several core workflows. Among the 448 firms using AI in any form, 64% name greater cost predictability as a benefit and 17% name reduced rework. The questions follow AI from estimating and design through the handoff to the field, and then ask what stops firms from scaling it and whom they can hire to run it.

What it does

The survey tests no product. Respondents work full time at AEC firms, at manager level or above, with authority or influence over technology decisions. The 750 US answers come from a wider study of 1,100. Each firm placed itself on a five-point adoption scale.

Where US AEC firms stand on AI, July 2026 survey (n=750)
Adoption stage	Share of US AEC firms (%)
AI embedded across multiple core workflows	7
AI supporting a few specific workflows	21
Piloting AI in limited or experimental use cases	32
Not using AI, but exploring future use	31
Not using AI, and no plans to	10

Source: bluebeam.com

“In production” is Bluebeam’s grouping of the top two rows, 7% and 21%. Add the 32% piloting and 60% of firms have started with AI in some form. (The five rows add to 101% from rounding.)

Most figures below come from one of three groups of respondents:

All 750 firms answered the adoption question.
The 448 firms using AI, pilots included, answered the questions on where they use it, what it delivers and what blocks scaling.
The 210 firms past the pilot stage answered what got them there and what they saved.

Bluebeam says this year’s 28% cannot be compared with the 27% in its 2026 edition, because the question wording and the set of countries both changed. The report page does not split results for contractors apart from architects and engineers. It does note that 58% of respondents are aged 35 to 44 and under 1% are 55 or older.

Evidence so far

The evidence is one survey of self-reported answers, commissioned and analyzed by a software vendor. Bluebeam sells Revu and the AI tier Bluebeam Max. Bluebeam also ran its own analysis of Hanover’s respondent-level file to compare firms that scaled with firms still piloting.

Where AI sits and what it delivers

Among firms using AI, the most common uses are office and preconstruction work.

Most common uses of AI among firms using it
Where firms use AI	Share of firms using AI (%)
Cost estimation	69
Design and modeling	69
Project planning	63
Document management	60

Source: bluebeam.com

Benefits reported from AI, firms using AI (n=448)
Benefit reported from AI	Share of firms using AI (%)
Greater cost predictability	64
Earlier identification of risks and issues	58
Reduced administrative burden	55
Enhanced team coordination knowledge	53
Reduced project delays	47
Improved output quality and consistency	43
Faster workflows and internal processes	38
Easier access to information and insights	27
Stronger compliance/audit readiness	27
Improved safety outcomes	25
Reduced rework/late-stage changes	17
Reduced cognitive load for staff	14

Source: bluebeam.com

Bluebeam groups the top three benefits as ways of knowing about a problem sooner. Reduced rework ranks 11th of 12. Bluebeam’s reading is that AI sits in estimating, design and planning, while rework happens in the field, where the people who could stop it have the least access to the systems that flagged it. The report puts the direct cost of rework near 5% of total construction cost, citing Construction Industry Institute research and a 2009 study by Hwang, Thomas, Haas and Caldas, and it adds that those studies lean toward industrial capital projects. Bluebeam CEO Usman Shuja says in the report that he has watched a model flag a clash three weeks early and seen the same clash get built anyway. “Prediction is a data problem, and we have largely won it,” he says.

Barriers to scaling

Barriers to scaling AI, by adoption stage (Bluebeam analysis of respondent-level data)
Barrier to scaling AI	Still piloting (n=238) (%)	Scaled past pilot (n=210) (%)
High implementation costs	50	50
Lack of internal AI skills	34	34
Resistance to change	27	17
Security and compliance	15	22
Moving beyond pilots	16	10
Identifying right use cases	10	18
Unclear business value	11	2
Data quality issues	9	4
Difficulty accessing data	5	2
Lack of trust in outputs	3	0

Source: bluebeam.com

Cost is named by exactly half of each group, so it does not separate the firms that scaled from the ones that stalled. The press release gives different figures for some rows, 22% for resistance to change and 18% for security, because it pools both groups. The arithmetic checks out:

27% of 238 piloting firms is about 64 firms, and 17% of 210 scaled firms is about 36.
That is 100 of 448 firms, or 22%.

The same weighting turns 15% and 22% on security into 18%, and 3% and 0% on trust into the 2% the release quotes.

Trust shows up twice in the survey. Only 2% call a lack of trust in outputs a barrier, while 20% of scaled firms list trust in AI outputs as something that carried them to production. The barrier question and the success question ask about different things, so a firm can count trust as a help without ever having found it missing.

What the scaled firms credit

What carried firms from pilot to production
Factor that carried AI from pilot to production	Share of firms past the pilot stage (n=210) (%)
Integration with existing systems/tools	47
Cross-team collaboration	29
Transparency of AI results	27
High-quality data	26
Employee training	26
Ease of use for employees	22
Trust in AI outputs/recommendations	20
Ability to tailor solutions to project needs	19
High vendor expertise	17
Clear value/return on investment	12
Strong leadership support	11
Ability to validate or audit AI-generated results	11
Adequate vendor support	10
Accessible data	8

Source: bluebeam.com

Integration with existing systems leads at 47%, and a clear return on investment sits at 12%. Bluebeam itself calls this question the weakest evidence in the study, since firms describe their own success after the fact. The use pattern lines up with it: 41% of firms using AI use it mainly to connect data across tools they already own.

How firms use AI relative to their existing tools (single response)
Use of AI relative to existing tools	Share of firms using AI (n=448) (%)
To connect and coordinate data across existing tools	41
To automate tasks within individual tools	37
To replace multiple tools with a single platform	23

Source: bluebeam.com

Where project workflows break down, US respondents
Where project workflows break down	Share of US AEC firms (%)
Design to construction handoffs	41
Between office teams and field teams	41
Document management	38
Construction to operations handoffs	32
Between internal teams and external partners	32
During reviews, approvals, or sign-offs	30
Between different software tools/platforms	24
Between disciplines	20

Source: bluebeam.com

Among the 210 firms in production, 40% estimate AI savings above $500,000 and another 31% between $100,000 and $500,000. The period and the baseline behind those estimates are not part of the published report.

Self-reported savings from AI, firms in production
Self-reported AI savings	Share of firms running AI in production (n=210) (%)
More than $500,000	40
$100,000 to $500,000	31

Source: bluebeam.com

Hiring and shortages

Capabilities firms prioritize in hiring and find hardest to hire
Capability	Prioritized when hiring (%)	Hardest to find in candidates (%)
Digital tools proficiency	93	77
Project management skills	82	40
Problem-solving and adaptability	74	31
Ability to work across multiple tools and platforms	73	43
AI-related knowledge/experience	69	59
Traditional technical/design expertise	58	33
Data analysis and interpretation	46	23

Source: bluebeam.com

Digital tools proficiency is both the most sought skill, at 93%, and the hardest to find, at 77%. AI knowledge is second-hardest at 59%. Only 16% of firms report no labor or skills shortage, and 82% of all respondents say shortages limit their ability to deliver projects to some degree. (Bluebeam’s hiring chart lists field and trade experience on two rows with different values, so this table leaves that capability out.)

Effects of labor and skills shortages
Effect of labor and skills shortages	Share of firms reporting a shortage (n=631) (%)
Increased workload for existing employees	74
Delayed project timelines	49
Increased project costs	39
Limited the number of projects taken on	36
Increased reliance on subcontractors/external partners	33
Reduced quality/increased rework needed	20
Increased safety risks	15

Source: bluebeam.com

How to measure it on your projects

The rework finding points at one number a contractor can track: the flag-to-fix rate. It is the share of coordination issues flagged before installation that were resolved before the affected work went in. Flags can come from clash detection or from an AI review tool, and the rate counts only whether the fix arrived before the work went in.

Build the baseline on one active job before any pilot:

Export the last three months of the coordination issue log, including clash items and the RFIs they generated.
For each issue, record the date flagged, the date a design response or RFI answer closed it, and the date the affected scope was installed, taken from daily logs.
Count the issues whose scope is already in place, and divide the ones closed before installation by that count.

Track the same rate on jobs where an AI review tool supplies flags, and keep the baseline job’s rate beside it. Pair it with RFI turnaround in calendar days, since a late answer is the usual reason a flagged issue gets built. To start, add an installed-date column to the coordination log on one job this week.

Availability and cost

The full US report is a free page on Bluebeam’s site. This year’s study covered three countries, reported as two separate markets, and this page is the US edition. Construction Briefing covered the US findings on November 2.

Zutec launches Handover AI to sort subcontractor closeout documents and track what is still missing

Mon, 02 Nov 2026 06:34:33 -0500

Zutec, a construction software company that sells mainly to contractors and building owners in the UK and Ireland, launched Zutec Handover AI on November 2, 2026. The announcement says the product is available now and is built for main contractors. It covers closeout: collecting O&M manuals, drawings, certificates and safety files from subcontractors and assembling the record the owner receives at practical completion, the UK term for substantial completion.

What it does

On most jobs this work falls to a project engineer or document controller on the main contractor’s team. Zutec describes the current process as chasing subcontractors for files, receiving wrong or incomplete documents and keeping status in Excel trackers, which ends in a rush at practical completion.

With Handover AI, subcontractors upload documents into Zutec’s platform through guided templates or by drag and drop. According to the announcement, Zutec’s own specialist staff chase missing submissions and quality-check each upload within 48 hours. The software then classifies, names and tags each file. Classification means the AI reads a document and assigns it to a slot in the handover structure, the way a document controller would file a pump data sheet under the mechanical O&M section.

The main contractor’s team reviews, approves or returns each item in a workflow that keeps comments, tasks and an audit trail in one place. Dashboards show what has been submitted, what is outstanding and what is complete, for one project or across a portfolio. A chatbot answers questions about the documents collected so far.

The output is the building manual. Zutec lists exports for UK requirements, including Building Safety Act and Regulation 38 fire safety packages, along with an asset register and a logbook. The Irish version of the release names the Digital Safety File and BCAR documentation instead. After handover, an owner can load the record into Building AI, a separate Zutec product launched in February that answers plain-language questions with a citation to the source document.

Evidence so far

The evidence is Zutec’s announcement, which includes comments from one main contractor and one subcontractor. These claims come from Zutec:

The product draws on more than 25 years of handover work and data from over 5,000 projects.
Zutec’s own research found that main contractors repeatedly cite supply chain chasing, incomplete information, slow responses and limited visibility as handover problems.
Asset register, Regulation 38, Fire and Emergency File, BREEAM and logbook features, which were paid add-ons before, are now part of the core product without raising its cost.

Jet Maal, a senior technical coordinator at main contractor Mount Anvil, says in the release that the dashboard shows a project completion percentage, an activity log and the items waiting for review. Maal adds that consistent AI naming and tagging cuts administrative work. Thalia Rothwell, a quality engineer at subcontractor William Hare, says the subcontractor’s team knows what to submit and where, and spends less time asking for status updates.

Time saved per project, the accuracy of the AI classification and the share of uploads that pass the 48-hour check on first submission are not yet public.

How to measure it on your projects

Track closeout completeness at substantial completion. Divide the number of required closeout documents that were received and accepted by the substantial completion date by the total number the contract requires. Take the list from the closeout submittal requirements in Division 01 of the spec, and count each O&M manual, warranty, as-built set and test report as its own line item.

Capture these baselines from three to five recently closed jobs of similar size:

The completeness percentage on the substantial completion date.
Calendar days from substantial completion to the owner’s acceptance of the full closeout package.
Hours the project engineer or document controller charged to closeout in the job’s last 60 days, if your timekeeping separates them.

During a pilot, log each document the AI classified and whether a reviewer had to rename or refile it. The share that needed a correction is the tool’s error rate on your own subcontractors’ files. Closeout documents are the last submittals on a job, and our post on Buildertrend’s AI Submittals covers a tool that drafts the submittal log at the start of a job.

A first step on one job: pull the closeout log from your most recently finished project and count how many required items were still open on the substantial completion date.

Availability and cost

Zutec says Handover AI is available now. The releases carry London and Dublin datelines, and the regulatory packages they name are UK and Irish ones. Availability for US projects could not be verified. Pricing is not listed in the announcement.

Eastern Construction says Tyce's AI bid review cut its preconstruction time at least 40%

Sun, 01 Nov 2026 06:34:11 -0500

Tyce, a Toronto software company, announced on October 19, 2026, that it raised C$2.8 million (about US$2 million) in pre-seed funding, an early round that comes before a seed round. BDC’s Thrive Venture Fund and the MaRS Investment Accelerator Fund led it. Tyce’s software reads a bid package, flags scope gaps and risky contract terms, and drafts proposals. In the same announcement, Ontario general contractor Eastern Construction says it rolled Tyce out across its preconstruction teams and cut the time spent bidding on and managing projects by at least 40%.

What it does

On a typical pursuit, an estimator reads the drawings, specs and addenda, writes a scope sheet for each trade, and then levels the sub bids against those sheets. Tyce’s CEO and co-founder, Lylia Djait-Paulien, told The Construction Broadsheet that the product is meant to catch the item on page 400 that nobody had time to read.

The inputs are the bid documents. Tyce pulls drawings, specifications, addenda and trade bids from Procore, Autodesk Construction Cloud or a direct upload. It links them into one project model that the team can question in plain language.

The software then flags scope gaps, exclusions and risky contract terms. It supports takeoff and bid leveling, and it drafts proposals. Each flag shows the sheet, spec section or page it came from, so the estimator can open the source and confirm it. Scope sheets, bid leveling and risk summaries export to Excel.

Tyce’s website adds two functions. It compares supplemental clauses with standard contract forms and flags risks against the company’s own risk profile. It also compares a new estimate with the firm’s past estimates and points out line items outside the usual range.

Djait-Paulien said large general contractors use Tyce across estimating, preconstruction, contract review and project management. Smaller firms with one or two estimators use it to add capacity, and a drywall or formwork sub can pull out only its own trade’s scope.

Evidence so far

The evidence is Tyce’s funding announcement, a quote in it from Eastern Construction’s CEO, and Djait-Paulien’s interview with The Construction Broadsheet.

These claims come from Tyce and Eastern Construction:

Eastern Construction has rolled Tyce out across its preconstruction teams and cut the time spent bidding on and managing projects by at least 40%.
Dean Walker, Eastern’s CEO, said in the announcement: “With Tyce, tasks that took our preconstruction team five hours now take 15 minutes.”
Djait-Paulien told the Broadsheet that the largest gains come from reviewing bid packages, leveling sub bids, reviewing contracts, and summarizing reports and meeting minutes.
She also said customers are bidding more work with the same staff.

Time savings reported for Tyce at Eastern Construction
Measure	Reported by	Before Tyce	With Tyce	Change
Unnamed preconstruction tasks	Dean Walker, CEO, Eastern Construction	5 hours	15 minutes	95% less (our arithmetic)
Time spent bidding on and managing projects	Tyce announcement	Not stated	Not stated	At least 40% less

Source: tyce.ai

Walker’s five hours to 15 minutes refers to preconstruction tasks that the quote leaves unnamed, and it works out to a 95% cut for those tasks. The 40% figure covers bidding and project management time together. The method, time period and number of bids behind it are not in the announcement.

Accuracy rates are not yet public. Tyce told the Broadsheet that it benchmarks its output against customers’ own estimates and plans to publish accuracy and win-rate figures as its customer base grows. Bricks & Bytes, a construction technology newsletter, said accuracy is the first question a chief estimator should ask.

How to measure it on your projects

Track review hours per bid. Define it as the hours estimators and preconstruction staff log against one bid, from receipt of the documents to issued trade scope sheets and leveled sub bids. Split the hours by task: document review, scope sheets, bid leveling and proposal writing. That split lets you check which of Walker’s tasks moved on your own work.

Capture the baseline on your next five to ten bids before the pilot starts, and note each bid’s size and delivery method. Review hours on a small tenant improvement and a large hospital bid differ widely, so compare pilot bids with baseline bids of similar size.

Track what the review catches alongside the hours. For each pilot bid, record how many flags the tool raised and how many held up when an estimator checked the cited sheet. Then count the scope items the estimator found that the tool missed. After award, tag any change order that traces back to a scope gap at bid time. Our posts on change orders by cause and bid-hit ratio show how to set up that tag and check whether the extra bids turn into wins.

A first step on one job: on the next bid that comes in, have each estimator log hours by task in a shared sheet until the bid goes out.

Availability and cost

Tyce sells an annual subscription, tiered by company size and usage, with no per-project fee, according to the Broadsheet. Prices are not listed. Djait-Paulien said setup, including onboarding and training, usually takes a few weeks, and that past estimates and cost data are optional at the start. She also said customer data is never used to train models for other firms.

Tyce says it serves customers in Canada and the United States, and it told the Broadsheet it expects the US to become its largest market. For a European product aimed at the same bid review work, see our post on Kuro’s AI tender review.

NCC makes Buildots a preferred supplier after AI progress tracking lifted one project's Percent Plan Complete 2.3 times

Sat, 31 Oct 2026 07:34:22 -0400

Buildots announced on October 28, 2026, that NCC’s Building Nordics business, which builds in Denmark, Finland and Norway, has signed a master services agreement naming Buildots a preferred supplier. Buildots sells AI progress tracking. The software compares what was installed on site with the 3D model and the schedule. The agreement follows several years of NCC projects in Finland and Denmark, and it affects the weekly progress check and the trade planning meeting that depends on it.

What it does

On most jobs, a superintendent or project engineer collects progress by walking the floors and asking each trade where it stands. Before Buildots, NCC says in its case study, project managers judged progress from manual, subjective reports collected from third parties.

With Buildots, a site team member walks the job wearing a 360-degree camera. Computer vision, software that recognizes objects in images, matches each captured element to the 3D model. Buildots then compares that status with the schedule and reports progress by trade and by area. NCC’s teams used a matrix view that showed, floor by floor and zone by zone, the percent complete for activities such as HVAC pipes, cable tray and first-fix plumbing.

The output goes into the weekly trade meeting. In one example from the case study, Buildots counted 45 climate beam units still uninstalled out of 150 expected by that date, while the trade’s own report said only a few were left. On another NCC job, the data showed HVAC work running slower than planned, and the project manager and HVAC subcontractor agreed to add labor.

Evidence so far

The evidence is Buildots’ press release, two Buildots case studies, a Buildots blog post, and comments from NCC staff quoted in those documents. Bricks & Bytes, a construction technology newsletter, labels the headline figure company-reported and not independently verified.

These claims come from Buildots and NCC:

NCC piloted Buildots on an 8,000-square-meter residential project in Helsinki. It then added three more projects, among them the 9,000-square-meter Fredriksberg D project and a hospital, for 68,500 square meters (737,000 square feet) tracked.
Tasks finished on time, measured by PPC, rose 2.3 times. The blog post puts it as a rise from a 30% baseline on past projects to almost 70%.
Time spent on manual reporting fell 70%.
A later case study covers six Finnish projects with 172,000 square meters of construction area, and says weekly PPC ratios improved over time.

The release describes the PPC gain on one project as a 230% increase. The case study figures, 30% to almost 70%, work out to a ratio of 2.3, which is an increase of about 130%. The table below shows PPC readings at points through the job, as Buildots charts them in the case study.

PPC on the NCC Buildots project compared with NCC's traditionally managed projects, by share of project complete
Share of project complete	PPC managed traditionally (%)	PPC managed with Buildots (%)
35%	27	52
40%	27	57
45%	27	71
50%	27	37
55%	27	32
60%	31	57
65%	22	72
70%	27	81
75%	27	90
80%	22	96

Source: pages.buildots.com

The Buildots line falls to 37% and 32% at the 50% and 55% marks, then rises to 96%. The cause of the dip and the number of projects behind the traditional line are not stated in the case study. NCC compared Buildots with three alternatives before choosing it, according to the case study.

Kati Tauriainen, who heads NCC’s Building Finland division, said in the release: “Our experience using Buildots on live projects gave us the evidence we needed.” NCC service manager Nithesh Frei Nadarajah says in the case study that his team spends less time in meetings and arguing with subcontractors.

How to measure it on your projects

Track weekly PPC by trade. Divide the number of tasks finished as committed by the number each foreman committed to in the weekly work plan. A task that is 90% done counts as not done. Record the reason for every miss, such as an open RFI, a late delivery or a predecessor trade that has not released the area.

Capture at least six weeks of baseline before the pilot starts, on the same job if you can. Log two other figures alongside it. The first is the hours the superintendent and project engineers spend each week collecting and writing up progress. The second is the number of times a trade’s reported percent complete differed from what a walk found. Score PPC by the same rule during the pilot, so any change can be traced to the tool. Our post on tracking total float across schedule updates covers the next step, since float shows how many days the misses cost.

A first step on one job: have the superintendent score last week’s lookahead commitments as done or not done, and keep that sheet every Friday.

Availability and cost

The agreement covers NCC Building Nordics operations in Denmark, Finland and Norway. Buildots says its other customers include HOCHTIEF, Bouygues, Intel and VINCI Construction. Prices are not listed in the release, and Buildots’ case study pages send buyers to a demo request. For a competitor that publishes per-square-foot pricing, see our post on Track3D’s progress tracking.

Bobyard adds AI submittals that build product data packages from a trade contractor's takeoff

Fri, 30 Oct 2026 07:35:30 -0400

Bobyard, a San Francisco company that sells AI takeoff and estimating software to trade contractors, announced AI-prepared submittals on October 29, 2026. The tool matches the materials in a contractor’s takeoff to manufacturer cut sheets and assembles a product data submittal package for the general contractor. The same release ties Bobyard’s takeoff, cost database, estimate table and proposal generator into one workflow, and adds drywall, insulation, doors and windows to the trades it covers.

What it does

On a subcontractor’s side, the product data submittal usually falls to the estimator or project manager who won the bid. They take the material list, find each product’s cut sheet on a vendor site or in a catalog PDF, mark up the right model on each page, and format a cover sheet by hand. Bobyard’s blog post describes that work as hours of screenshots and cross-referencing.

The tool follows the order of that work. A team uploads its manufacturer catalogs once, and Bobyard splits each catalog into individual products tagged with manufacturer, model and size. The library is shared across the company, so a catalog one estimator uploads is there for every later job.

For each job, the user uploads the spec or the submittal requirements. Bobyard’s product page says the AI then lists the required packages, such as product data, warranties and test reports, with spec references. Materials come in from the Bobyard takeoff with descriptions filled from the legend. Teams that did their takeoff elsewhere can paste in a list instead.

Bobyard’s AI assistant, Bob GPT, then searches the library and the web for each item. It returns candidate cut sheets ranked by a confidence score, which is the model’s own estimate of how likely a match is correct. It also highlights the model number on the page. The user confirms every product and every page before it goes into the package. Bobyard’s FAQ says the AI does not choose the product.

How Bobyard describes the submittal workflow, step by step
Step	What goes in	What Bobyard does	Who checks it
Build the product library	Manufacturer catalogs uploaded once	Splits each catalog into products and tags manufacturer and model and size	Anyone on the team who uploads
Set up the packages	The spec or the submittal requirements	Lists the required packages such as product data and warranties and test reports with spec references	Project manager or estimator
Load the items	Materials from a Bobyard takeoff or a pasted list	Creates submittal items and fills descriptions from the takeoff legend	Estimator
Match product data	Each submittal item	Searches the library and the web and ranks candidate cut sheets with confidence scores	The user confirms every product and page
Package and send	Confirmed cut sheets	Builds a branded cover page and table of contents and exports one PDF	Project manager before sending to the GC
Resubmit	GC review marks per item	Builds a resubmittal with only the items marked Resubmission Required	Project manager

Source: bobyard.com

The output is one PDF with a branded cover page, a table of contents and the marked-up spec sheets. Each item can then be marked Approved, Approved as Noted, or Resubmission Required. A resubmittal package includes only the flagged items. For plants and other organic materials, which have no manufacturer cut sheet, the tool generates a product data page with an image and cited sources for the user to check.

Bobyard describes the tool for product data packages, and its pages do not cover shop drawings or samples.

Evidence so far

The evidence so far is Bobyard’s announcement, its product pages, and comments from unnamed participants in its beta program. Bobyard identifies the quoted participants as landscaping contractors.

These claims come from Bobyard:

In a survey of beta participants, most said a single submittal took 2 to 5 hours to prepare, and some put it at 10 hours or more.
The tool cuts submittal prep time in half.
Customers cut takeoff time by an average of 65% and submit three to five times more bids per estimator, according to the release.
Building the submittals feature took one engineer about two months, according to founder and CEO Michael Ding.

One landscaping contractor in the beta said the tool was about twice as fast as the old method for irrigation and soil submittals. Another wrote, “Search accuracy on our irrigation catalogs was flawless, and setting up the package and items was straightforward.” Both comments come from irrigation and soil work, where the catalogs are narrow. Match rates on drywall, doors or MEP equipment, and the size of the beta survey, are not yet public.

How to measure it on your projects

Track prep hours per product data package. Start the clock when someone opens the material list for the package and stop it when the PDF goes to the GC. Log rework hours on a resubmittal separately.

Capture the baseline before the pilot:

Pick three to five recent product data packages from one trade, with a similar item count.
Ask the person who built each one for the hours it took, and count the line items in each.
Record whether each package came back Approved, Approved as Noted, or Resubmission Required.

On the pilot job, log the same three figures for each package built in Bobyard. Count the line items where the top-ranked match was wrong and the user had to pick another or search again. That count tells you how much of the saved time goes to checking. The first package will also include catalog uploads, so log those hours as setup.

A first step you can take on one job is to pull the last approved product data package and its resubmittal from the project file and count the line items the GC sent back.

Availability and cost

Bobyard’s blog post on the feature is dated October 1, 2026, and tells existing customers to ask their customer success manager to add submittals to their plan. New customers book a demo, and Bobyard prices through that demo request. Teams can use submittals without Bobyard’s takeoff. The company’s earlier releases on Business Wire include a $35 million Series A led by 8VC.

Bluebeam Max puts AI drawing checks and a Claude connection inside Revu for $590 per user a year

Thu, 29 Oct 2026 07:34:57 -0400

On October 29, 2026, the Nemetschek Group grouped its AI products for architecture, engineering and construction under one name, Max. The construction member is Bluebeam Max, a premium plan for Bluebeam Revu that has been on sale worldwide since May 19. It adds AI checks of drawing sets, AI comparison of revisions, and a connection that lets an AI assistant such as Anthropic’s Claude carry out markup and reporting tasks in Revu. The work it touches is drawing review, revision comparison and takeoff markup. Bluebeam says more than 4 million professionals use its software.

The three Max products Nemetschek announced on October 29, 2026
Product	Discipline	Status on October 29 2026
Bluebeam Max	Construction	Available globally since May 2026
Allplan Max 2027	Engineering	Available globally, launched the week of October 26
Archicad Max	Architecture	Available with Archicad 30, rolling out by region

Source: nemetschek.com

What it does

Start with the preconstruction manager who gets a reissued drawing set before bid day. Today that person overlays the old and new sheets, scans for clouds that were missed, and checks that tags and gridlines still line up before pricing the changes. Bluebeam Max puts AI into three steps of that review.

Smart Review scans a set for problems that Bluebeam lists by name: missing sheets, missing door tags, missing plumbing fixture tags and gridline inconsistencies. It uses computer vision, software that recognizes shapes and symbols in an image, and returns its findings as markups on the sheets, a dashboard and trackable issues. Bluebeam says Smart Review is in preview and is optimized for US vertical commercial construction.

What Bluebeam says Smart Review looks for
Item Smart Review checks	Where Bluebeam lists it
Missing sheets	Bluebeam Max page and pricing page
Missing door tags	Bluebeam Max page and pricing page
Missing plumbing fixture tags	Bluebeam Max page and pricing page
Gridline inconsistencies	Bluebeam Max page and pricing page
Design discrepancies, scope gaps and missing information	May 19 launch release and October 29 Nemetschek release

Source: bluebeam.com

Smart Overlay compares sheet sets across disciplines and drawing scales and reports what changed. It produces visual overlays, reports and trackable comparisons. It is also in preview.

The third piece connects Revu to an outside AI assistant through the Model Context Protocol (MCP), an open standard that lets an assistant call another program’s functions. A user with the Claude desktop app can type a request, and Claude reads the open PDF’s text and markup data and then creates or edits markups, creates custom columns or builds a report. The Bluebeam University course on the feature walks through four examples:

Changing the subject on a group of markups.
Building an Excel summary of markups, filtered by page label and author.
Writing a narrative of changes from an existing drawing overlay.
Checking title block information against the ISO 7200 standard with a saved Claude project.

Nemetschek’s release says Bluebeam Max also works with GitHub Copilot CLI and AnythingLLM. Bluebeam’s own FAQ says the plan connects to Claude first and is built to connect to other assistants as they add MCP support to their desktop apps.

The other Max features handle sheet layout and markup work. Stitching, which Bluebeam also labels as AI, joins sheets from different parts of a project into one continuous view, and Bluebeam pitches it for large and linear infrastructure projects. Magic Markups duplicates a markup as another type, converts a shape into a measurement, or offsets a markup a set distance. Connected Studio Sessions link a Bluebeam markup to the matching spot in a Revit sheet or 3D view.

Bluebeam Max features, where their data goes, and whether an admin can turn them off
Feature	What it does	Where the data is processed	Admin can turn it off
Smart Review	Checks pages, gridlines, doors and similar items and builds a dashboard of marked issues	Uploaded PDFs processed on AWS servers in the UK	Yes
Smart Overlay	Finds design changes and discrepancies across drawing sets and disciplines	Uploaded PDFs processed on AWS servers in the UK	Yes
AI connection through MCP	Lets an AI assistant such as Claude run Revu tasks from typed prompts	With user permission, prompts, PDF text, markup data and Studio file metadata go to the AI client	Yes
Stitching	Combines multiple drawing sheets into one navigable view	On the user's machine	No
Magic Markups	Duplicates, converts or offsets existing markups	Not listed on the AI transparency page	Not listed

Source: bluebeam.com

Evidence so far

The evidence is Bluebeam’s May launch release, Nemetschek’s October 29 release, and comments from two beta customers. Both releases come from the company selling the product.

These claims come from Bluebeam and Nemetschek:

More than 2,000 people took part in the Bluebeam Max beta, and Nemetschek says those early adopters reported significant productivity gains in preconstruction, design coordination and construction management.
Martin-Harris Construction, a Las Vegas construction manager, uses Smart Overlay and Smart Review to compare revisions for pricing and impact analysis during design and construction. James Miranda, its director of preconstruction, said, “Features like Smart Overlay help us instantly see what’s changed and where to focus, which is critical when you’re managing complex projects at scale.”
At KPFF’s Los Angeles civil office, an associate said the offset tool cuts down on trips back to CAD for parallel utilities and curved geometry.

Nemetschek’s October release gives one percentage, and it covers a different product. The company says early beta users of Allplan Max, its engineering offering, finished selected tasks up to 36 percent faster. Detection rates for Smart Review and Smart Overlay, and measured time savings for Bluebeam Max, are not yet public.

Bluebeam’s training material sets limits on the Claude connection. The course says results depend on the document text and markup data that MCP exposes, that users should review generated reports rather than treat them as a guarantee of accuracy or compliance, and that AI provider usage limits and company network controls can affect the workflow.

How to measure it on your projects

Track reissue review hours: the staff hours from receipt of a revised drawing set to the point where the changes are logged and priced. Pair it with a count of drawing discrepancies found at review, so faster reviews cannot hide misses.

Capture the baseline before turning on Smart Review or Smart Overlay:

Pull the last three reissued sets on a comparable job and record the review hours your preconstruction or project team logged for each.
Count the discrepancies your reviewers flagged on each set, such as missing tags, mismatched gridlines and unclouded changes.
Count RFIs from the following 60 days that trace back to a drawing problem in those sets, since each one marks something the review missed.

During the pilot, have the reviewer mark each AI flag as a real problem, a false flag or something already found. The share of flags that turn out to be real problems is the tool’s precision. The share of your reviewer’s own findings that the tool also flagged is its recall. A first step on one job is to run Smart Review on a set your team has already reviewed and compare the two lists, using a trial seat.

Earlier posts covered Document Crunch’s drawing review against contracts and specs, BuildCheck’s funding for AI drawing review and EarthCam’s connectors to ChatGPT and Claude.

Availability and cost

Bluebeam Max is available globally and sold per user on annual billing. The introductory price is $590 per user per year, $150 more than the Complete plan. Buyers at that price keep it through their 2027 renewal, and Bluebeam will announce the standard list price for later renewals. Existing users can start a trial through Org Admin, and since September 2026 the plan is free for eligible students and educators.

Bluebeam plan prices per user per year, October 2026
Plan	Price per user per year (USD)	What the plan adds
Basics	$260	PDF markups, length and area measurement, joining Studio sessions
Core	$330	Perimeter, count, angle and volume measurement; overlay and compare; hosting Studio sessions
Complete	$440	Dynamic Fill, Quantity Link to Excel, markup reporting, batch hyperlinks and slip sheets, scripting
Max (introductory)	$590	AI drawing review and comparison, AI connection through MCP, Stitching, Magic Markups, Connected Studio Sessions with Revit

Source: bluebeam.com

Plans can be mixed, so a firm can put Max on its estimators and preconstruction staff and leave field users on a lower plan. Revu runs on Windows. Revu 20 reached end of support on July 31, 2026, and reaches end of life on December 31, 2026, so firms still on that version will pick a subscription plan before year-end.

NetSuite plans AI agents that turn estimates into job budgets and draft pay applications

Tue, 27 Oct 2026 07:40:43 -0400

Oracle NetSuite announced on October 26, 2026, at its SuiteWorld conference in Las Vegas, that it plans a construction accounting product called NetSuite for Construction Businesses. The product includes two AI agents. One sets up a new job from the estimate, and the other prepares the monthly payment application. Both work on the handoff between estimating, project management and billing at general contractors, subcontractors and specialty-trade firms, the three groups NetSuite names as users.

What it does

Job setup after award usually falls to a project accountant or a project manager. They take the final estimate, build the schedule of values for the owner contract, and load a budget by cost code into the accounting system. NetSuite says its Estimate Import Agent converts an estimate into the contract, the schedule of values and a cost-coded job budget. Each imported line traces back to the estimate, so a reviewer can check it against the source.

An AI agent here means software that carries out a sequence of steps, such as reading an estimate and creating the records it implies, and then hands the result to a person for review. NetSuite says users reach both agents through Ask Oracle, its chat assistant, by typing requests in plain language.

The Payment Application Agent works on the monthly billing cycle. According to the announcement, it prepares a pay app from verified progress, approved change orders and retainage terms. The project manager confirms percent complete, and the billing team reviews and approves the application before it posts.

The rest of the planned product supports those two agents. NetSuite lists:

A Construction Workbench showing budget, committed cost, actual cost, billed to date and estimated cost to complete for each job.
Change orders that carry into the schedule of values and later pay apps.
AIA-style payment applications built from the schedule of values, including stored materials.
WIP accounting that compares earned and billed revenue, adjusts for over- and underbilling, and tracks retainage through final collection.
Construction payroll with certified payroll, wage determinations, multi-rate overtime and garnishments.

A contractor that keeps its WIP schedule in Excel will recognize those workbench columns from our work-in-progress schedule walkthrough.

Evidence so far

The evidence so far is NetSuite’s press release and its construction product page. These claims come from NetSuite:

The Estimate Import Agent automates job setup and gives each job a consistent financial baseline.
The Payment Application Agent cuts manual work and speeds up billing.
The Construction Workbench helps project managers and controllers spot margin problems earlier.

Evan Goldberg, NetSuite’s founder, says in the release that small gaps between project activity and financial records can quickly hurt profit and cash flow. The release describes planned capabilities and carries Oracle’s future-product disclaimer, which says features, timing and pricing may change. Customer results for the construction agents, such as the accuracy of the estimate conversion or the time saved per pay app, are not yet public.

How to measure it on your projects

Track pay application cycle time: the calendar days from the end of the billing period to the day the pay app goes to the owner or the GC. Record it per job, per month. A pay app that goes out later gets paid later, and our DSO post shows how that lag feeds days sales outstanding.

Before a pilot, pull the last six billing cycles for five to ten active jobs. For each pay app, record:

The billing period end date and the submission date.
Whether the owner or architect returned it for corrections, and how many days the resubmission took.
Hours the project manager and billing staff spent on it, if your timesheets break that out.

During the pilot, log the same items. Also log every line where the agent’s draft differed from what the project manager approved. The share of lines changed is the error rate that matters on your jobs. Job setup can be measured the same way, as hours from award to a loaded budget and the count of budget lines corrected in the first month.

A first step on one job: pick the job with the most change orders this year and write down the submission dates of its last six pay apps.

Availability and cost

NetSuite says the product will start with a limited US rollout, with broader availability expected later. The construction page says NetSuite sells an annual subscription made up of the core platform, optional modules and user count, plus a one-time implementation fee. Launch dates and pricing for the construction product and its agents are not in the release. NetSuite says it has more than 44,000 customers in 220 countries and dependent territories, and its construction page notes that some capabilities described there require partner integrations.

Document Crunch, now part of Trimble, adds drawing sets to its AI review of contracts and specs

Mon, 26 Oct 2026 07:43:17 -0400

Document Crunch, the contract review software Trimble agreed to buy in April, announced on October 21, 2026, that its AI can now read construction drawing sets alongside contracts and specifications. The announcement says the feature is available to all Document Crunch users. It affects the preconstruction document review, the stage where an estimator or project manager reads the bid set for scope gaps and conflicting requirements before the price goes in.

What it does

Today that review means someone reads the contract, then the spec book, then pages through the drawings and tries to connect a general note on one sheet to a requirement in the specs. Josh Levy, a Document Crunch co-founder who is now Trimble’s vice president of risk and compliance solutions, says in the announcement that teams work across hundreds of drawing sheets on a project.

The new feature starts with the drawing set. Document Crunch says its CrunchAI engine extracts and indexes the text on each sheet, including notes, allowances and the language that defines scope. Indexing here means the software builds a searchable record of every word and where it sits, the way a spec book’s table of contents points to a section. The announcement describes text extraction only.

Once the set is indexed, a user asks questions in a chat window across drawings, contracts and specs together. Each answer links back to the source page so a reviewer can open the sheet and check it. Document Crunch lists these checks:

Verify the sheet index and flag missing sheets or spec sections.
Flag discrepancies between disciplines.
Find requirements that conflict between drawings, contracts and specs, along with scope gaps and unclear responsibilities.
Run the same company checklist on every drawing set.

Findings can be turned into review items, assigned to a person and tracked inside the project. The users are the preconstruction group, the estimators pricing a package and the project manager who inherits the job after award.

Evidence so far

The evidence is Document Crunch’s announcement and one customer comment inside it. These claims come from Document Crunch:

General-purpose AI models have trouble with large, text-dense drawing sheets and lack the project context to connect requirements across documents.
The drawing review finds conflicts and scope gaps early enough for a team to clarify, price or resolve them before they turn into rework or disputes.

Kristen Ortiz, vice president of project controls at Hill & Wilkinson, describes a renovation where the team had to redo plumbing and the drawings did not show the concrete cutting and backfill that work would require. “Document Crunch caught those additional requirements, helping our team identify work that might otherwise have been overlooked,” she said in the announcement.

Earlier figures cover the contract side of the product. In its June platform release, Document Crunch said it had been deployed on more than 10,000 customer projects that served more than $350 billion in annual construction volume. In the same release, a Ben Hur Construction vice president estimated that contract and scope review time fell 80%. That figure predates the drawing feature. Accuracy figures for the drawing review, such as the share of real scope conflicts it flags, are not yet public.

How to measure it on your projects

Track scope-gap change orders as a percent of original contract value. Count only the change orders your team coded to a scope gap or a conflict between documents, the items a pre-bid review is meant to catch. Our change order rate post walks through coding a log by cause.

Capture two baselines before the pilot:

From five to ten closed jobs of similar type, total the scope-gap and document-conflict change orders and divide by the original contract value.
On your next two or three bids, have the estimator log hours spent on document review and the number of clarifications or exclusions written from it.

During the pilot, run the tool on new bid sets and log every item it flags. Mark each one as a real issue, a duplicate of something your team already caught or a false alarm. The share marked real is the tool’s precision on your documents, the fraction of its flags worth an estimator’s time. The change order percentage will take months to show up, so review hours and the flag log are the early read.

A first step on one job: pick a recently awarded project, run its bid set through the tool and compare the flags with the change orders that job has already logged.

Availability and cost

The announcement says drawing review is available now to all Document Crunch users, inside their existing workflows. The June release said the project-level platform is available in the United States and Canada. Trimble announced its agreement to acquire Document Crunch on April 2, 2026, and Document Crunch now calls itself a Trimble company. The April release said Trimble plans to bring Document Crunch’s contract rules into Trimble Construction One project management and ERP workflows. Pricing is not listed in the announcements.

BRG survey finds 68% of North American construction executives spent $250,000 or more on AI and 15% call it fully integrated

Sun, 25 Oct 2026 07:35:40 -0400

The consulting firm BRG published AI in Construction: Measurable Results, Unclear Benchmarks on October 9, 2026, the fourth report in its AI Industry Spotlight Series. In July 2026 BRG surveyed 109 executives who work on their organizations’ AI efforts, and the report’s figures describe the 65 North American respondents unless it says otherwise. Of that group, 68% spent at least $250,000 on AI over the previous 24 months, and 15% say AI is fully integrated across their projects and operations. The questions run through estimating, bidding and contract formation, procurement, scheduling, field operations and warranty work, and they end with how firms measure the return.

What it does

The survey tests no product. BRG split respondents into builders, who deliver projects (58 respondents across all regions), and owners, who develop and commission them (51). Construction Dive reports that the builders were mainly small and midsize general contractors and construction managers, and the owners mostly larger developers and REITs. For each function, a respondent said whether AI is implemented, planned, wanted with no current plan, or not being explored.

AI implementation status by function, North American respondents, July 2026
Function	Implemented (%)	Planning to implement (%)	Aspire with no current plans (%)	Not exploring (%)
Planning and feasibility	58	32	2	8
Project management and scheduling	52	32	14	2
Warranty and post-completion services (builders only)	52	36	8	4
Corporate and administrative functions	48	35	12	5
Procurement and supply chain	46	37	15	2
Field operations and site execution	46	31	18	5
Bidding and contract formation	40	38	20	2
Design and engineering	38	45	12	5
Risk, claims and dispute management	37	37	20	6

Source: media.thinkbrg.com

Planning and feasibility leads at 58%, and BRG says the share reaches 80% among builders. On the builder side, the report lists cost estimating, schedule modeling, feasibility analysis and permitting support as that function’s uses. Design and engineering sits at 38% implemented, with the largest share in the table planning to implement, at 45%.

Of the North American respondents, 37% spent $1 million or more over 24 months, and 6% made no dedicated AI investment. On integration, 29% expect to get there within twelve months of July 2026 and 31% within one to two years, while 5% do not expect AI ever to be fully integrated.

Evidence so far

The evidence is one survey of self-reported answers, run by a firm whose services include economics, disputes and investigations work. Every respondent is involved in AI at their organization, so the adoption rates describe firms already doing AI work. Builders and owners are pooled in most figures. A margin of error is not stated, and with 65 North American respondents, one answer moves a percentage by about 1.5 points.

Respondents reported savings by project phase, and BRG counts the share reporting savings above 10%.

AI-delivered cost savings by phase, share of respondents. The last column is BRG's published figure.
Phase	Less than 5% (%)	5 to 10% (%)	11 to 20% (%)	21 to 30% (%)	More than 30% (%)	Too early to quantify (%)	Above 10% (%)
Pre-construction	4	36	31	18	9	2	58
Construction	4	29	38	25	4	0	67
Business operations	9	27	31	22	9	2	62

Source: media.thinkbrg.com

Read the savings figures with these limits:

The report does not say what the savings percentages are measured against, such as the cost of the task, the phase budget or the project.
For post-construction, the text gives 58% above 10%, while the bars in BRG’s chart add to 49%. This table leaves that phase out until the figure is settled.

Anyway, the workforce answers lean toward reorganizing teams. In BRG’s results, 18% added headcount in some areas because of AI and 11% reduced it, 38% reskilled existing staff, and 51% changed how much they rely on outside vendors, consultants or contractors.

The risk questions matter most for anyone drafting an AI policy. Liability for errors in AI-generated work is the concern most often cited as restricting use, at 43%, and 40% point to unclear treatment of AI in contracts. Policies covering human oversight and accountability exist at 42% of firms, and 34% have policies on accuracy and reliability. Construction Dive notes that construction had the lowest policy coverage of the four industries BRG has surveyed. (The other three were retail, energy and financial institutions.)

How to measure it on your projects

So, on measurement, respondents split across methods. The largest group, 37%, relies mainly on financial metrics. Another 23% track usage, which shows whether a tool is in use and leaves the job’s outcome unmeasured.

How respondents primarily evaluate ROI from AI
Primary way ROI from AI is evaluated	Share of respondents (%)
Financial metrics (cost savings, margin improvement, reduced change orders)	37
Operational metrics (schedule performance, labor productivity, rework or error reduction)	29
Usage or adoption metrics (active users, tools deployed across projects)	23
Qualitative assessment (team feedback, post-project reviews)	5
Not currently evaluating ROI	6

Source: media.thinkbrg.com

“Contingency is a useful construction-specific measure,” says Ben Nolan, a BRG managing director, who suggests reading AI’s return in uncommitted contingency at stages of the project. A contractor can turn that into one number. The uncommitted contingency ratio is the contingency not yet committed to a change or a buyout overrun, divided by the original contingency, read at 25%, 50% and 75% complete.

Build the baseline from closed jobs before the pilot starts:

Pick five to ten closed jobs of similar type and size.
From the monthly cost reports, record the original contingency and the uncommitted balance at each of the three milestones.
Average the ratio at each milestone.

Then track the same ratio on jobs where AI handles a defined task, such as drawing conflict review or estimate checks, and compare at each milestone. Pair it with change order rate by cause, since a higher ratio is only useful if the commitments it avoided would have become changes. Start by pulling the contingency history for one closed job this week.

Availability and cost

The report is a free PDF on BRG’s site with no registration form. It also compares regions: 18% of Asia-Pacific respondents and 9% of European respondents say AI is fully integrated, against 15% in North America.

Procore says its Magpie 3.0 engine completed 78% of 200-plus construction tasks, 18 points above the closest general AI model

Sat, 24 Oct 2026 07:36:21 -0400

Procore shared a preview of its own construction AI benchmark at Groundbreak 2026, its annual customer conference, and posted the results on its blog on October 22, 2026. A benchmark is a fixed set of tasks scored the same way for every system, and this one has more than 200 tasks drawn from RFI, submittal, drawing, schedule and change work. Procore says Magpie 3.0, the new reasoning engine behind Procore AI, completed 78% of them. That is 18 percentage points above the closest general-purpose frontier model, meaning the large commercial models from OpenAI, Anthropic and Google. On October 21, at the same conference, Procore put Procore AI Go, a voice app for field crews, into open beta.

What it does

Start with the project engineer reviewing an emergency generator submittal. That person reads the package against the mechanical spec section and the drawings, then writes a response. A requirement sitting in another division, such as an electrical section that names the transfer switch manufacturer, gets caught only if the reviewer thinks to look there.

Procore describes three layers in Procore AI that do that cross-referencing:

A multimodal index, meaning one that reads images as well as text, opens drawings, site photos and meeting minutes. On a drawing it picks out gridlines, fixture symbols, dimension strings and revision clouds.
A construction ontology, a map of how project records relate to each other, links RFIs, submittals, observations, daily logs, meetings, drawings, photos and change events to the same trade, assembly and location.
Magpie 3.0 works through a question in steps across those linked records. It chooses a tool for each step, such as an image analyzer or a document parser, and cites the governing drawing detail or spec section in its answer.

For submittals, two Procore tools described in a September 18 product post sit on top of this. Submittal Builder reads the spec book and drafts the submittal log, including closeout, warranty and O&M items. The Submittal Reviewer agent checks returned packages against the specs and the contract, flags deviations and unapproved substitutions, and drafts a response. The project engineer or PM reviews the draft and makes the call.

Procore AI Go takes the question-and-answer part to the field. A foreman presses to talk, asks about drawings, specs, RFIs, submittals, daily logs, observations or inspections, and hears a spoken answer drawn from the project’s Procore data. Procore says later releases will let crews create daily logs and observations and draft RFIs by voice, and will add its Daily Log Agent and Site Safety Agent to the app.

What Procore announced around Groundbreak 2026 and where each piece stands
Capability	What it does	Status in October 2026
Multimodal index	Reads drawings, site photos and meeting minutes as content, such as gridlines, symbols and revision clouds	Generally available in all Procore AI packages
Magpie 3.0	Reasons across linked project records and cites the governing drawing, spec section or record	Open beta for Procore AI customers through Procore Explore
Construction ontology	Links RFIs, submittals, observations, daily logs, meetings, drawings, photos and change events	Built in; drawing-photo linking and rework pattern detection roll out over the coming year
Procore AI Go	Voice questions and spoken answers on drawings, specs, RFIs, submittals, daily logs, observations and inspections	Open beta in North America on Digital Coworker Starter, Pro and Enterprise
AI Go on Meta AI glasses	Hands-free questions and first-person photos	Closed beta
Daily log and observation creation in AI Go	Create daily logs and observations and draft RFIs by voice and images	Announced as coming

Source: procore.com

Evidence so far

The evidence is Procore’s benchmark preview, one customer example in the same post, a customer comment in the AI Go release, and one customer figure from a July release. Procore built the tasks and ran the test. Procore says Thomas Kommer of the California Center for Construction Education at Cal Poly San Luis Obispo reviewed the task scenarios, scoring rubrics and correct answers.

These claims come from Procore:

Superintendents, project managers and project executives contributed the questions behind the tasks. The tasks fall into five workflow families and four difficulty tiers, from Easy to Expert.
A task passed only if the answer was correct and supported by the project record. Speed and cost were measured separately.
The general models got the same project data and tasks through Procore’s public API, the route a customer would use to connect them today.
Magpie 3.0 passed 78% of tasks. The closest frontier model scored 18 points lower, which works out to about 60%.

Task success rate on Procore's construction benchmark, early preview, October 2026
System	Task success rate (%)	Basis
Procore AI with Magpie 3.0	78%	Reported by Procore
Closest general-purpose frontier model	60%	Computed here as 78 minus the 18-point margin Procore reports

Source: procore.com

The five workflow families in Procore's benchmark
Workflow family	Example tasks Procore lists
RFI management	Investigating root causes, identifying linked drawing callouts, catching duplicates, evaluating schedule impacts
Submittal management	Cross-checking submittals against spec divisions, identifying non-approved substitutions, flagging lead-time risks
Drawings and specifications	Resolving revision-sensitive details, tracing section cuts across sheets, detecting constructability clashes
Schedule and delay	Evaluating float, identifying critical-path impacts from field observations, analyzing trade sequencing
Cost and change	Identifying unapproved scope changes, calculating budget exposure across cost codes, validating potential change orders

Source: procore.com

Procore calls these results an early preview and says it will publish the full set. Scores by model, by difficulty tier and by workflow family are not yet public, and neither are the speed and cost measurements. The post names the three model families tested and gives one score for the closest of them.

The field example comes from Bernards, a California builder. Procore says project manager Joseph Kim ran a generator submittal through Procore AI, which found that the package met the mechanical spec but missed an electrical spec requirement that the automatic transfer switch come from the generator’s manufacturer. Kim’s comment: “It caught something from a different spec.”

Procore’s July 23 release says Haskell cut submittal reviews from seven days to 10 minutes with Procore AI. The number of submittals and the scope of the reviews behind the figure are not yet public.

Submittal review time at Haskell, as Procore reports it
Customer	Task	Before Procore AI	With Procore AI	Reported by
Haskell	Submittal review	7 days	10 minutes	Procore press release, July 23, 2026

Source: procore.com

For AI Go, the evidence is a comment from Andrew Baker, CIO of Sequoyah Electric, who says his foremen ask questions while looking at the problem and get an answer on the spot. Procore says it is part of a study with Moss Construction, designed by SBK Technologies, of how agents, AI Go and AI wearables affect administrative time. The study’s results are not yet out.

How to measure it on your projects

Track submittal review cycle time: calendar days from the date your team receives a submittal package to the date the response goes back to the subcontractor. Next to it, count spec conflicts caught at review, so a shorter cycle cannot hide missed requirements.

Capture the baseline before you turn on the agent:

Pull the last 30 to 50 closed submittals on a comparable job and record the received and returned dates for each.
Count the revise-and-resubmit responses, and mark the ones whose reason came from a spec section outside the package’s own division.
Count change events or field issues from the last six months that trace back to an approved submittal with a missed requirement.

During the pilot, have the reviewer log each agent flag as a real problem, a false flag or something already caught. The share of flags that were real problems is the tool’s precision. A first step on one job is to run the agent on 10 submittals your team has already approved and compare its flags with what your reviewer found.

Earlier posts covered Autodesk’s Submittals Evaluation Agent, Buildertrend’s AI Submittals and how to measure RFI turnaround before an AI pilot.

Availability and cost

Procore says the multimodal index is generally available across all Procore AI packages, and Magpie 3.0 is in open beta for Procore AI customers through Procore Explore. Ontology features shown at Groundbreak, including linking drawings and photos and automated rework pattern detection, roll out over the coming year.

AI Go is in open beta for North American customers on the Digital Coworker Starter, Pro and Enterprise packages. The Meta AI glasses version is in closed beta. Each Digital Coworker package includes Control Tower, which shows AI usage and credit consumption by agent, project and team member. Package prices are not listed in the releases.

EarthCam connects its jobsite camera AI to ChatGPT and Claude and adds seven detection models

Wed, 21 Oct 2026 07:40:24 -0400

EarthCam, the New Jersey company that rents and installs construction cameras, announced on October 20, 2026, a set of connectors that pass its jobsite detections to general AI assistants. A project team can ask ChatGPT or Claude about deliveries, safety conditions or site activity on a job, and the assistant answers from analysis EarthCam’s own software has already run on the camera imagery. The release also adds seven new detection models, several of them aimed at the safety review and delivery log.

What it does

On many jobs, camera footage gets reviewed after something happens. A superintendent or safety manager scrubs through footage when an incident, a disputed delivery or an owner question comes up, and the answer depends on someone finding the right hour of video.

EarthCam’s software analyzes the imagery as it comes in. According to the release, its computer vision (software that recognizes objects and events in images) runs on EarthCam’s servers and in some cases on the camera itself. The detections land in Control Center 9, EarthCam’s existing project platform.

The connectors are the new piece. They are built on the Model Context Protocol, an open standard that lets an AI assistant request data from outside software. EarthCam says the connector is available as an app in ChatGPT and as a connector in Claude, and that it can be added to other assistants that support the protocol, naming Gemini, Copilot and Perplexity. The assistant reads EarthCam’s finished detections, with links back to the high-resolution images, timestamps and metadata behind each answer.

EarthCam says teams can combine that camera data with weather and IoT sensor readings, plus schedules, budgets and specs from other connected systems. Where the assistant supports scheduled tasks, a team can set it to produce safety reviews, delivery summaries and executive updates on a recurring basis. The people most likely to use it are the superintendent, the safety manager and the project executive who writes the owner update.

The new models, as EarthCam describes them:

AI Trench Safety Analytics records workers and equipment entering and leaving defined excavation areas.
AI Housekeeping Hazard Detection flags visible debris, clutter and trip hazards.
Thermal AI Analytics pairs thermal imaging with object recognition to spot people and activity in darkness.
AI Delivery Verification recognizes boxes and pallets as they arrive and keeps time-stamped evidence.
AI Dump Truck Analysis counts truck movements and reads load status.
Visual Occupancy Analysis and Retail Signage Detection are the other two.

The release also mentions expanded schedule automation that compares planned milestones with observed site conditions.

Evidence so far

The evidence so far is EarthCam’s October 20 announcement, its product pages and one customer comment. These claims come from EarthCam:

Its AI analyzes jobsite imagery ten times faster than a customer’s own AI agent can interpret the raw images.
Agents working from EarthCam’s results use fewer AI usage credits than agents processing the images themselves.
In EarthCam’s benchmark testing, its models identified construction assets and objects more accurately than leading general-purpose models on the same images, with no hallucinations, the AI term for output a model makes up, such as a reported pallet that was never delivered.
Its models were trained on billions of construction images from its own cameras over three decades.

The image set, the scoring method and the accuracy rates from that benchmark are not yet public.

The customer comment comes from Scott Travis, director of northern operations at GBR Roofing. His crew used EarthCam’s detection to set exclusion zones around fall-through hazards while replacing rusted roof panels, and got an alert when someone crossed into one so they could check that fall protection was in place. “AI will alert you immediately when someone approaches a fall hazard,” Travis said in the release. His example describes the exclusion-zone alerts inside EarthCam’s platform.

How to measure it on your projects

Start by testing whether the detections are correct, since every answer the assistant gives depends on them. Track the confirmed alert rate: the share of EarthCam alerts or detections that a person, checking the linked image, agrees are correct. Run it per model, because trench entry and delivery recognition can perform differently on the same job.

For a time measure, track the hours a project executive or superintendent spends preparing the weekly owner or safety update. Capture the baseline before turning on scheduled summaries:

For four weeks, log the start and end time of each weekly update, including time spent pulling photos and footage.
Count the camera-related questions from the owner or the safety manager that week, and how long each took to answer.
For delivery verification, pull the delivery tickets for two of those weeks so you can later check each one against a camera detection.

During the pilot, log every summary the assistant drafts and every correction made before it goes out. Earlier posts covered viAct’s PPE detection on Alliad’s site cameras and Miter’s safety reports built from crew photos, and the confirmed alert rate works for comparing camera tools side by side.

A first step is to pick one camera on one job, pull a week of its detections and check twenty of them against the images.

Availability and cost

EarthCam says the connectors are available now to its customers through ChatGPT and Claude. Pricing is not listed on EarthCam’s product pages, which direct buyers to schedule a demonstration. Custom detection models are available on request from its data science team.

How to measure RFI turnaround before an AI pilot, with a 10-RFI example

Tue, 20 Oct 2026 07:38:59 -0400

A project engineer gets a question from the drywall foreman about a soffit detail that doesn’t match the reflected ceiling plan. The engineer checks the drawings and the specs, writes the RFI, attaches a marked-up sheet and sends it to the architect. Days later the answer comes back and goes out to the field. AI tools for RFIs work on both halves of that sequence, so a contractor who wants to test one needs a record of how long each half takes today.

This post sets up that record. It uses the definitions and benchmarks in a 2013 research paper from the Navigant Construction Forum, which the Construction Management Association of America hosts, and the settings in Procore’s RFI tool.

The definitions

Navigant defines an RFI as a written request from the contractor asking the owner to interpret or clarify a requirement of the contract documents. Its sample contract clause asks the contractor to state the issue, why and when a response is needed, and the contractor’s own reading of the requirement.

The same clause starts the response clock on the date the owner stamps the RFI received from the contractor and stops it on the date the response is stamped out to the contractor. The sample allows 10 working days. A footnote warns owners to say “working days” when they mean it, because most contracts count calendar days.

Procore’s RFI tool asks an administrator to set how many calendar days after creation an RFI is due, and it sends overdue email notices to assignees. Procore also tracks which party has the ball in court on each open RFI.

What the published benchmarks show

Navigant’s figures come from Aconex, which mined its records for 1,362 projects begun between 2001 and 2012. Each project had at least 100 RFIs, and together they logged 1,083,807 RFIs, an average of 796 per project. About 79% of the projects were in Australia and New Zealand, and 29 were in the Americas, so the averages describe that mix.

Across the sample, the average first reply took 6.4 days and the median reply took 9.7 days. A median is the middle value once the RFIs are sorted by time, so one RFI that waits two months for a structural engineer moves it far less than it moves the average. Projects of one year or less had a median reply of 7 days. Projects of two years or more ran close to 10.

On cost, Navigant’s estimate came from interviews with three US design and construction management firms. They put the typical RFI at about 8 hours to receive, log, review and answer, and Navigant priced that at an average of $1,080 per RFI.

The paper’s case study shows why the RFI log needs categories. A claim on a wastewater treatment plant laboratory cited about 4,000 RFIs. About 800 turned out to be repeats of questions already answered, and about 300 were answered the day they arrived by pointing to a specification section, drawing or detail.

Where an AI tool would show up

Procore’s Draft RFI with AI feature, in beta, shows the two places a tool can act. The user types the question and any cost or schedule impact, and the feature suggests a subject, a question, a cost impact and a schedule impact. A second step searches the project documents for information that could answer the question, which Procore says helps the user see whether the question has already been answered.

Suggested wording saves the project engineer drafting time. The document search can stop some RFIs from reaching the architect. The design team’s answer time sits outside both steps, so a pilot of a drafting tool should be measured on the contractor’s side of the clock. Tools that review drawings before the RFI stage, such as Buildcheck and UpCodes, would show up in the count of RFIs per million dollars of work.

How to measure it on your projects

Track three intervals for every RFI, in calendar days unless your contract counts working days:

Days to draft: from the date the question reached the project engineer to the date the RFI was submitted.
Days to answer: from submission to the date the response came back, the interval Navigant’s clause measures.
Total days: from the question to the answer reaching the field.

Add a category to each RFI. Navigant’s list for an RFI coordinator includes “answered by reference to specification or drawing” and “previously asked and responded to.” Those two categories count the questions a document search might have answered before submission.

A worked example

The figures below are an example with round numbers. A general contractor exports 10 closed RFIs from one job, adds the date each question first came in from the field, and assigns a category. The contract allows 10 calendar days for a response.

Example RFI log, 10 closed RFIs from one job (illustrative figures)
RFI	Category	Days to draft	Days to answer	Total days
001	Design clarification	2	6	8
002	Answered by reference to documents	1	1	2
003	Incomplete plans or specs	3	12	15
004	Design coordination	2	9	11
005	Answered by reference to documents	2	2	4
006	Differing site condition	1	14	15
007	Design clarification	4	7	11
008	Previously asked and answered	1	1	2
009	Constructability	2	10	12
010	Design clarification	3	8	11

Example baseline computed from the log above (illustrative figures)
Measure	Example baseline
Median days to draft	2.0
Median days to answer	7.5
Mean days to answer	7.0
Median total days	11.0
RFIs answered past the 10-day contract time (%)	20
RFIs answered from existing documents (%)	30

Days to answer has a median of 7.5 and a mean of 7.0. Two RFIs, the incomplete-spec question at 12 days and the site condition at 14, ran past the 10-day contract time. Three of the 10, or 30%, were answered from the existing documents, and those three account for 4 of the 21 drafting days and 4 of the 70 answer days.

If a drafting tool cut the median days to draft from 2 to 1, the typical RFI would reach the architect about a day sooner. If its document search caught the three RFIs answered from the documents before they went out, the architect would receive 7 RFIs. Navigant’s estimate of $1,080 per RFI is a 2013 figure from three firms, so price your own review hours before carrying it into a pilot budget.

Answer days depend on the design team, so a change in the architect’s staffing can move them during a pilot. RFI volume follows the job too: Navigant reports that projects of five years or more averaged slightly more than 1,400 RFIs, against 270 for projects of a year or less. Compare months at a similar stage of the work, or use a second job of similar size as a control.

RFIs that are late on activities with little float carry the schedule risk. The total float log shows which activities those are, and Navigant’s best practices ask contractors to submit RFIs at least 10 days before the response is needed.

Start with one active job. Export the closed RFIs from the last three months, add the date each question first came in from the field, assign a category, and compute the median of each interval before anyone turns on the new tool.

SketchDeck 3.0 adds connection material and revision matching to the AI steel takeoff that more than 500 estimators use

Sun, 18 Oct 2026 07:35:03 -0400

SketchDeck.ai released SketchDeck 3.0 on October 15, 2026. It is a rebuilt version of the company’s AI takeoff software for structural steel fabricators and erectors, the tool that reads a PDF drawing set and returns the member counts, weights and bill of materials a steel estimator would otherwise build by hand. Version 3.0 adds connection material, markup tools and bid tracking to the takeoff, and automates the matching of revised sheets to the originals. The company says more than 500 estimators use the software on hospitals, stadiums and data centers.

What it does

Steel takeoff by hand is sheet-by-sheet work. Maccabee Industrial, a fabricator in southwestern Pennsylvania, described its old method in a SketchDeck customer story: the estimator opens the drawing, finds the W8x10, notes that it is 10 feet long, and marks it off.

SketchDeck runs that work in three steps, according to its homepage:

The estimator uploads the PDF set, which can run to hundreds of sheets. SketchDeck sorts the sheets and reads the scale.
A detection model, meaning software trained to find and label one kind of object on a page, here steel members, locates the beams, columns and bracing and sizes and measures them. It marks what it found and what it could not identify. The estimator reviews those marks, fixes members with a click, and adds connections, counts and markups.
The bill of materials exports to Tekla PowerFab, STRUMIS, FabTrol, E.J.E. or Excel. The 3.0 page also shows Bluebeam as an export target.

The release lists what changed in 3.0. With connection material, the estimator selects a group of members, enters the plates and bolts once, and SketchDeck applies them to every member in the group. The release says the new automations also match revised drawings, pull schedules out of tables and identify drawing scales.

What SketchDeck 3.0 includes, by status
Feature	Status	What it does
Automated takeoff	Carried over	Finds beams, columns and bracing on a PDF set, sizes and measures them, and builds a bill of materials
Revision management	Carried over	Compares a revised drawing set with the original and marks each member as added, removed or changed
Connection material	New in 3.0	Applies plates and bolts entered once to every member the estimator selects
Markup tools	New in 3.0	Areas, polylines, counts and annotations, with quantities that update as the estimator counts
Setup automations	New in 3.0	Matches revised sheets to originals, reads scales, classifies sheets and extracts schedules from tables
Section-cut linking	New in 3.0	Opens the section or detail a callout points to without leaving the plan
Bid tracking	New in 3.0	Shows where each bid stands, with assignments, due dates, roles and permissions
3D modeling	In beta	Turns the 2D takeoff into an interactive 3D model of the frame
Column automation	In beta	Follows a column through every level with the column schedule beside the plan
Page stitching	In beta	Shows multi-page drawing sets in one view
Parametric connection material	In beta	Listed by SketchDeck as a beta feature

Source: sketchdeck.ai

Revision management carries over from earlier versions. An estimator loads the new drawings, lines them up with the old set, and sees every member that was added, removed or changed, ready for a change order. SketchDeck’s blog post on the release explains why it is pushing that feature. The company says GCs are selling jobs at earlier design stages to lock in steel prices, so changes come after award and the fabricator absorbs them. It wants the takeoff to become an early advanced bill of materials (ABM, a bill of materials accurate enough to order steel from) so a shop can buy before detailing is finished.

The direct users are estimators and chief estimators at steel fabricators and erectors. A general contractor would see the effect in the fabricator’s response to a revised set, as steel pricing and a change order that lists members added and removed.

Evidence so far

The evidence so far is SketchDeck’s October 15 release, its product pages, and customer stories and testimonials on its website. These claims come from SketchDeck and its customers:

Results SketchDeck and its customers report
Customer	What was measured	Before	With SketchDeck	Who reported it
Steelworks of the Carolinas	Takeoff throughput		More than tripled	Jordan Engelking, estimating manager, in SketchDeck's October 15 release
Maccabee Industrial	Takeoff time on a 600-ton project		75% less	Dawn Hargraves, estimator, in a SketchDeck customer story
Maccabee Industrial	Overall takeoff speed		50% faster	Dawn Hargraves, estimator, in a SketchDeck customer story
SSE (Southern Services & Equipment)	Estimating time on one large project	About 8 hours	1.5 hours	SketchDeck customer story
SSE (Southern Services & Equipment)	Estimating time overall		50% to 80% less	Justin Airhart, COO, in a SketchDeck customer story
King Steel	Estimating time on complex projects	4 days	2 days	SketchDeck customer story
MotionSteel	Estimates completed per month	30 to 40	70	Jay Livesay, general manager, in a SketchDeck customer story
MotionSteel	Team time freed for detailed reviews		40% to 50%	SketchDeck customer story
MSE	Time on beam takeoffs for larger projects		Up to 95% less	MSE's estimate, in a SketchDeck customer story
MSE	Accuracy rate on large projects		95% to 99%	MSE's estimate, in a SketchDeck customer story
MSE	Time saved		About one work week per month	MSE's estimate, in a SketchDeck customer story
S&H Steel	Time for one project takeoff	A day and a half	About 2 hours	Tim Allen, estimator, on SketchDeck's homepage
NAFCO	Members corrected on a drawing with 100 beams		Fewer than 5, a minute or two each	Matt Chance, estimating, on SketchDeck's homepage

Source: sketchdeck.ai

Two of the stories give a before and after in the same unit, which allows a direct comparison.

Reported estimating time before and with SketchDeck
Customer	Task	Unit	Before	With SketchDeck	Cut in percent (computed by Construction Metrics)
SSE (Southern Services & Equipment)	Estimate for one large project	hours	8	1.5	81.3%
King Steel	Estimate for a complex project	days	4	2	50.0%

Source: sketchdeck.ai

The customer stories were posted in December 2025, and several refer to the product under its earlier name, LIFT. They describe the takeoff and revision tools that 3.0 carries over. The connection material, bid tracking and beta features in 3.0 are new, and the stories predate them.

The measures also differ from story to story. MSE’s figures are the team’s own estimates, and its time cut covers beam takeoffs on larger projects. Maccabee’s 75% applies to one 600-ton job, and its 50% to takeoff work overall. SSE’s story adds that the company grew from $8 million to an anticipated $40 million in annual revenue over two years, and links that growth to bidding more work.

Two customers describe how much checking the output still takes. Matt Chance of NAFCO says on SketchDeck’s homepage that on a drawing with 100 beams he adjusts fewer than five, at a minute or two each. Dawn Hargraves at Maccabee said in her company’s story, “I actually appreciate that it’s not 100% perfect because it keeps me engaged and checking the work.” An accuracy rate measured across customers, and error rates for connection material, are not yet public.

For comparison, Togal.AI’s recent takeoff benchmark tested general-purpose AI models on drawing tasks. SketchDeck’s FAQ says its own detection model is built only for structural steel drawings, and that estimators keep the final say on every number.

How to measure it on your projects

Track estimator hours per steel takeoff, recorded with the tonnage and sheet count of each bid. Hours per takeoff is the time from opening the drawing set to a finished bill of materials ready for pricing. Tonnage and sheet count let you compare jobs of different sizes on the same basis.

Capture the baseline before a pilot:

Pull the last ten steel bids. For each, write down the tonnage, the sheet count and the estimator hours from timesheets.
Count bids submitted per estimator per month over the last six months, and the win rate on those bids. Maccabee’s story says its win ratio was 3% to 5%, which is why bid volume mattered to that shop.
For jobs with a revised drawing set, record the calendar days from receiving the revision to sending the change order price.

During the pilot, track a correction rate as well: members the estimator added, deleted or resized per 100 members SketchDeck detected. That number shows how much review the tool needs on your drawings, and it can differ from one drawing set to the next. Run the same bid both ways on at least one job and compare the two bills of materials line by line. General contractors can track the third measure from their side, as days from issuing a revised steel set to receiving the fabricator’s change order pricing.

The same baseline works for other takeoff tools covered here, including Quotr and STACK IQ.

A first step is to pull your last ten steel bids and write down the tonnage, sheet count and estimator hours for each.

Availability and cost

SketchDeck 3.0 is available now to fabricators and erectors. Customers can try the beta features, which are column automation, 3D modeling, page stitching and parametric connection material, by asking during a demo. The company says shops with one estimator and teams with dozens both use the software.

SketchDeck discusses plans and pricing on a 30-minute demo call, where it runs a takeoff live on a drawing set the shop brings and sets up a free trial if the fit is good. Its blog says the rebuilt platform is designed to ship new tools every few weeks. SketchDeck.ai, which its release says was founded in 2021 and is headquartered in Edmonton, Alberta, will show 3.0 at FABTECH in Las Vegas from October 21 to 23, at the CISC Canadian Steel Conference in Vancouver on November 3 and 4, and at Trimble Dimensions in Las Vegas from November 9 to 11.

Autodesk names Diana Colella to lead AEC, the product family that brought in $3.58 billion in fiscal 2026

Sat, 17 Oct 2026 11:50:33 -0400

Autodesk named Diana Colella executive vice president of its architecture, engineering and construction business on October 6, 2026. She succeeds Amy Bunszel, who is retiring and stays through November 2 to support the transition. The job covers Forma, Autodesk’s industry cloud for AEC, and desktop products such as Revit and Civil 3D. AECO, the product family Autodesk reports for this business, brought in $3.58 billion in the fiscal year that ended January 31, 2026, which was 49.7% of company revenue. On October 14, Bricks & Bytes published its panel’s reaction, and the three panelists disagreed on whether an internal pick from Media & Entertainment fits the job.

What Autodesk announced

The release says Colella will advance the Forma strategy while she continues to develop Revit and Civil 3D and connects them more closely to Forma workflows. She keeps Media & Entertainment (M&E) until she hires a new leader for it, who will report to her. Her executive bio carries both titles and lists AutoCAD in her AEC portfolio.

Colella said in the release: “My priority is to listen to them, build on the strong foundation Amy and the team have created, and make sure the work we’re doing across Forma and AI translates into meaningful improvements in how projects are designed, built and operated.”

ENR reports that CEO Andrew Anagnost’s internal email to employees gave her a mandate to speed up the embedding of AI in critical workflows, and called her “an internal disruptor and challenger who is willing to question how we’ve always done things while staying close to customers.”

Who Diana Colella is

Autodesk says she has nearly three decades at the company. She joined through its acquisition of Discreet Logic after working at KPMG, and her roles have covered finance, strategy, sales, operations and product management. She ran the AutoCAD business and launched One AutoCAD. Most recently she led M&E, where Autodesk credits her with moving that business to a connected platform through Autodesk Flow and with the acquisitions of Moxion, PIX and Wonder Dynamics. Bricks & Bytes notes that none of those roles was running an AEC product line.

Revenue by product family

AECO stands for architecture, engineering, construction and operations. Autodesk reports AutoCAD and AutoCAD LT as a separate family, so that revenue is not included in the AECO figure even though the bio places AutoCAD in her portfolio.

Autodesk net revenue by product family, fiscal year ended January 31, 2026. Shares computed by Construction Metrics from Autodesk's figures and may not sum to 100 because of rounding.
Product family	FY2026 revenue ($M)	Share of total	Change from FY2025
AECO	$3,583	49.7%	22%
AutoCAD and AutoCAD LT	$1,787	24.8%	14%
MFG	$1,379	19.1%	16%
M&E	$332	4.6%	5%
Other	$125	1.7%	6%
Total	$7,206	100.0%	18%

Source: investors.autodesk.com

AECO grew 22% from fiscal 2025, the fastest of the four families, and Autodesk’s CFO singled out construction in the fourth quarter. M&E grew 5% and made up 4.6% of revenue, so the business Colella ran most recently is less than a tenth the size of the one she takes over. The figures come from Autodesk’s February 26 earnings release, which labels its financial statements unaudited.

What the Bricks & Bytes panel said

The reaction comes from an episode of the Bricks, Bucks & Bytes show, written up by Owen Drury. The other voices were Martin Piekarz and Dustin DeVan, CEO of Ediphi, who worked at Autodesk after it bought his previous company, according to the article.

DeVan said he has never met Colella and is sure she is capable. In his view Autodesk already lacks clarity on where AEC is going, and a leader from a small vertical unrelated to construction makes a clear direction harder to set. (M&E was 4.6% of Autodesk’s fiscal 2026 revenue.) Bricks & Bytes quotes him: “If you don’t have a vision for what AEC should be in five years, you can’t efficiently manage your resources.” DeVan also said he doubts a full outside search took place, although Autodesk said it would run one when Bunszel’s retirement was announced. The article presents this as his view as a former insider and says Autodesk has neither confirmed nor denied it publicly.

Piekarz argued that running a product business takes experience building product, and said he could not see the connection between creative-industry software and the AEC roadmap. Drury took the other side. A strong executive can adapt to a new industry, Drury said, and Colella may have seen something work in M&E that she believes can be repeated in AEC.

How to track this on your Autodesk account

Bricks & Bytes lists four things to watch over the next two quarters:

Who Autodesk hires to run M&E, and whether that person reports to Colella.
Whether the first large Forma announcement under Colella keeps investment in Revit and the other desktop products. Our earlier post covers Autodesk’s plan for Revit and Forma.
Whether senior AEC product leaders leave in the next six months.
Pricing.

Pricing is the item that shows up on a contractor’s invoice. Autodesk already sells token-based access through Flex, where you prepay tokens and spend them at a daily rate per product, and through enterprise business agreements, which give token access to a pool of products over a contract term. At Autodesk University in September, Autodesk previewed a next-generation Autodesk Assistant and said its availability, rollout timing and subscription terms will be announced in 2027. Bricks & Bytes has reported a prediction from Buro Happold’s CTO that AI billing will shift from seats toward compute.

So the number to set up now is your annual Autodesk cost per active user, in three buckets. The first two follow the Design and Make categories in Autodesk’s revenue reporting.

Desktop subscriptions, such as the Architecture, Engineering and Construction Collection, Revit and AutoCAD.
Cloud construction products, such as Autodesk Build and BuildingConnected.
Token spending through Flex or an enterprise business agreement.

Add up 12 months of Autodesk invoices in each bucket and divide by the number of people who used those products in that period. If the 2027 terms move AI features into token billing, you will have a per-user baseline to price them against. A first step is to get the list of assigned users by product from whoever manages your Autodesk account and mark the ones who have not used their seat this year.

Punch raises $2 million for spoken daily reports that one drywall superintendent says take 3 minutes instead of 30

Sat, 17 Oct 2026 07:34:45 -0400

Punch, an Atlanta construction software company founded in 2025, announced on October 15, 2026, that it raised $2 million from returning investor Gutter Capital. Its product targets the superintendent’s end-of-day paperwork. A super talks to the Punch phone app during the day, and Punch assembles the daily report, flags delays and turns material requests into purchase orders that follow the contractor’s approval rules.

What it does

Today the daily report usually gets written last. The superintendent walks the job, remembers what happened, and types it into a form in the truck or the trailer after a long day.

Punch moves that capture into the walk itself. The super speaks short updates and takes photos as things happen, or forwards an email. According to Punch’s product page, each capture files into the day’s record under labor, deliveries, delays or issues, and the app asks a follow-up question when something is unclear. By the end of the day the report is drafted. The super reviews it and sends it as a PDF on the company’s template, with the photos and AI-written captions inside.

The same updates feed other work. Punch says a spoken material request becomes a purchase order priced from the contractor’s supplier list and checked against the budget, and a scope question becomes a task assigned to the project manager. The rules come from a company “playbook,” Punch’s term for approval limits, preferred vendors and required fields that the contractor writes down once. One example on its homepage routes any change order over $10,000 to the PM before it reaches the client.

Punch sits between the estimating and accounting systems and sends cost data on to the accounting side, which stays the system of record. Its website names Excel, PlanSwift and On-Screen Takeoff on the estimating side, and QuickBooks, Sage, Foundation and Trimble on the accounting side. The company says it is built for contractors doing $5 million to $1 billion a year.

Evidence so far

The evidence so far is Punch’s October 15 release and the customer pages on its website. Most of the measured figures come from one person, Dustin Gibbs, general superintendent at Piedmont Drywall in Atlanta. These claims come from Punch and its customers:

Results Punch and its customers report
Claim	Figure	Who reported it
Time to complete one daily report	30 minutes in the previous field app, 3 minutes in Punch	Dustin Gibbs, general superintendent at Piedmont Drywall, in Punch's release
Time saved per day across daily reporting and material orders (the release also lists payroll)	About 2.5 hours	Gibbs, in Punch's release and customer story
Projects behind those daily reports	6	Punch's daily reports product page
Reduction in time spent on paperwork	About 90%	Punch, on its website
Expected annual savings at one drywall subcontractor	$178,000, a projection	Punch case study of a business case the subcontractor mapped with Punch's team

Source: businesswire.com

The 30-minute baseline is Gibbs’s time in the field app his company used before, and Punch’s customer story says the times were reported by the customer. In that story, one spoken walk covered three buildings across five floors, with photos, flagged delays and a scope question that went to the PM as a task. Gibbs describes the change this way in the release: “I’m saving that time by not having to stop and fill out forms and search through emails all day.”

The $178,000 figure measures something different. It is a projection a drywall subcontractor built with Punch’s team, and most of it comes from expected savings on wrong material orders, labor sub overbilling and problems caught earlier. Only $13,000 of it is the cost of the daily report app the sub expected to drop.

How often supers edit the drafted report before sending it, and results from Punch’s other listed customers, are not yet public. Punch’s homepage shows logos for Rado Mechanical Group, Atlas Demolition and KR-Witwer alongside Piedmont Drywall.

How to measure it on your projects

Track superintendent minutes per daily report, along with the share of reports filed the same day. Minutes per report is the time from opening the form, or the app, to sending the finished report. Same-day filing matters because the daily report is the record a contractor relies on in a delay claim or a backcharge dispute.

Capture the baseline before a pilot:

For two weeks, have each participating super note the start and end time of every daily report.
Count reports filed after midnight on the day they cover, and divide by the total.
Pull ten reports at random and count the entries for crew headcount, deliveries, delays and photos, so you can compare completeness later.

During the pilot, log every report where the super changed the drafted content before sending. That edit rate tells you how much review the tool still needs on your jobs. The labor productivity method built from daily logs and timecards is a second check, since better crew counts in the report make that calculation more reliable. ServiceTitan is piloting a daily log agent for commercial contractors as well, so the same baseline works for comparing tools.

A first step is to time the next five daily reports on one job and write down the minutes.

Availability and cost

Punch is available now. It prices by proposal, with no fixed tiers or per-seat fee, and its pricing page says each proposal targets a return of five to ten times the software’s cost. Punch says a team can go live on one workflow, such as daily reports, in less than 48 hours, and that a broader rollout across purchase orders, invoices, subcontracts and scheduling takes weeks. It plans to spend the new funding on its engineering and customer teams.

Procore opens Compass Q Score prequalification to every general contractor and subcontractor on its platform

Fri, 16 Oct 2026 07:38:30 -0400

Procore has named Compass, the subcontractor prequalification platform from Toronto-based Bespoke Metrics, its preferred prequalification tool, Bespoke Metrics announced on October 15, 2026. The program, called Prequal for All, makes Compass prequalification available to every general contractor and subcontractor working in Procore, starting that day. It touches the prequalification packet a risk manager reviews before a trade partner can bid, and the bid list a preconstruction team builds in Procore’s Bidding tool.

What it does

Today a subcontractor fills out a separate prequalification form for each general contractor it works with, and someone in the GC’s risk or preconstruction group reads the financial statements, EMR letter and insurance certificates in each one. Compass replaces those forms with one standardized questionnaire, the 1Form. The subcontractor completes it once, keeps it current and chooses which general contractors can see it. Bespoke Metrics says more than 150,000 subcontractors already have a profile.

Compass checks each submission in two passes. Its product page says machine learning flags possible problems in the data first, and then a person checks the supporting documents. Data that could not be verified carries a warning.

From the verified data, Compass produces the Q Score, a rating from Q1 (lowest risk) to Q7 (highest risk), along with a recommended prequalification outcome. Its published methodology weights three sections:

How the Compass Q Score is weighted, per Bespoke Metrics' methodology
Section	Weight in Q Score	Examples of what Compass measures
Business	50%	Employee turnover, revenue diversification, backlog relative to revenue, legal history, years in business
Finance	40%	Liquidity (current ratio, cash ratio), leverage (debt to equity, interest coverage), turnover, profitability
Health and safety	10%	Incident rate, experience modification rate, recordable incident rate, safety policy, dedicated safety lead

Source: compass.bespokemetrics.com

Each ratio is graded strong, average or weak against a fixed threshold, and the same thresholds apply to every subcontractor regardless of trade, region or size. Severe findings, such as an active insolvency or financial statements older than 24 months, cap the score however strong the ratios are. Missing, expired or unverified data counts as weak, and Compass treats financials as expired six months after fiscal year end. The health and safety inputs include the experience modification rate and the recordable incident rate.

The Compass Q Scale
Q Score	Color	Risk level	What Compass's methodology says it calls for
Q1	Green	Very low	Minimal review
Q2 to Q3	Green	Low	Minor issues with above-average overall performance
Q4 to Q5	Yellow	Moderate	Further evaluation or risk mitigation may be required
Q6 to Q7	Red	High	Additional due diligence or information before selection

Source: compass.bespokemetrics.com

The work splits between two groups. The GC’s risk team sets the standards and reviews exceptions inside Compass. Project teams see the Q Score and the prequalification result in Procore, where they invite bidders and level bids. The subcontractor’s financial statements stay in Compass, and only the score and the outcome reach Procore.

The connection runs over the Model Context Protocol (MCP), an open standard that lets AI software pull data from outside systems. Bespoke Metrics says a Compass Risk Agent, which will show project teams where every subcontractor stands from bid to award, is expected later this year.

Evidence so far

The evidence is Bespoke Metrics’ announcement, a statement from Procore inside it, ENR’s report and Compass’s own published methodology. These claims come from Bespoke Metrics:

Compass connects more than 400 general contractors and more than 150,000 subcontractors in 33 countries, representing more than $1 trillion in annual construction volume.
The 1Form meets the data collection standards of more than 70 general contractors across North America.
The 1Form satisfies more than 92% of surety applications.

ENR reports that Compass has been available on Procore for more than two years. What changes now is its reach to every Procore customer and its status as the default provider. Procore’s Geoff Lewis, senior vice president of product, said the partnership will “replace passive, static forms with active, modern risk intelligence.”

The methodology says Compass backtests the Q Score by comparing past scores with what happened later. Default or failure rates by Q Score band, the figure that would show how well the scale separates strong subcontractors from weak ones, are not yet public.

How to measure it on your projects

The metric to track is prequalification cycle time: calendar days from the day you send a subcontractor the prequal request to the day your risk team records an approve, approve-with-conditions or decline decision. When a review runs past the bid date, a subcontractor can end up on the bid list before anyone has read its financials.

Capture the baseline before you switch:

Pull the last 20 to 30 prequalification files your risk team closed and record the request date and decision date for each.
Count how many of those needed a second request for missing documents, such as an expired insurance certificate or last year’s financials.
On two or three recent bid packages, count the bidders who were invited before their prequalification was complete.

During a pilot, record the same three numbers, split between subcontractors who already have a Compass profile and those who have to create one. Log every case where your reviewer’s decision differs from Compass’s recommended outcome, with the reason. That log shows how closely the Q Score matches your own risk tolerance. The marketplace listing says a risk manager can also build a custom risk model with Compass to fit the company’s thresholds.

A first step on one job is to export the bidder list from your next bid package and check how many of those companies already have a Q Score in the Procore Construction Network.

Availability and cost

The announcement says Prequal for All is available to Procore customers now. Procore’s support guide for the existing Bidding integration lists it for the Enhanced Bid Management experience in the United States and Canada, with an active Procore Bid Management subscription required. The fastest match uses each company’s EIN, and Compass syncs to Procore twice a day, so new data can take up to 12 hours to appear. The guide says companies without a Compass account can still see the Q Score for companies listed in the Procore Construction Network.

The Procore Marketplace listing says the integration requires a Compass subscription, that a free trial is available, and that Bespoke Metrics prices by quote. Pricing specific to Prequal for All could not be verified.

Relantic's RealityBench finds the best AI model reads a tape measure in a solar install photo within 1/8 inch 36% of the time

Thu, 15 Oct 2026 07:38:43 -0400

Relantic, a company in Columbus, Ohio, posted RealityBench to arXiv on October 9, 2026, and revised it on October 14. It is a benchmark, a fixed test scored the same way for every model, that checks whether vision-language models (AI models that answer questions about images) can review the photos crews take to prove work was done. Seven of its eleven scenes come from rooftop solar installation and four from vehicle repair intake. Across 26 models, the best read the tape measure at a solar module frame within 1/8 inch on 36% of photos.

What it does

Each item is one photo and one prompt. The model first judges acceptability, meaning whether the photo meets the stated requirement well enough to serve as proof of work. It then answers a documentation question about what the photo shows, such as the number on the tape or whether a lag bolt is present.

The photos are rendered in Blender, a 3D modeling program, through a simulated phone camera, and 30% of frames have missed focus or hand shake. Because every scene is built, the true answer is always known. The solar scenes follow an install in order:

A close-up of one L-foot showing its flashing and lag bolt.
The low point of the wire under the modules.
A tape measure standing from the roof to the module frame.
A digital angle gauge on the module.
Equipment nameplates, the skyline from the ridge, and the panel’s circuit directory card.

That list matches the photo checklist a solar or electrical subcontractor’s QA reviewer works through, and the closeout photos a GC’s project engineer collects.

Evidence so far

The evidence is the authors’ paper and their leaderboard on Hugging Face. Relantic built the scenes, wrote the scoring rules, ran the models at list API prices, and supplied the human raters (four authors and two colleagues). Results from an independent rerun are not yet public. The paper states that the benchmark measures third-party models and no Relantic product.

RealityBench leaderboard, selected rows. Paper runs were September 13 to 23, 2026; October rows were added after the paper.
Model	Run	Acceptability F1	Documentation score	Cost for all 1378 items (USD)
GPT-6 Astra	Paper	0.61	0.83	30.05
Gemini 4 Argon	October	0.60	0.86	4.96
GPT-6.1 Sol	October	0.59	0.82	6.22
Claude Opus 5.5	Paper	0.58	0.83	16.36
GLM-5.3 Flash	Paper	0.57	0.75	1.30
Gemini 3.8 Flash	Paper	0.54	0.84	7.16
Gemma 4 31B	Paper	0.52	0.79	0.70
GPT-6 Luna	Paper	0.43	0.62	0.69
Single human rater	Paper	0.59
Always answer "adequate"	Paper	0.33

Source: huggingface.co

The acceptability column uses F1, a score from 0 to 1 that penalizes both bad photos a model accepts and good photos it rejects. Answering “adequate” every time scores 0.33. The best models land near 0.6, about where a single non-author rater scored (0.59) on a separate set of rated frames. Raters disagree with each other, too. An individual rater matched the others’ majority on 76% to 92% of frames.

Documentation splits by skill. The authors report that eight of ten fields asking a model to read text, such as nameplates and directory cards, scored 0.90 to 1.00 averaged across models. Questions about small parts, spatial relations and measurements scored lower.

Share answered correctly, solar questions. Tape leaning is scored on all tape photos, the rest on acceptable photos.
Question about a solar install photo	Mean of 26 models (%)	Best model (%)
Is the lag bolt present	67%	82%
Does any wire touch the roof	57%	74%
Is the tape leaning	55%	82%
True module-to-roof gap within 1/4 in	23%	48%
Tape reading at the module frame within 1/8 in	11%	36%

Source: arxiv.org

On every acceptable tape frame the tape stood within 5 degrees of upright, yet the models called it tilted 75% of the time. In that task, 97% of the models’ wrong rejections came from misreading the scene, while the models applied the requirement consistently to what they reported seeing.

The authors also tested real photos, including 133 tape readings on 84 photos of a tape held against everyday objects, which they released. GPT-6 Astra read 44% of those within 1/8 inch and Claude Opus 5.5 read 35%. Model rankings carried over from rendered to real photos on all three questions tested, with Spearman rank correlations (a measure of how closely two rankings agree, where 1 means identical) of 0.63 to 0.79.

The paper priced one human pass over the 1,378 photos at $11 to $165, using $10 an hour as a low bound and $35.91 an hour from BLS wage data for construction and building inspectors as the high end. Claude Opus 5.5 cost $16.36 for the same set.

How to measure it on your projects

Anyway, your own install photos are the better test. If you pilot a model as a first-pass photo reviewer, track the photo retake rate: the share of required photos in a job’s closeout packet that the reviewer sends back for a reshoot.

Capture the baseline before the pilot:

Pull the photo packets from three to five recent jobs and count required photos, photos sent back, and reviewer minutes per packet.
Run the same packets through the model and record its verdict on each photo.
Count photos the model accepted that your reviewer rejected, and the reverse, as separate numbers.

Leave tape readings and gauge values with the reviewer during the pilot, since measurement scored lowest. A first step on one job is to take one closeout packet from a recent solar or electrical job and compare the model’s accept and reject calls with your reviewer’s, photo by photo. ConSynth-X is another test built on synthetic jobsite images.

Availability and cost

The rendered benchmark is a free download on Hugging Face after you agree to share contact information. The frames, ground truth and model results carry a CC BY 4.0 license and the scoring code an MIT license, and the scripts run against any OpenAI-compatible endpoint. Across the 28 priced models on the leaderboard, one run of the full set cost from $0.56 to $45.99 at list prices. The simulated scenes themselves, which can generate new tasks, are available by contacting Relantic.

Blueprints AI launches V2, which drafts architectural, structural and MEP sheets from a sketch or a written description

Wed, 14 Oct 2026 07:39:21 -0400

Blueprints AI, a Santa Monica startup, launched the second version of its drafting software at TechCrunch Disrupt in San Francisco, which runs October 13 to 15, 2026. V2 takes a sketch, a PDF or a written description of a building and produces a drawing set with architectural, structural, mechanical, electrical, plumbing and interior sheets, according to the company and a TechCrunch interview with its founders. The work it touches is construction document production, the drafting between an approved schematic design and the set that goes to the building department for permit.

Founder Al Dram runs an architecture firm, Drams Design-Build, and added a construction business in 2021 after earning a general contractor’s license, TechCrunch reports. The company says it trained its models on eight years of approved drawing sets from that firm.

What it does

On most projects a drafter or junior architect turns the designer’s sketches into sheets in CAD or BIM, and a licensed architect or engineer reviews and seals them. On a design-build job, the contractor’s design manager coordinates that work and waits on it before permit. Blueprints AI says its tool does the drafting step. The steps below follow the company’s own description.

How a project runs in Blueprints AI, as the company describes it
Step	What happens in Blueprints AI
1. Onboarding	About two weeks. The firm uploads its standards and past projects, and Blueprints AI fine-tunes the model on them
2. Start a project	Upload hand sketches, site photos, PDFs of existing drawings, point cloud scans or a plain-text description
3. Generate	The tool drafts the sheets and applies IBC, IFC, NFPA, ADA and local zoning rules, citing the sections it used
4. Revise	Ask for a change in plain language, upload redlines, or export to CAD or BIM and edit the files there
5. Review and seal	A licensed architect or engineer reviews the set and seals it where the jurisdiction requires, then submits for permit

Source: blueprints-ai.com

Fine-tuning, in step 1, means further training the model on one firm’s past drawings so the output uses that firm’s title blocks, sheet layouts and detail library.

The main change in V2 is coordination across disciplines. TechCrunch reports that one design change now updates the whole project and is checked against local building codes. The first version produced architectural documents only and did not allow editing inside the tool. The company’s site also lists Title 24 energy compliance and fire and life safety sheets, with civil and landscape sheets on Enterprise plans.

Blueprints AI lists single-family, multifamily, ADU, commercial and light industrial work, and says repeatable project types that resemble a firm’s past work benefit most. It recommends settling planning and zoning decisions during design, because those codes vary widely and change often.

Evidence so far

The evidence so far is the company’s product description, its founders’ interview with TechCrunch, and customer comments on its website. These claims come from Blueprints AI:

Manual construction document production takes one to four months. With the tool, the company says, a user spends hours.
Licensed professionals spend about 70% of their time on drafting, a figure the company attributes to the AIA 2023 Firm Survey. Dram told TechCrunch the share is about 80%.
Delivery is up to 10 times faster. The company calls its times illustrative and says they depend on the plan, the project type and its complexity.
More than 4,000 companies have signed up.
Early customers were using V2 before the public launch, Dram told TechCrunch.

Manual drafting compared with Blueprints AI, as stated on the company's website
Measure	Manual drafting	With Blueprints AI
Time spent producing a construction document set	1 to 4 months	Hours, including chat, markup and revisions
Share of professional time spent drafting	70%	15%
Share of time on design and business development	20%	75%
Share of time on admin and operations	10%	10%

Source: blueprints-ai.com

On the website, the president of Beaumont & Associates in Emeryville, California, says the tool freed significant hours from documentation. An unnamed construction company owner in Manitoba says the main bottleneck is getting from a concept to drawings ready to stamp. Drafting hours per set measured on customer projects, and the number of plan-check corrections on sets the tool drafted, are not yet public.

How to measure it on your projects

The metric to track is drafting hours per permit set. It is the total hours that drafters, designers and the reviewing architect or engineer log from an approved schematic design to the first permit submission. Count the licensed reviewer’s hours, because a tool like this moves time from drafting into review.

Capture the baseline from projects you have already permitted:

Pull timesheets for three to five recent projects of the type the tool targets, such as single-family homes or small multifamily buildings.
Record drafting hours, review hours, the sheet count, and calendar days from schematic approval to first submission.
Count the comments in the building department’s first review letter, so a faster set that draws more corrections shows up. Our post on UpCodes AI Plan Review explains how to count them by category.

Blueprints AI does not offer a free trial, only a live walkthrough. A first step on one job is to bring the sketch from a project your city has already permitted to that walkthrough, ask the team to generate the set, and have your architect mark up the output against the approved set. Record the redlines per sheet.

Availability and cost

The company says V2 is live, and it is offering a Founding 100 group of firms priority access with 30% more credits and two extra seats. TechCrunch reports that the first-generation software costs $3,200 to $8,500 a month depending on plan, with custom pricing for enterprises, and that the company may add a one-time fee for single projects.

The company’s pricing page lists monthly and yearly plans, with yearly billing saving up to 20%, and asks buyers to contact sales for prices. Enterprise engagements include an on-premise deployment option. The company says most professional liability policies cover AI-assisted work once a licensed architect or engineer seals the drawings, and tells buyers to confirm with their carrier.

UpCodes says its AI Plan Review flagged 180,000 potential code issues across 1,300 plan sets

Tue, 13 Oct 2026 07:40:48 -0400

UpCodes, an Austin company that sells building code research software, said on October 13, 2026, that its AI Plan Review feature has checked more than 1,300 construction plan sets since it launched in June and flagged more than 180,000 potential issues. Plan Review reads an uploaded drawing set and compares it with the codes and amendments adopted where the project will be built. UpCodes named Dallas-Fort Worth Airport, BDG Architects and Hereford Dooley Architects among the organizations now using it.

The workflow it touches is the code check on a drawing set before it goes to the building department for permit. On a contractor’s side, that is the design manager on a design-build job or the preconstruction team reviewing a set before bid.

What it does

UpCodes describes most QA/QC today as manual review, often done by one person working through the sheets against the adopted code. Plan Review takes the same set and returns a list of issues. UpCodes’ June launch release describes the steps in the order a user takes them.

How a Plan Review run works, as UpCodes describes it
Step	What happens in Plan Review
1. Set up the project	Enter the jurisdiction, code year and building type once
2. Upload drawings	The drawing set goes into the UpCodes QA/QC platform
3. Pick the analysis	A general code review, or a checklist review built on the firm's own QA/QC template
4. Pick disciplines	Optional runs for architecture, structural, MEP, fire protection, life safety, and accessibility and energy
5. Read the issues	Issues sorted by severity and category, each linked to the drawing page and the governing code section, with plan and spec discrepancies flagged
6. Work the list	Filter, comment on, assign and track issues, and ask the Copilot assistant for a code interpretation
7. Rerun and export	Re-upload revised drawings to run the check again, and export the report for consultants, clients or government officials

Source: prnewswire.com

The October release says Plan Review runs on multimodal AI, meaning a model that reads the drawing images along with the text on them. That lets it read a dimension string or a door location on a floor plan together with the notes. UpCodes says each run checks against 11 million sections of locally adopted codes and amendments, with amendments from more than 6,000 jurisdictions.

Each issue in the output carries a severity, a category, the sheet where it appears and the code section behind it. The tool also flags places where the plans and the specifications disagree, the kind of conflict a project engineer would otherwise write up as an RFI once the job is in the field.

UpCodes calls Plan Review a first-pass analysis and says the tool does not replace review by a licensed design professional or approval by the authority having jurisdiction. Someone with code knowledge still has to read each flag and decide whether it is real. A similar tool from Buildcheck, which raised $12 million in September, has its own staff check each flag before a customer sees it.

Evidence so far

The evidence is UpCodes’ own count of every plan set run through the tool since the June launch. The three new users the release names are an airport owner and two architecture firms. These figures come from UpCodes’ October 13 announcement:

UpCodes Plan Review results, June to October 2026, as reported by UpCodes
Measure	Value	How UpCodes states it
Plan sets reviewed since the June 2026 launch	1,300	more than 1,300
Potential issues flagged	180,000	more than 180,000
Jurisdictions covered by the plan sets	73	73
Code editions covered by the plan sets	14	14
Critical issues per plan set	21	average
Share of critical issues involving life-safety systems (%)	46	nearly half (46%)
Share of critical issues that were cross-discipline coordination conflicts (%)	10	roughly one in ten
Estimated potential construction error cost per project (USD)	1,000,000	more than $1 million on average, estimated

Source: prnewswire.com

The plan sets ranged from single-family homes to commercial and medical buildings. Missing exit signage was the most common critical finding. The life-safety group covers means of egress, fire protection and fire-resistance-rated construction. The coordination conflicts are clashes between the architectural, structural and MEP sheets.

The release gives totals and averages, and two more figures follow from them.

Dividing 180,000 flagged issues by 1,300 plan sets gives about 138 issues per set. With 21 of those rated critical, about 15% of what the tool flags is critical. Both inputs are “more than” figures, so treat the ratio as approximate.
Applying the 46% life-safety share to 21 critical issues gives about 10 critical life-safety findings per set. About two per set were coordination conflicts between disciplines, and the two groups can overlap.

Issues per plan set, derived from UpCodes' reported averages
Item	Per plan set	How it was derived
All flagged issues	138.0	180,000 issues divided by 1,300 plan sets (both are lower bounds from UpCodes, so approximate)
Critical issues	21.0	Average reported by UpCodes
Life-safety critical issues	9.7	46% of 21 critical issues
Cross-discipline coordination conflicts	2.1	Roughly one in ten of 21 critical issues

Source: prnewswire.com

The $1 million figure is an estimate. UpCodes built it from the average issue counts and its estimated cost to resolve comparable issues during design, permitting and construction. The share of flags that a reviewer confirmed as real code issues, and the time a reviewer spent working the list, are not yet public.

How to measure it on your projects

The metric to track is permit review comments per submission. It is the number of comments in the building department’s plan review letter for one permit submission, counted by category such as egress, fire protection, accessibility, structural, MEP and energy. A tool that catches code issues before submission should lower that count and the number of resubmission rounds.

Capture the baseline from projects you have already permitted:

Pull the plan review comment letters for three to five recent projects in the same jurisdiction and of a similar building type.
Count the comments in each letter by category, and record the number of review rounds and the calendar days from first submission to permit issue.
Note which comments turned into RFIs or change orders after construction started.

During a pilot, log each flag the tool raises and whether your reviewer accepts it. The share accepted is the tool’s precision, the portion of its flags that are real code issues. A low share means your reviewer spends more of the pilot clearing flags that turn out to be fine. Then, when the building department’s letter arrives, count how many of its comments the tool had already flagged. That share is the tool’s recall, the portion of real issues it caught.

A first step on one job is a backtest. Take a set that has already been through permit, with the comment letter in hand, ask the vendor to run it, and count how many of the building department’s comments show up in the tool’s list.

Availability and cost

UpCodes’ pricing page lists AI Plan Review in early access on the Enterprise plan only, priced through its sales team. The Enterprise plan also lists bulk discounts for teams of 25 or more.

UpCodes plans and list prices, October 13, 2026
Plan	Per user per month billed annually (USD)	Per user per month billed monthly (USD)	Includes AI Plan Review
Essentials	39	45	No
Professional	59	68	No
Enterprise			Yes, in early access, priced by sales

Source: up.codes

The Professional plan adds UpCodes’ Copilot assistant, which answers code questions limited to a chosen jurisdiction and cites the code sections it used. Essentials and Professional come with a two-week free trial. The pricing page does not list a trial for Plan Review.

Track3D prices AI progress tracking at $0.15 per square foot and adds site photo documentation at no extra cost

Mon, 12 Oct 2026 07:41:47 -0400

Track3D, a Milpitas, California, company that sells AI progress tracking to contractors, changed how it charges on October 6, 2026. Contractors now pay $0.15 per square foot per project for progress tracking, and Track3D includes visual documentation, the searchable photo record of site conditions, at no additional cost. The workflow it targets is the weekly progress check that feeds the schedule update and the percent-complete figures on a pay application.

What it does

Today a superintendent or field engineer verifies progress by walking each floor with each trade, marking up drawings and comparing notes with foremen. In Track3D’s case study of a Hensel Phelps project at San Francisco International Airport, a mechanical scope walk took 30 to 60 minutes, and the team repeated it for framing, electrical, finishes and the other trades.

With Track3D, a field engineer captures the site at the end of the workday. On the Hensel Phelps job, those captures came from the field engineers’ routine 360-degree site walks. The software runs image segmentation on each frame, which means it labels the parts of every photo by what they show, such as installed ductwork or framed walls. It then maps what it finds to the scope of work and reports quantified progress by level, trade and area, compared against the schedule. Track3D says the output shows what got built, which activities are slipping and which areas have stalled.

Track3D says it works from standard 2D drawings and uses a BIM model when one exists. Visual captures are available in as fast as 30 minutes and progress data in as fast as six hours, according to the company.

The people who would use it are superintendents and field engineers who run progress walks, project managers who check percent complete before approving a sub’s pay application, and schedulers updating the look-ahead. In the case study, the Hensel Phelps QC team also used the photo history to check in-wall work without opening walls.

Evidence so far

The evidence so far is Track3D’s own case study of the Hensel Phelps SFO project, a six-story office and operations center built between two active terminals, plus comments from a Hensel Phelps superintendent and the company’s director of innovation. Hensel Phelps deployed the software during the project’s second phase.

These claims come from Track3D’s case study:

2,964 total hours saved, which the press release describes as manual progress-tracking work.
$342,000 in labor cost savings.
A reduction of more than 50% in documentation time.
Three major reworks avoided and 26 subcontracted trades coordinated.

Dividing the two headline figures gives about $115 per hour saved. The length of the period behind these totals and the method used to count the hours are not public.

The case study also reports what changed in the superintendent’s week:

Progress verification on the Hensel Phelps SFO project, as reported by Track3D
Task on the SFO project	Before Track3D	With Track3D
Hours per week verifying progress	20	1
Job walks per month	20+	1

Source: track3d.ai

Area superintendent Garin Anderson is quoted in the case study: “We used to spend 20 hours a week just trying to verify progress.” Midway through the job the team added lighting fixtures to what the software tracked, and Track3D says lighting progress was being measured within days. Thai Nguyen, Hensel Phelps’s director of innovation, says in the press release that a six-hour turnaround makes the data early enough to change what happens on site.

Track3D also says 16 of ENR’s Top 30 contractors use the product.

How to measure it on your projects

Track progress verification hours per week. Count the hours your superintendent, field engineers and project engineers spend walking the job to confirm progress, marking up drawings, chasing trade updates and assembling the progress report. Leave out safety walks and QC inspections, so the number covers only the work the tool claims to replace.

Capture the baseline before a pilot:

Add a timesheet code for progress verification on one active job, or have each person keep a simple daily log.
Record four weeks of hours, and note how many job walks happen in that time.
Record how many days pass between a walk and the updated progress report reaching the scheduler.

During the pilot, keep logging the same hours and compare the pay application percent complete for two or three subs against what the software reports. The SPI and CPI post shows how percent complete from pay applications feeds schedule and cost indexes, and the labor productivity post uses the same daily log and timesheet records.

A first step this week is to add the progress-verification timesheet code on one job.

Availability and cost

The new pricing is available now to new and existing customers, subject to Track3D’s terms and eligibility criteria. Progress tracking costs $0.15 per square foot per project. Customers on progress tracking get visual documentation across their projects at no additional cost for up to five times the area they contract for progress tracking. Track3D sells through its contact page.

ServiceTitan opens its Max AI pilot to commercial contractors with daily log and invoice agents

Sun, 11 Oct 2026 07:38:33 -0400

ServiceTitan opened a limited pilot of its Max package to commercial and roofing contractors on October 6, 2026, at Pantheon, its user conference in Orlando. Max is ServiceTitan’s top tier of AI agents, meaning software that carries out tasks inside the system as well as answering questions. Until this week it was sold only to residential in-home trades. ServiceTitan says the agents already working in the field for the pilot draft foremen’s daily logs, review invoices and write up site walk findings, and agents for accounts payable, collections and applications for payment (AFPs) are due later.

What it does

ServiceTitan already has a daily log in its Field Mobile App. Its help documentation says the foreman enters notes, issues, materials, equipment, site safety and weather. An office employee then clicks Generate Log to get an AI summary, edits any field, approves the log and downloads a PDF to send to the GC or owner. After approval the foreman can no longer edit it from the phone.

The Daily Log Agent moves the drafting to the field. At the commercial keynote, ServiceTitan said the agent builds the foreman’s log from voice notes and jobsite photos. When it spots a conflict in the field, it calls in an RFI agent or a change order agent. ServiceTitan’s published materials describe that handoff in a single sentence, so how a conflict is detected and who signs off on the RFI draft could not be verified.

The invoice review agent works after the job closes. In a demo at the residential keynote, it caught a smart thermostat that was installed but never billed and drafted the correction for one-click approval.

Two agents shown at the commercial keynote serve the service side of the business. The Findings Agent reads completed work orders in the office and recommends findings technicians may have missed. The Equipment Agent fills in manufacture dates and links compatible filters, belts and refrigerant type from a unit’s model and serial number.

Agents ServiceTitan lists for the commercial and roofing Max pilot
Agent	What ServiceTitan says it does	Status ServiceTitan gives
Daily Log Agent	Builds the foreman's daily log from voice notes and photos, and calls in an RFI or change order agent when it spots a field conflict	Live in the field
Invoice review agent	Checks invoices against the work done and drafts fixes for approval	Live in the field
Site walk and findings copilot	Writes up findings from a site walk	Live in the field
Findings Agent	Reviews completed work orders and recommends findings technicians may have missed	Shown at the commercial keynote
Equipment Agent	Fills in manufacture dates and links compatible filters, belts and refrigerant type from a unit's model and serial number	Shown at the commercial keynote
Agreement review agent	Reviews service agreements and flags problems with a recommended next step	Coming soon
Touchless AP autopilot	Named in the pilot announcement for accounts payable	Coming soon
Collections agent	Named in the pilot announcement for collections	Coming soon
AFP recommendation advisor	Named in the pilot announcement for applications for payment	Coming soon
Construction estimate variance agent	Forecasts project variance from historical budget and actual data	Planned by next year

Source: servicetitan.com

Evidence so far

The evidence so far is one customer result ServiceTitan presented at its commercial keynote and figures its executives gave on stage. These claims come from ServiceTitan:

At Gulfshore Air and Heating, invoice prep time fell from 30 minutes to under 5 minutes with the Invoice Agent.
Gulfshore now gets paid in under five days, and days to payment are down 70%.
The first time the agreement review agent ran on a live Gallo Mechanical service agreement, it flagged a problem with the first two quarterly invoices.
Max builds on more than 100 new foundational features delivered over the past year.
CEO Ara Mahdessian said the Max customer base will approach 700 by the end of the year.

Gulfshore Air and Heating invoice figures, as ServiceTitan reported them
Measure	Before	With Invoice Agent
Invoice prep time per invoice	30 minutes	Under 5 minutes
Days to payment	Not stated (ServiceTitan says the figure fell 70%)	Under 5 days

Source: servicetitan.com

The number of invoices and the period behind the Gulfshore figures are not yet public, and neither are results for the Daily Log Agent on construction projects. PLOT Daily, which covered the launch, noted that the Gulfshore figures come from ServiceTitan’s own customer example.

How to measure it on your projects

Track billing lag, the calendar days from the date work is complete to the date the invoice reaches the customer. On a service call, start the clock when the tech closes the job. On a progress-billed project, start it at the pay period cutoff and stop it when the pay application goes out.

Capture the baseline before the pilot starts:

Export every job completed in the last 90 days with its completion date and invoice sent date.
Compute the median lag, the middle value, which a few stuck invoices barely move.
Have the billing clerk time five invoice preps from open to send.
Count credit memos and rebills issued in the same 90 days, so you can tell later whether faster invoices also carry more errors.

Run the same report on pilot jobs each month and compare all three figures. A first step you can take this week is to pull that 90-day export into Excel and find the median.

Availability and cost

The pilot is limited, and ServiceTitan’s live blog tells contractors to register for it. Pricing for Max and the pilot is not listed in the announcement or the live blog. Atlas, ServiceTitan’s AI assistant, is in closed alpha for admins, and ServiceTitan says it has not set Atlas pricing and packaging. Max customers get Atlas first.

The existing daily log feature needs account configuration through ServiceTitan support. By next year, ServiceTitan plans to add a construction estimate variance agent for commercial customers, along with roofing agents for insurance supplements, crew and delivery scheduling, and subcontractor work orders and pay. ServiceTitan’s survey of commercial service contractors from last month covers how many of those firms already use AI.

Volve adds an AI layer that checks spec quantities against tender drawings and the IFC model

Sat, 10 Oct 2026 07:38:12 -0400

Volve, a tendering software company whose customers so far are Nordic contractors, has added what it calls an orchestration layer to its platform. AEC Magazine reported the release on October 6, 2026, and Volve describes it as the largest upgrade since the platform launched. The first task it takes on belongs to the bid team’s estimator: checking that the quantities in the specification match the drawings and the BIM model.

What it does

On most bids that check is manual today. A quantity can read one way in the spec, another on the drawings and a third in the model, Volve says, and someone has to open all three to find the difference.

The work in Volve starts with the tender set. The team drags the documents in or pulls them from SharePoint or Procore, and Volve’s homepage says it reads PDFs, Word and Excel files and keeps track of document versions. For the model, the release names IFC, the open file format that most BIM tools can export.

Then the estimator types the task in plain language. AEC Magazine gives Volve’s example request, which asks the software to check that the quantities in the specification match the drawings and the IFC model. The orchestration layer is the part that turns that request into steps: it decides which documents to compare, runs the comparison across the whole tender, and checks its own result before it reports back, according to Volve.

The output is a list of places where the quantities differ, with the source page or model element behind each one. Volve says the same layer can check that scope matches what was estimated and that an addendum has not moved something without anyone noticing.

An estimator or preconstruction manager still reviews each flag. A confirmed mismatch can go into the price, into a bid qualification, or into a question to the owner before bid day. Volve’s argument is that these mismatches tend to surface during construction as change orders and disputes, after the price is fixed, and that catching them in the tender lets the team price them.

The same cross-document problem is the target of Kuro, a Berlin tender-review company covered here on October 2. How well general AI models read drawings is a separate question, and Togal.AI’s takeoff benchmark measured it this week.

Evidence so far

The evidence so far is Volve’s release note, its website, its August funding announcement, and AEC Magazine’s report, which relays Volve’s description of the product. These claims come from Volve and from customer comments it publishes:

Figures Volve and its customers report
Claim	Figure	Who reported it
Projects Volve has been used on	More than 2,500	Volve, in its October release note and its August funding announcement
One public infrastructure tender	213 documents, 20 risk findings, 73 subcontractor bids compared	Volve, in its August funding announcement
Time on early mitigation and optimization	60% more	Volve's homepage
Rework	30% less	Volve's homepage
Bid quality scores	100% from public evaluators across five consecutive wins	A tender manager at an unnamed contractor, on Volve's pricing page

Source: volvetech.com

The figures in the table describe the platform as a whole and predate the orchestration layer. The 213-document tender is the most specific case Volve gives, and the announcement counts what the software surfaced there, which was 20 risk findings.

Volve’s homepage also carries a comment from Christoffer Hernæs, Skanska’s CDO at the time, who says his team saw time saved and better decision support from the first day.

Results for the new quantity check, such as how many real mismatches it finds and how many of its flags turn out to be wrong on review, are not yet public.

How to measure it on your projects

Track quantity reconciliation hours per bid. That is the total time your estimating team spends comparing quantities across the spec, the drawings and the model for one bid, from opening the set to closing the last open line.

Capture the baseline before a pilot:

On your next three to five bids, give quantity reconciliation its own line on the estimators’ timesheets.
For each bid, record the hours and the number of quantity mismatches the team found.
After award, tag any RFI or change order on those jobs that traces back to a quantity mismatch already in the bid set.

During the pilot, log the same hours, including the time spent reviewing the tool’s flags. Count how many flags your estimators confirmed and how many they rejected. The confirmed share is the tool’s precision, meaning the fraction of its flags that pointed to a real mismatch. A first step on one job is to add the timesheet line on the next bid that comes in with a model.

Availability and cost

Volve’s pricing page asks buyers to book a demo, where the company matches a plan and any enterprise add-ons to the buyer’s size and project portfolio. The onboarding page says the first session happens in the first week and runs on a live bid, followed by weekly follow-ups through week four.

Volve raised a USD 3 million seed round in August 2026, led by Skyfall Ventures, and said it was expanding into the UK and continental Europe. Availability for US contractors could not be verified. The homepage shows ISO 27001 and GDPR marks.

Speedchain says its AI Receipt Agent cut one contractor's month-end close time 95%

Fri, 09 Oct 2026 07:38:57 -0400

Speedchain, which sells a commercial credit card and spend platform built for contractors, said on October 7, 2026, that customers using its AI Receipt Agent have cut month-end close time by as much as 95%. The agent handles the receipts behind company card purchases. Those receipts feed job cost, and Speedchain aims the agent at the month-end work of matching them to card charges.

What it does

The work starts at the supply house counter. A foreman or superintendent buys material on a Speedchain card, and the platform sends a text asking for the receipt. The cardholder replies with a photo.

The agent reads the vendor and the amount from the image. Speedchain says it also extracts job cost details, matches the receipt to the card transaction, and suggests the project and cost code. The cardholder confirms the coding with one tap, and the transaction moves to an approver.

Approved transactions sync to the contractor’s accounting system. Speedchain’s website names Sage, QuickBooks, Viewpoint and Acumatica as the systems it works with. A comment from Fite Building Company on the site says the Sage 300 integration lets the company track card spend against its own job cost codes.

Field employees who carry cards answer the texts, and Speedchain says they no longer file traditional expense reports. AP and project accounting staff review the coded transactions and close the month. The card itself carries spend controls set by an administrator, and the website says purchases outside those rules get stopped at the counter.

Evidence so far

The evidence so far is Speedchain’s October 7 press release, which quotes two customers, and the testimonials on Speedchain’s website. These claims come from the release:

Results Speedchain and its customers report for the AI Receipt Agent
Claim	Figure	Who reported it
Time Marcus Construction's accounting team spends on month-end close	Down 95%	Taylor Marcus, CEO of Marcus Construction, in Speedchain's release
Time G&H Contracting's accounting team spends on expense entry	Down 50%	Joseph Shepherd, CFO of G&H Contracting, in Speedchain's release
Projected annual labor savings at G&H Contracting	More than $40,000	Joseph Shepherd, in Speedchain's release
Share of field coding that is inaccurate and needs rework by finance	More than 70%	Speedchain, citing contractors it calls leading

Source: prnewswire.com

The two customer figures measure different tasks. The 95% applies to the month-end close at Marcus Construction, according to its CEO, Taylor Marcus. The 50% applies to expense entry at G&H Contracting, and the $40,000 is a projection of G&H’s combined labor savings for a year, according to its CFO, Joseph Shepherd.

The release gives these reductions as percentages. The hours behind them, the card volume at each company, and the share of receipts the agent codes correctly without a person editing the suggestion are not yet public.

Speedchain attributes the 70% figure to contractors it describes as leading ones. Its website describes the problem in plain terms: when nobody in the field knows the job number, the charge lands in overhead and inflates costs on the next bid. A charge coded to the wrong job also moves cost to date on two jobs at once, and cost to date drives percent complete on a work-in-progress schedule.

How to measure it on your projects

Track accounting minutes per card transaction at month-end. Count the hours AP and project accounting staff spend collecting card receipts, coding or recoding card charges, and reconciling the card statement to the ledger, from the statement date until the books close. Convert to minutes and divide by the number of card transactions in that period, so a busy month and a slow month compare fairly.

Capture the baseline before a pilot:

Have AP and project accounting log card work to a separate time code for two closes.
Count the card transactions in each of those months from the card statement.
Count the transactions accounting moved to a different job or cost code after the field first coded them. Divide by the total to get a recode rate, which lets you test Speedchain’s 70% figure against your own books.
Record the calendar days from the statement date to the close.

During a pilot, keep the same time code and log every agent suggestion that a person changed before approval. The share of suggestions changed is the agent’s error rate on your cost code structure. The gross margin fade report is a second place to watch, since miscoded material shows up there as cost moving between jobs.

A first step is to pull last month’s card statement for one division and count the charges that were recoded after they were first entered.

Availability and cost

Speedchain sells the AI Receipt Agent as part of its card and spend management platform for contractors. Its website lists no prices and starts sales with a demo request. The release says customers also get access to global and industry-specific rewards and discounts through a program called Speedchain Advantage.

Buildcheck raises $12 million for AI drawing review it says has flagged 115,000 issues across 150,000 drawings

Wed, 07 Oct 2026 07:39:57 -0400

Buildcheck, a company whose software reviews construction drawings, has raised a $12 million Series A. CEO Alexander Michalatos announced the round in a letter on the company’s site dated September 9, 2026, and the site’s banner says Telescope Ventures led it. AEC Magazine covered the raise on October 7. The work it targets is design review: the back-check a general contractor’s preconstruction, quality or MEP team runs on each drawing set to catch errors before they turn into RFIs and change orders.

What it does

The input is the drawing set as 2D PDFs. AEC Magazine reports that Buildcheck needs no BIM model, metadata or special export. Its computer vision models (software trained to recognize objects in images, here tags, dimensions, pipes and columns on a sheet) read each drawing for its components, geometry and spatial relationships. Buildcheck says this lets it catch a pipe drawn through a column, which an AI model that reads only the text on a sheet would miss.

AEC Magazine lists four kinds of checks:

Drawing health, such as missing tags and dimensions, sheet index mismatches and broken detail references.
2D clash detection, such as ceiling devices that collide, pipes through columns and slab edges that do not line up.
Omissions and rule checks, such as a pump with no power, an under-rated door in a fire-rated wall, or an electrical panel that appears on the floor plan and differs on the single-line diagram.
Spec-to-drawing checks, where the drawings and the written specification contradict each other.

Michalatos’s letter says the product can also turn a contractor’s own QA/QC playbook into custom review agents, isolate the changes between drawing revisions, identify code issues and point out value-engineering opportunities. The output is a list of comments. Buildcheck says its in-house experts review every AI finding before the customer sees it, confirming each issue and setting its severity.

So the daily users are the people who mark up drawing sets today: preconstruction managers, project engineers, MEP coordinators and quality managers. Owners and architects use it too, according to the customer logos and testimonials on Buildcheck’s website.

Evidence so far

The evidence so far is the CEO’s Series A letter and the claims and customer testimonials on Buildcheck’s website. These are the figures Buildcheck has published:

Figures Buildcheck reports for its drawing review
Claim	Figure	Where Buildcheck states it
Drawings reviewed to date	More than 150,000	Series A letter
Comments flagged to date	More than 115,000	Series A letter
Projects with significant life-safety comments	2 in 3	Series A letter
Paying customers	More than 110	Series A letter
Return on investment	10x to 35x	Website
Relevant issues caught compared with manual review	50% or more	Website
Primex Investments savings	$541,000 and 27 days	Website
Fitzrovia savings	$221,000 and 2 days	Website

Source: buildcheck.ai

The volume figures work out to roughly three comments for every four drawings reviewed. The letter names DPR Construction, EllisDon, Power Construction, IMC Construction and EMJ Construction among the paying customers, with work across data centers, healthcare, commercial and residential projects.

Buildcheck offers the Primex and Fitzrovia case studies to readers who contact it. How the ROI range and the 50% comparison were calculated, and the AI’s false-positive rate (the share of its raw flags that turn out not to be real issues) before expert review, are not yet public.

The customer comments describe the work the software takes on. Kevin Weckman, quality support manager at Power Construction, says on the site: “Every new drawing set, we spend several hours manually back-checking every sheet.” DPR’s Zach Murphy says the software helps DPR’s reviewers spend more of their time on critical issues. An earlier post here covered ARCH-B, an academic benchmark that tests how well AI models read building drawings.

How to measure it on your projects

Track design-error RFIs per 100 sheets. Count the RFIs on a job that trace to an error, omission or conflict already present in the issued drawings, divide by the number of sheets in the set, and multiply by 100. Dividing by sheet count lets a small tenant improvement and a large hospital sit in the same comparison.

Capture the baseline before a pilot:

Pick 5 recently completed or late-stage jobs with full RFI logs.
Tag each RFI by cause, and keep the ones caused by the documents.
Record the sheet count of the set those RFIs were written against.
Record the hours your team spent back-checking each drawing revision, if your timesheets carry a code for it.

In a pilot, run the tool on a new set before it goes out for bid or construction, log which of its comments your team would have missed, and then track the RFI rate on that job as it builds. The change order rate by cause uses the same tagging and pairs well with this metric.

A first step this week is to pull the RFI log on one active job and tag each RFI as design-caused or not.

Availability and cost

Buildcheck works from 2D PDF drawing sets, so a contractor can run it on a set from any design team. Its website starts sales through a scheduled call, and prices are not posted there. The site’s banner says the round will expand four new products, and Michalatos’s letter says the company is hiring engineers, salespeople, researchers and builders.

Punchcard survey finds 64% of Canadian construction workers trust AI to read drawings and 87% get no formal AI support

Tue, 06 Oct 2026 07:40:05 -0400

Punchcard Systems, an Edmonton consultancy that builds software and AI tools, released The State of AI at Work in Canada on October 6, 2026. Angus Reid Group ran the survey for Punchcard from September 2 to 11, polling 1,100 working Canadians and recruiting extra respondents in heavy industries, construction among them. Punchcard reports that 64% of construction workers trust AI to read technical drawings and 63% trust it to interpret Canadian codes and standards, provided they can review what it produces, while 87% have received no formal AI support from their employer. The tasks in the survey run from reports and job paperwork up to drawing review and decisions that affect worker safety.

What it does

The survey asks workers how they use AI and which tasks they would hand it. It tests no tool. For each task, respondents could say they trust AI completely or trust it with a person checking the output, and the report counts the two answers separately. The public landing page shows five of the eight heavy-industry tasks in the full report.

Heavy-industry AI users who trust AI with each task. Total is the sum of the two published shares.
Task	Trust completely (%)	Trust with review (%)	Total (%)
Routine procedures, product specs or how-to information	9%	68%	77%
Reports, logs or job paperwork	5%	66%	71%
Technical drawings, manuals or spec sheets	4%	49%	53%
Safety, compliance or code requirements	3%	49%	52%
Decisions affecting worker safety or physical risk	2%	27%	29%

Source: punchcard.io

Construction comes out ahead of the heavy-industry average on the technical work. Across heavy industry, 53% of AI users trust AI with technical drawings, manuals or spec sheets when both answers are counted, against Punchcard’s 64% for construction on drawings. Only 4% of heavy-industry users would take a drawing answer from AI without checking it. The release says 36% of construction workers would trust AI on decisions that affect worker safety with review, and none would trust it completely there.

The figures are most useful to an operations lead or CFO writing an AI policy for field and office staff. Punchcard counts AI training, approved tools, written rules and a named person to ask for help as formal support. Construction had the highest share of workers with none of it. (Only 6% of AI users say their organization has given them formal responsibility for AI governance or safe use.)

Workers who report no formal AI support from their employer
Group	No formal AI support (%)
Construction	87%
Agriculture	77%
Heavy industry overall	68%
Manufacturing	67%
Transportation and logistics	65%
Other industries	54%

Source: businesswire.com

Punchcard says 52% of all Canadian workers in the sample use AI for work, 49% in heavy industry and 57% in the service sector. In construction, 31% of AI users say it has affected their role, the lowest of the five heavy industries on the landing page (agriculture is next at 32%).

Evidence so far

The evidence is one survey of self-reported answers, commissioned by a firm that sells AI training and custom AI tools. Respondents were members of the Angus Reid Forum panel. Punchcard gives a margin of error of plus or minus 3.0 percentage points, and labels it as what a probability sample of 1,100 would carry, for comparison only.

The construction figures rest on a smaller group. The number of construction respondents is not on the public pages, so a reader cannot compute the margin on the 87% or the 64%, though it is wider than the full-sample figure. The bases also differ between the two documents. The release says “construction workers,” while the landing page labels its trust chart as answers from AI users. The likeliest reading is that the 64% and 63% describe construction respondents who use AI, and the full report would settle it.

So, the time and error figures cover all industries together:

Of AI users who answered the question, 65% say AI saves them time, at an average of 6.1 hours a week, and 35% report no time saved.
In the release, 94% say they rework AI output before they can use it, and 19% say reviewing AI work is a regular part of the job.
Among AI users, 24% say they caught an AI error before it caused trouble, and 18% say AI output did cause a problem at work, including 3% who report a serious one such as a safety incident or a financial loss.

How to measure it on your projects

Reports, logs and job paperwork rank second on the trust list at 71%, which makes the daily report a reasonable first pilot. Track the correction rate, the share of AI-drafted daily reports that the superintendent or project manager changes for a factual error, such as a wrong crew count, quantity or location, before the report goes out. Pair it with minutes per report, since 94% of AI users in the survey rework output before they use it.

Capture the baseline before the tool goes live:

For two weeks on one job, have the superintendent log the minutes spent writing each daily report.
Have the reviewer note every factual correction made to those reports.
Divide reports with at least one correction by all reports filed.

Run the AI drafts on the same job for the next two weeks and compare both numbers. Write down who checks each draft before the pilot starts, which also gives the job one of the written rules Punchcard counts as formal support.

Availability and cost

The full report is a free download from Punchcard’s site after a registration form. It adds trust figures for all eight heavy-industry tasks and compares AI support across seven heavy industries.

Miter raises $40 million as its contractor payroll platform turns crew photos and voice notes into safety reports

Sun, 04 Oct 2026 09:27:29 -0400

Miter, a San Francisco company that sells payroll, HR, time tracking and field software to contractors, announced a $40 million Series B round on September 30, 2026. Battery Ventures led the round, with Bessemer Venture Partners and Coatue taking part, and Miter has now raised $78 million in total, according to its release. The release names two places where the AI features sit: back-office processing of invoices, receipts and pay rate tables, and field safety reporting, where a crew member’s photo or voice note becomes an observation or incident report.

What it does

Miter began as construction payroll, and that is still the core. The software handles job costing to cost codes, certified payroll reports for prevailing wage jobs, union rules and tax withholding for crews that cross state lines. The release says the AI reads invoices, receipts and pay rate tables and shows labor and job cost trends while a job is still open. The daily users are the controller or payroll administrator and the project accountant.

The safety product launched on May 4, 2026, and is the newer piece. A foreman or crew member photographs a hazard or records a voice memo on a phone, and Miter turns it into a structured observation or incident report with the hazard type, severity and description filled in. A worker can dictate in Spanish and the office receives the report in English, and the app works offline and syncs later. The safety manager reviews the draft, then assigns a corrective action with an owner and a due date.

Safety leads can also generate a toolbox talk from a topic such as heat illness, or a job hazard analysis from a description of the task. Because hours worked already sit in the same system, Miter calculates TRIR and DART from actual hours. TRIR, the total recordable incident rate, is the count of OSHA recordable cases times 200,000, divided by hours worked. BLS explains that 200,000 hours stands for 100 employees working 40 hours a week for 50 weeks. DART uses the same formula for cases with days away, restricted duty or job transfer. Miter also generates OSHA Forms 300, 301 and 300A.

Customers named in the release include Clayco and Haugland Group.

Evidence so far

The evidence so far is Miter’s press release, a post by CEO Connor Watumull, and customer stories on Miter’s website. These are the figures Miter has published:

Figures Miter reports
Claim	Figure	Who reported it
Series B size and total raised	$40 million round, $78 million raised in total	Miter press release
Contractors on the platform	More than 2,000	Miter press release
Share of US construction workers paid through Miter	Roughly 2 of every 100	Miter press release and CEO post
People paid through Miter each week	About 100,000	Miter CEO post
Customer count since the May 2025 Series A	Tripled (release), nearly tripled (CEO post)	Miter press release and CEO post
Products launched in the last six months	Safety, Performance, Data Studio, Accounts Payable	Miter CEO post
Weekly payroll hours at Almendariz Consulting	About 30 before Miter and about 2 after	Almendariz controller in a Miter customer story

Source: prnewswire.com

The share of workers and the weekly headcount do not line up neatly when you check them against BLS.

BLS puts construction employment at 8,364,000 in September 2026, seasonally adjusted and preliminary.
Two of every 100 would be about 167,000 workers.
Watumull’s post describes the system as one that “pays a hundred thousand people every week.”

One reading is that the ratio counts everyone paid over a longer period, such as a year, which runs higher than a weekly count on a seasonal workforce. Another is that Miter used a different base than the BLS construction sector. The release gives the ratio without its base, so either reading fits.

The measured results on Miter’s site cover payroll. At Almendariz Consulting, a California traffic control contractor with 70 to 100 employees, the controller says payroll fell from about 30 hours a week to about 2. Other customer stories listed on the safety page report payroll time cuts of 60% at Green Mechanical and 75% at W.J. O’Neil. Those stories describe payroll and onboarding automation.

For the AI features themselves, the evidence is product description and one customer comment. An executive at Rubicon Applied Divers, a commercial diving and marine contractor, is quoted on the safety page saying, “The transparency into TRIR and who signed off on safety meetings is bar none the best.” Accuracy rates for the photo and voice reports and the invoice reading, and time saved per report, are not yet public.

How to measure it on your projects

Track safety observations logged per 1,000 crew hours, by job. Count every hazard and behavior observation submitted in a month, leave out incident reports, divide by the hours worked on that job that month, and multiply by 1,000. If the photo and voice capture lowers the effort of reporting, as Miter says it does, this rate should rise on the pilot job.

Capture the baseline before the pilot:

Pick one job with a stable crew size for the next three months.
Count the observation cards or app entries from that job for the last three months.
Pull the hours worked on that job from payroll for the same months, and compute the rate for each month.

During the pilot, keep the same count. Also pull 20 AI-drafted reports and note how many fields the safety manager changed before closing each one.

A first step this week is to pull last quarter’s observation cards for one job and the matching payroll hours.

Availability and cost

Miter Safety has been available since its May launch. Miter’s site sends buyers to a demo request, and the pages opened for this post list no price sheet. The Almendariz customer story gives one data point, a cost of about $2,200 a month for that 70 to 100 person firm’s use of Miter’s payroll and HR product. The safety page lists integrations with Procore, Sage, QuickBooks and NetSuite, among others, and a REST API.

Top AI model matches building drawings to photos 83.9% of the time in the ARCH-B benchmark

Sat, 03 Oct 2026 09:05:11 -0400

Two researchers from Harvard and Northeastern University posted ARCH-B to arXiv on September 28, 2026. It is a benchmark, a fixed set of test questions, with 354 multiple-choice items that ask an AI model to match photographs, floor plans, elevations and sections of the same building. Kieran Sagar Parikh and Jose Luis Garcia del Castillo y Lopez ran it on 25 multimodal models, meaning models that read images as well as text, and Gemini 3.1 Pro Preview scored highest at 83.90%. Untrained people recruited online averaged 35.35%. The skill under test comes up in tools that relate a drawing set to the building, such as software that pins field photos to a sheet or AI takeoff that reads plan PDFs.

What it does

Every question has a short instruction, one or more reference images, and four answer images with one correct choice, so a model that guesses scores 25%. The 354 questions fall into 11 task types. Most ask for the image that matches the references, in either direction between drawings and photos. A typical instruction reads, “Choose the photograph that corresponds to the building shown in this/these floorplan/s.” Two task types ask for the one image that belongs to a different building.

The 11 ARCH-B tasks and how models and people scored on each
Task	Questions	Best model (%)	Best model	Average of 25 models (%)	Untrained people (%)
Pick the floor plan that matches 2 to 4 photos	23	95.7	Gemini 3.1 Pro Preview	58.8	34.7
Pick the photo that matches 1 to 4 floor plans	56	76.8	GPT-4.1	44.1	36.4
Find the photo from a different building	46	84.8	Gemini 3.1 Pro Preview	41.2	21.6
Pick the photo that matches three others	24	100.0	GPT-4.1	60.2	32.0
Pick the interior that matches 4 exterior photos	25	96.0	Gemini 3.1 Pro Preview	52.6	35.9
Pick the exterior that matches 4 interior photos	20	85.0	Claude Opus 4.5	50.0	29.7
Pick the elevation that matches the floor plans	32	90.6	Gemini 3.5 Flash	50.3	37.0
Pick the floor plan that matches the elevations	30	90.0	Gemini 3.5 Flash	57.9	32.3
Find the odd one out among a plan, elevation, section and photo	37	100.0	Claude Fable 5	64.2	33.9
Pick the photo that matches a plan, elevation and section	29	96.6	Gemini 3 Flash Preview	59.9	42.4
Pick the exterior photo that matches 2 elevations	32	96.9	Gemini 3.5 Flash	61.0	56.2

Source: arxiv.org

The images come from real-estate listings on Zillow and Realtor.com and from project pages on ArchDaily and Dezeen, collected daily from November 2024 to October 2025. That gave the authors 3.9 million images, 58,000 of them floor plans, grouped by the listing or project they came from. The grouping is what makes an answer checkable, because every image in a group shows the same building.

The wrong answers are deliberately similar to the right one. For each question, the authors picked the three images from other buildings that an image model rated most similar to the correct one, so each decoy is a similar-looking building or drawing. They then cut the pool down in three steps:

Generate 2,200 candidate questions, 200 per task type.
Run them through four screening models (Claude Sonnet 4, Gemini 2.5 Pro, Pixtral Large and GPT-4.1) and keep the 1,013 that two or fewer of the four answered correctly.
Have two researchers review every remaining question by hand and drop any with duplicate images, mislabeled drawing types, or no single answer a person could defend.

From image collection to final benchmark
Stage	Count
Images collected (Nov. 2024 to Oct. 2025)	3,900,000
Floor plans among them	58,000
Real-estate listings	95,000
Architectural projects	41,000
Candidate questions generated	2,200
Candidates answered correctly by 2 or fewer of 4 screening models	1,013
Questions in the final benchmark	354
Human answers collected	5,830

Source: arxiv.org

In a contractor’s office, the likely user is a VDC manager or technology lead comparing vendors whose products read drawings. The released runner lets you score a given model on the same 354 questions. The source images are listing photos and published design work, and the set contains no construction document sets or progress photos. The authors name construction documents and images from different stages of construction as possible additions in future versions.

Evidence so far

The results come from the authors’ own runs of the 25 models and a crowdsourced human study, all reported in the paper. A replication by another group is not yet public.

ARCH-B accuracy by model, 354 questions
Rank	Model	Developer	Accuracy (%)	Valid response rate (%)
1	Gemini 3.1 Pro Preview	Google	83.90	97.46
2	Gemini 3.5 Flash	Google	83.05	98.87
3	Claude Fable 5	Anthropic	80.23	98.87
4	GPT-4.1	OpenAI	77.97	100.00
5	GPT-5.5 reasoning	OpenAI	76.55	99.72
6	Gemini 3 Flash Preview	Google	75.99	98.02
7	GPT-5.4 reasoning	OpenAI	74.29	99.72
8	Gemini 2.5 Pro	Google	71.47	100.00
9	GPT-5.5	OpenAI	70.90	99.72
10	Claude Opus 4.6	Anthropic	70.34	100.00
11	Claude Opus 4.5	Anthropic	65.25	99.15
12	GPT-5.4	OpenAI	64.41	98.87
13	Grok 4.20 reasoning	xAI	57.06	99.15
14	Grok 4.3	xAI	50.28	98.87
15	GPT-5.4 mini reasoning	OpenAI	50.00	100.00
16	GPT-5.4 mini	OpenAI	46.61	99.72
17	Gemma 3 27B	Google	38.98	99.72
18	Claude Sonnet 4.6	Anthropic	37.85	60.45
19	Claude Sonnet 4.5	Anthropic	34.46	68.08
20	Pixtral Large	Mistral	32.77	97.18
21	GPT-5.4 nano reasoning	OpenAI	29.10	99.72
22	Claude Haiku 4.5	Anthropic	17.51	94.63
23	Gemini 2.5 Flash	Google	17.51	24.29
24	GPT-5.4 nano	OpenAI	16.95	99.44
25	Gemma 3 4B	Google (run on Ollama)	10.45	100.00

Source: arxiv.org

Read the low end of the table with the valid response rate in view. A response that timed out or did not contain a valid answer ID counted as wrong. Gemini 2.5 Flash returned a usable answer on 24.29% of questions, Claude Sonnet 4.6 on 60.45% and Claude Sonnet 4.5 on 68.08%, so their scores partly measure formatting failures. Across all 8,850 model-question runs, 37 ended in a provider error and 559 produced no valid answer ID.

The more useful finding is in the task table. Averaged across all 25 models, accuracy was 58.78% when the model had photos and had to pick the floor plan, and 44.14% when it had floor plans and had to pick the photo. People scored about the same in both directions, 34.74% and 36.41%. The authors read this as models finding it easier to check a plan’s layout against views they can see than to picture how a plan will look once built.

The best model on each task beat the human average by 40.38 to 68.00 percentage points. That human baseline needs a caveat the authors state themselves. Participants were English-speaking US adults on the Prolific research platform who each answered 10 questions from one task type (they were paid $2 per session whatever their score), and the authors call them a non-expert baseline. They suggest future comparisons with architecture students and practitioners. A superintendent or architect who reads elevations every day would be a different reference point.

The authors also checked whether screening with four models had made the benchmark hard only for those four. They grouped questions by how many screening models got them right and scored the models outside the screening set on each group.

Accuracy of the models outside the screening set, by screening result
Screening models that answered correctly	Questions	Accuracy of the other models (%)
0 of 4	33	28.65
1 of 4	77	41.38
2 of 4	244	59.02

Source: arxiv.org

Accuracy rose from 28.65% to 59.02% as screening difficulty fell, which the authors take as evidence that the hard questions are hard in general. Per question, average model accuracy was 53.36%, with a range from 4% to 96%. No question was answered correctly by every model, and none was missed by all of them.

How to measure it on your projects

So the paper’s split by direction is worth copying. If you pilot a tool that ties photos to drawings, such as a reality capture product or a photo log that places shots on sheets, track the location match rate. It is the share of a fixed sample of items that the tool placed on the correct sheet and in the correct room or grid bay, as checked by the project engineer who knows the job.

Measure it both ways, since the models in ARCH-B scored differently in each direction:

Photo to plan. Take field photos the tool has already placed and check each placement against the sheet.
Plan to photo. Pick rooms or grid bays on a sheet, ask the tool for the matching photos, and check that each returned photo shows that area.

Capture the baseline before the pilot starts. Pull a sample from last month’s photo log, record how many photos carry a correct location tag, and time how long the project engineer spends tagging a week’s photos by hand. AI takeoff tools that read plan PDFs, such as the one Quotr raised a seed round for, can be checked the same way by comparing a sample of extracted quantities to an estimator’s count. A first step on one job is to pick a single floor and one drawing sheet, and have the project engineer mark the correct location for each photo taken there since the last pay application.

Availability and cost

ARCH-B is free. The question set, a JSON schema and the evaluation runner are posted as ancillary files with the arXiv paper, and the README and runner are under the MIT License. The question file links to each image at its original public address with a SHA-256 hash, a fingerprint the runner uses to confirm the image has not changed, and the authors do not redistribute the images themselves. The runner calls each model provider’s API, the paid connection a program uses to send questions to the model, and the README notes that provider use can incur charges. The authors also host a version of the human-study quiz at archbench.codecolab.org, where anyone can try sample questions.

Kuro raises €10 million for AI tender review that customers say saves about 90 hours per bid

Fri, 02 Oct 2026 07:39:49 -0400

Kuro, a Berlin company that sells AI software to general contractors, announced a €10 million funding round on September 30, 2026. UVC Partners led the round, and Foundamental, Realyze Ventures and helloworld also invested, according to the announcement on UVC’s website. The software works on bid-stage document review, the job an estimating team does before it prices a project: reading the drawings, specifications and bills of quantities in a tender and finding where they disagree.

What it does

The contractor’s team uploads the whole tender package. Kuro’s platform page lists the formats it reads, including PDFs, Office files, emails, images, IFC models and GAEB files, the German exchange format for bills of quantities. The software sorts and versions the documents. When an owner reissues a package, it shows which files changed between the two uploads and removes duplicates.

Each customer then loads what Kuro calls a rulebook, made up of its own building components, review criteria and work packages. With that in place, the software answers a standard list of questions about each new project and checks the drawings, the specs and the bill of quantities against each other. It rates each contradiction it finds by risk, and every answer links back to the source page. When an estimator corrects a result, the correction stays in the rulebook for the next job.

Kuro’s website describes further modules:

A room data agent that reads each room’s requirements, from flooring to ventilation, from the drawings and specs and writes them back into the Revit model.
A bid check that converts subcontractor bids sent as PDFs into GAEB files and flags where a sub changed quantities, wording or line items from the bid form.
An agent that screens public tenders against the company’s references and pre-fills the applications to bid.

So the daily users are estimators and preconstruction managers, with BIM staff on the Revit side. Kuro says it started in estimating and plans to extend the platform to procurement, accounting and compliance.

Evidence so far

The evidence so far is the funding announcement, Kuro’s website and a case study of one customer, Köster GmbH, a turnkey general contractor in Osnabrück. These are the figures Kuro has published:

Figures Kuro reports for its software
Claim	Figure	Who reported it
Time saved per bid	About 90 hours	Kuro, citing customer reports, in the funding announcement
Users who say Kuro found information they would otherwise have missed	70%	Kuro, in the funding announcement
Estimating time saved at Köster	6 hours per estimator per week on average	Survey of Köster estimators, in Kuro's case study
Size of a tender package at Köster	About 4,000 A4 pages on average, over 100,000 on the largest projects	Kuro's case study
Customers	12, of which 6 are among the 20 largest general contractors in the German-speaking region	Kuro, in the funding announcement
Pages processed	Over 1 million per month	Kuro's website

Source: uvcpartners.com

Each of the headline figures counts something a little different.

The 90 hours per bid comes from customer reports, as the announcement describes it.
The 70% is a share of users who say Kuro found information they would have missed. It records what users believe the tool caught.
The Köster figure comes from a survey of Köster’s estimators, so it is self-reported time, averaged per estimator per week.

A detection rate would count something else: the share of real conflicts in a tender package that the software flagged. Software teams call that recall. The bid sample behind the 90-hour figure and the software’s miss and false-alarm rates are not yet public.

The case study adds a detail about how Köster tested the product. Lea Scherer, Kuro’s CEO, says in it that Kuro first had to show on Köster’s past projects that it found problems Köster’s own team had not. Nils Köster, Köster’s CEO, describes the payoff in the announcement: “What matters is that we spot risks in the tender quickly, before we submit our price.” (Kuro was founded in November 2024 and has 16 employees in Berlin.)

How to measure it on your projects

Track the pre-bid catch rate for document conflicts. Count every conflict between drawings, specs and bid forms on a job. The catch rate is the share your team flagged before submitting the price. The rest surface later as RFIs, change orders or cost the job absorbs.

Köster’s trial of Kuro started with past projects, and the baseline can too:

Pick 5 to 10 recently awarded jobs that are past the early RFI phase.
Go through each job’s RFI log and change order log, and tag every item that traces to a conflict already present in the bid documents.
For the same jobs, count the conflicts your estimators flagged in bid clarifications or qualifications.
Divide the flagged count by the flagged count plus the tagged count.

In a pilot, run a tool like Kuro on the bid documents of those same jobs. Count how many of the tagged conflicts it flags, and how many of its flags were false alarms on review.

A first step this week is to pull the RFI log on one recently awarded job and tag the RFIs that came from conflicts in the bid set.

Availability and cost

Kuro’s website is in German, and the customers it names are in Germany, Austria and Switzerland: Goldbeck, Hochtief Infrastructure, Köster, Implenia, Geiger and Riedel. Kuro says it hosts project data in the EU and holds ISO 27001 certification. The platform reads data from and writes data back to Revit, IFC, iTWO, SharePoint and Office files.

Kuro sells by inquiry, through a contact address on its website. Availability for US contractors could not be verified.

Quotr raises $4 million for AI takeoff software that it says cut 20-hour takeoffs to 1 to 2 hours

Thu, 01 Oct 2026 07:39:18 -0400

Quotr, a Berkeley company founded in 2024 that sells AI takeoff and estimating software to contractors, announced a $4 million seed round on September 30, 2026. Llama Ventures led the round, with strategic angel investors taking part, according to Quotr’s release on Business Wire. The product covers the estimating workflow from plan set to proposal: quantity takeoff, pricing, the proposal document and, if the contractor wants it, material procurement.

What it does

An estimator uploads a plan set as PDF or image files. Quotr’s software sorts the sheets, then counts symbols, measures lengths and calculates areas across the whole set. Quotr says it uses computer vision for this, meaning software trained to recognize objects in images, here the outlets, fixtures and openings drawn on a sheet. Every extracted quantity stays editable, and the estimator reviews and corrects the counts before pricing.

Pricing comes from the contractor’s own cost database or from Quotr’s average costs by US zip code. Line items stay linked to the markups on the drawing, and a finished estimate exports as a proposal PDF with scope, exclusions and pricing. A chat assistant, which Quotr calls the Quotr AI Agent, answers plain-language questions about the plan set, such as a symbol count or a room dimension.

The software page lists 23 trades, from electrical and drywall to structural steel and earthwork. Quotr’s four published case studies cover an electrical subcontractor, a Bay Area homebuilder, a modular manufacturer in South Carolina and a hospitality general contractor in Raleigh. So the daily user is an estimator at a specialty contractor or in a GC’s preconstruction group. Developers can instead send drawings to Quotr’s estimating desk, which the company says pairs AI analysis with review by its own estimators.

The release says the money will go toward enterprise deployments, a larger supplier network and a move into heavy industrial and mission-critical work, data centers included.

Evidence so far

The evidence so far is Quotr’s press release, its website, and a testimonial from an estimator at RL Electric, the electrical subcontractor in the case studies. These are the figures Quotr has published:

Results Quotr reports for its software
Claim	Figure	Who reported it
Takeoff time before and after Quotr	Roughly 20 hours before and 1 to 2 hours after	Quotr press release, and an RL Electric estimator quoted on Quotr's website
Reduction in takeoff time	Up to 80%	Quotr press release
Bids submitted per month	40% more	Quotr press release
Construction value estimated on the platform	Over $1.2 billion cumulative	Quotr press release

Source: businesswire.com

The takeoff percentage and the example attached to it describe the same result in different terms.

The release and the RL Electric testimonial both describe takeoffs that took roughly 20 hours and now take 1 to 2 hours.
Going from 20 hours to 2 is a 90% cut. Going to 1 hour is a 95% cut.
The release labels that result “up to 80%.”

The headline percentage comes out lower than the example offered with it. One reading is that 80% is a figure across more customers and the 20-hour job is the best case. The release text fits either reading, and the RL Electric case study page describes its outcomes in words, such as reduced manual counting. Quotr’s ROI calculator on the software page uses an 80% cut as a fixed input, so its sample result of $160,322 a year for two estimators at $65 an hour follows from that assumption.

The base period and customer sample behind the 80% and 40% figures, and count accuracy by trade, are not yet public. (One of the three testimonials on Quotr’s homepage comes from a principal at Llama Ventures, the round’s lead investor.)

How to measure it on your projects

Track takeoff hours per sheet, including review. Start the clock when the estimator opens the plan set and stop it when quantities are ready for pricing, then divide by the number of sheets. Count the time spent checking and correcting the software’s counts, since Quotr’s workflow has an estimator review every quantity.

Capture the baseline first:

Pull the last 10 bids in one trade, such as electrical rough-in or drywall.
Record the takeoff hours for each from timesheets, or have the estimator reconstruct them.
Record the sheet count for each bid.

During a pilot, run the next 10 bids in that trade through Quotr and log the same hours per sheet. On two of those bids, recount a sample of symbols by hand and write down how many the software missed or counted twice. A 10-bid pilot will likely run past the 7-day free trial, so budget at least one paid month of a Lite seat.

A first step this week is to time the takeoff on the next bid that comes in and divide by its sheet count.

Availability and cost

Quotr sells the software per seat. Lite costs $79.90 per seat per month and includes a basic AI agent, on-screen takeoff, symbol and area detection and a custom cost database. Plus costs $299.90 per seat per month and adds advanced AI, Quotr’s full cost database, two hours of guided onboarding, project sharing, 2,000 square feet of takeoff credits a month and 10% off procurement. Enterprise pricing is custom.

The managed estimating service is priced by area, at $0.25 per square foot under 50,000 square feet and $0.10 per square foot at 50,000 and above. Quotr’s pricing page says development cost estimates arrive in 3 to 4 business days. The procurement program is separate from the software and sources windows, doors, cabinetry, flooring and bath fixtures from more than 50 audited manufacturers in Foshan and Guangdong, China.

Buildertrend's AI Submittals drafts the submittal log from an uploaded spec book

Wed, 30 Sep 2026 07:38:45 -0400

Buildertrend announced a set of AI and accounting features on September 29, 2026, aimed at builders taking on commercial and larger residential work. The one with the most bearing on project engineers is AI Submittals. Buildertrend says it generates the submittal log from the project documents and then routes each submittal through review and approval. The same release added AI Bill Capture for accounts payable, a Sage Intacct integration with two-way sync of bills and invoices, a Bluebeam integration, and AIA-style progress invoicing with a schedule of values and retainage.

What it does

Building the submittal log is usually a project engineer’s job at the start of a project. They read the spec book section by section, list every product data sheet, shop drawing and sample the specs call for, and assign each one a CSI division, a responsible sub and a due date. A line missed at this stage can surface months later as a product installed without approval.

Buildertrend’s help article describes the AI step. On the job’s Submittals dashboard, a user picks Generate Submittals and uploads a spec book. Buildertrend drafts the submittal log and returns it as a CSV file, which the user reviews and then imports. Buildertrend’s AI page says the tool also pulls in plan sets and organizes both into an importable format.

The CSV step matters for control. The project engineer can open the draft in Excel, delete lines that do not apply, add the ones the tool missed, and fix division codes before anything reaches the log. Division codes come from a standard CSI list set at the job level.

After import, the help article describes a fixed routing workflow. Each submittal goes to the assigned sub to upload documents, then to one or more approvers, and returns to the submittal manager to publish. Approvers answer with approved, approved as noted, revise and submit, or rejected. Step due dates are calculated from the days set for each step, and the log shows who holds the ball in court. Architects, engineers and subs can respond from an emailed link without a Buildertrend login.

AI features Buildertrend lists in the September 29 release
Feature	Who uses it	What goes in	What comes out
AI Submittals	Submittal manager or project engineer	A spec book uploaded to the job	A draft submittal log delivered as a CSV to review and import
AI Bill Capture	Accounts payable staff	Bills and receipts	Vendor details and line items and cost codes pulled into the bill
AI Plans (Page Naming)	Project manager or document controller	An uploaded plan set	Sheets named and organized with slip sheet revisions tracked
AI Client Updates	Project manager	Daily job entries already in Buildertrend	A drafted client update for review before sending

Source: buildertrend.com

Evidence so far

The evidence so far is a customer comment in the release. Jacob Wittenberg of Edgemont Builders, a Baltimore firm that builds single-family, multifamily and light commercial projects, said, “AI Submittals is the clearest example; it’s become core to how we run a job now, especially on our commercial projects.” Accuracy rates, time savings and pilot results for AI Submittals are not yet public.

These claims come from Buildertrend:

AI Submittals keeps architects, owners, subs and the field aligned on commercial and high-end residential projects.
The Bluebeam integration gives faster submittal and RFI turnaround and lowers rework risk.
Over 20,000 builders and over half of U.S. home builds run on Buildertrend.

The launch date is unclear. The release lists AI Submittals among its new capabilities, but Buildertrend’s AI page carried a Submittals entry before September 29. When the feature first reached customers could not be verified.

How to measure it on your projects

Track log recall, meaning the share of required submittals on a job that the AI draft included. Count a line as required if it appears in the final log for the job. Divide the lines the draft caught by all required lines.

Capture the baseline on a recent job before you pilot:

Take the final submittal log from a finished commercial job and count its lines.
Count the lines added after the log was first issued, since those are the items the manual pass missed.
Ask the project engineer who built the log how many hours it took, and write the figure down.

On the pilot job, upload the spec book before anyone builds the log by hand. Save the CSV as it came out, then record the hours spent cleaning it up and every line added later. Compare recall and hours against the baseline job. Count the lines you deleted as well, since each false line costs review time.

A first step you can take this week is to pull the final log from one closed job and mark which lines were added after the first issue.

Availability and cost

Buildertrend prices by custom quote. Its pricing page asks for builder type and annual volume, and it says the subscription includes unlimited users. The AI page says there is no separate AI application to purchase. Whether AI Submittals is included in every plan could not be verified.

ServiceTitan survey finds 62% of commercial service contractors have tried AI and 15% of users report clear ROI

Tue, 29 Sep 2026 07:39:00 -0400

ServiceTitan released its 2026 Commercial State of the Trades report on September 24, 2026. Thrive Analytics ran the survey for ServiceTitan online from July 10 to July 28, polling 1,020 owners, executives and general managers at commercial service contractors, most of them mechanical, electrical and plumbing firms working on commercial buildings. ServiceTitan says 62% of those firms have piloted or deployed AI, and that 15% of the contractors using it report a significant positive impact with clear ROI. The workflows in question are service scheduling and dispatch, where respondents expect AI to matter most, and the invoicing and collection cycle that the same survey tracks.

What it does

The report covers the service side of MEP work, meaning maintenance agreements and repair calls. ServiceTitan sells field service and construction software and now describes its product as an agentic operating system for the trades, so the sponsor has a stake in the AI answers. (An agentic system is software that carries out a chain of steps by itself, for example booking a technician and then sending the invoice when the job closes.)

The full report sits behind a download form. This post works from the press release and the report’s landing page, which carry the headline figures. The table below collects the AI questions.

AI questions in the 2026 Commercial State of the Trades survey
Survey answer	Asked of	Share (%)
Have piloted or deployed AI	All 1020 respondents	62%
Actively using AI or have embedded it across the business	All 1020 respondents	33%
Name AI as a top technology priority	All 1020 respondents	33%
Report a positive impact from AI	Contractors using AI (count not published)	59%
Report a significant positive impact with clear ROI	Contractors using AI (count not published)	15%

Source: servicetitan.com

Watch the bases. The first three rows cover all 1,020 respondents. The last two cover only “contractors using AI,” and ServiceTitan does not say how many respondents that is.

These figures are most useful to a service manager choosing where to pilot a dispatch tool, or to a CFO weighing a software budget against cash flow. Respondents picked scheduling and dispatch (37%) and predictive maintenance (31%) as the areas where they expect AI to have the greatest impact. The release also names estimating, remote diagnostics and customer inquiries, without shares.

Where respondents expect AI to have the greatest impact
Area	Share expecting AI's greatest impact here (%)
Scheduling and dispatch	37%
Predictive maintenance	31%

Source: servicetitan.com

Evidence so far

Every figure here is self-reported, so the ROI answers are each respondent’s own judgment of whether AI helped. The margin of error and the way Thrive Analytics recruited the panel are not in the release.

The 15% figure depends on a base the release leaves open. There are two readings:

If “contractors using AI” means the 33% who are actively using it or have embedded it, about 337 respondents, then 15% of them is about 5% of the full sample.
If it means the 62% who have piloted or deployed it, about 632 respondents, then 15% is about 9% of the full sample.

Under either reading, fewer than one respondent in ten reported a clear return. The landing page says “only 15% can point to a clear, proven return on investment so far,” and does not restate the base.

The year-over-year comparisons come from ServiceTitan’s annual survey of commercial service contractors, which drew on the same kind of MEP-heavy sample in May 2025. The 2026 release reports three of them.

Changes ServiceTitan reports against its 2025 survey
Survey answer	2025 (%)	2026 (%)	Change (points)
AI is a top technology priority	15%	33%	18
Improving cash flow is among the top three business goals	28%	40%	12
More than half of commercial customers are on service or maintenance agreements	42%	46%	4

Source: servicetitan.com

One of those comparisons does not match last year’s release. The 2026 release puts the 2025 share of contractors with more than half their commercial customers on service or maintenance agreements at 42%. ServiceTitan’s July 14, 2025 release reported 63% of contractors with more than half their customer base on preventative maintenance agreements. The question wording may have changed between years, and the gated reports would show that, but neither public release explains the difference. Treat the rise from 42% to 46% as unconfirmed.

So, on cash flow, 40% of respondents now rank improving cash flow among their top three business goals, up from 28% in 2025. Increasing net profit margins ranks first at 42%, and acquiring new customers trails at 29%.

Business goals in the 2026 survey
Business goal	Share of contractors (%)
Increase net profit margins	42%
Improve cash flow	40%
Acquire new customers	29%

Source: servicetitan.com

The billing questions measure the time from finishing work to getting paid. Eighty-two percent send invoices within three days of completing work, yet 96% wait at least 15 days to be paid and 30% wait more than 30 days. The 2025 release reported averages instead, two days to invoice and 28 days to get paid, so the two years cannot be lined up question by question.

Invoicing and payment timing, 2026 survey
Billing and collection	Share of contractors (%)
Send invoices within three days of completing work	82%
Wait at least 15 days to be paid	96%
Wait more than 30 days to be paid	30%

Source: servicetitan.com

The field questions describe records that are missing at the job. Sixty-nine percent cite warranty coverage and agreement details as a top obstacle for technicians. The 2025 release put the lack of fast access to warranty and service agreement data at 70%, parts availability at 57% and retrieving service and repair history on site at 55%. The 2026 release names parts and service histories as challenges and gives no shares for them.

Obstacles for technicians in the field
Obstacle for technicians in the field	2025 (%)	2026 (%)
Warranty coverage and agreement details	70%	69%
Parts availability	57%
Service and repair history on site	55%

Source: servicetitan.com

ServiceTitan ran a separate survey this year of more than 1,000 commercial construction leaders at commercial building firms, released March 30, 2026. It reported that 38% of contractors saw measurable business impact from AI, up from 17% in 2025, with cost estimation and budgeting (24%) and bid management (22%) as the leading uses. Only 20% ran on a single platform. The samples differ, so the March figures describe a different population from the September ones.

ServiceTitan's March 2026 survey of commercial construction firms
Survey answer	2025 (%)	2026 (%)
Report measurable business impact from AI	17%	38%
Apply AI to cost estimation and budgeting		24%
Apply AI to bid management		22%
Run the business on a single platform		20%

Source: servicetitan.com

How to measure it on your projects

Scheduling and dispatch topped the list, so a dispatch pilot is where most respondents expect a return. Track first-time fix rate, the share of service calls closed on the first technician visit with no return trip for parts or information. It ties to the field obstacles in the survey, since a technician who arrives without the warranty terms or the unit’s service history usually has to come back.

Capture the baseline before the pilot starts:

Pull 90 days of closed work orders from your service system.
Divide the calls closed on the first visit by all calls closed.
Record the reason for every return trip.

Run the tool on one branch or one service agreement portfolio for the same 90 days, then compare the rate and the reasons. If the pilot is on invoicing, track calendar days from work completion to cash received for each invoice, and count how many cross the 15-day and 30-day marks the survey uses. If your service system has no return-trip reason field today, add one on a single branch this month.

Availability and cost

The full report, titled 2026 Commercial Service Market Report on ServiceTitan’s site, is a free download after a registration form.

viAct says AI on Alliad's existing site cameras cut recorded PPE violations 41.7% in three months

Mon, 28 Sep 2026 07:39:44 -0400

viAct, a Singapore-based computer vision company, published a case study on August 28, 2026, saying its video analytics cut recorded PPE non-compliances by 41.7% in three months on a hospital project that the contractor Alliad is building in Tiassalé, Côte d’Ivoire. TNGlobal reported the results on September 28. The system handles the safety manager’s routine PPE and exclusion-zone checks, and it ran on cameras Alliad already had on the job.

What it does

On most jobs a safety manager or superintendent checks PPE by walking the site and writing up what they see. viAct’s software takes the live feeds from a site’s existing IP cameras over RTSP, a standard protocol IP cameras use to stream video. A computer vision model, software trained to recognize objects in images, looks at each worker in frame and flags any required item that is missing.

On Alliad’s site the model checked four items: hard hat, vest, safety shoes and gloves. The same feeds flagged entry into restricted zones and signs of fire or smoke. viAct’s product page says alerts can go out by site speaker, SMS, WhatsApp or email, and a dashboard rolls the detections into daily, weekly and monthly reports.

The site had little fixed telecom service, so the cameras reached the analytics over mobile data networks, according to the case study. It describes the system as an addition to Alliad’s existing site supervision, and the safety team decides what to do with each alert.

Evidence so far

What viAct and Alliad report from the first three months
Item	What the companies report	Reported by
PPE non-compliances	41.7% fewer in the first three months	viAct and Alliad
Danger-zone intrusions	46.3% fewer over the same period	viAct and Alliad
Fire and smoke detection	Running on all connected cameras from day one	viAct and Alliad
PPE items checked	4 (hard hat, vest, safety shoes, gloves)	viAct
New cameras added	None	viAct and Alliad

Source: viact.ai

Every result in the table comes from viAct and Alliad. TNGlobal lists the details that are not yet public: the size of the deployment, the baseline violation counts, the number of cameras and any independent audit.

The baseline deserves the closest look. A 41.7% reduction needs a starting count, and the sources do not say which period supplied it or how it was counted. If it came from the system’s own first weeks, the figure measures how behavior changed once workers knew the cameras were checking. The case study credits the drop to workers and supervisors adjusting to continuous monitoring, and that change is the outcome a PPE program is after. The count still covers only what the cameras can see, so a worker without gloves outside camera view appears in neither number.

These claims come from viAct and Alliad:

PPE non-compliances fell 41.7% in the first three months.
Danger-zone intrusions fell 46.3% over the same period.
Fire and smoke detection ran on every connected camera from the first day.
The pilot gives Alliad a model it can repeat across its 10 to 20 concurrent sites.

viAct’s product page also claims a 92% reduction in PPE non-compliance and related fatalities after deployment, without naming a project or a measurement period.

Precision, here the share of alerts that a safety manager confirms as a real violation, is not reported for Alliad’s site. It decides how much of that person’s day goes to clearing alerts, and lighting and weather affect it. Our September 27 post on ConSynth-X covered a dataset researchers built to test jobsite camera models in rain, fog and darkness.

How to measure it on your projects

Track the PPE non-compliance rate from a manual walk audit. Divide the number of workers observed missing required PPE by the number of workers observed, and express it per 100 observations. A fixed route and the same observer keep the count independent of the cameras under test.

Capture the baseline before the system goes live:

Walk the same route at two set times each workday for two weeks, and record observations and violations by PPE item.
Mark which stretches of the route fall inside camera view.
Keep walking the route during the pilot, and report in-view and out-of-view rates separately.

For the first month of the pilot, have the safety manager mark each alert as confirmed or false. That gives you a precision figure for your own site, lighting and camera angles.

A first step is to pick one job with cameras already streaming and start the two-week walk audit before you call any vendor.

Availability and cost

viAct says its software connects to existing IP cameras without new hardware and can run in the cloud, on premises or in a hybrid setup. It lists more than 200 detection modules. viAct’s pages do not list prices.

Autodesk says Forma will replace Revit over time, without a forced switch

Sun, 27 Sep 2026 10:25:58 -0400

Autodesk’s FAQ for its Autodesk University 2026 announcements asks whether Forma, its cloud platform for architecture, engineering and construction, will replace Revit. The answer, in a September 15 article by Autodesk executive vice president Amy Bunszel, is “Over time, yes.” She writes that more of the design, analysis, coordination and documentation work done in Revit will move into Forma, that Autodesk will keep investing in Revit, and that it will not force a sudden transition. At a general contractor, the change reaches the VDC coordinator who pulls design models into clash detection and the document controller who manages the project’s file set.

The question has circulated since the conference. Hayley Swords, who hosts the Built For Tomorrow construction technology leadership podcast, asked it on LinkedIn. She wrote that Autodesk is not saying Revit is going away, and she described a gradual shift in which Revit becomes one part of a larger connected system of tools. Autodesk’s FAQ also describes a gradual change, and it answers the replacement question yes.

What changed

Autodesk has moved in steps over the past year.

On September 16, 2025, Autodesk announced that Revit would become the first Forma Connected Client, its label for a desktop product tied directly into Forma. That announcement said, “Forma is not replacing Revit, it’s extending it.”
On March 24, 2026, Autodesk Construction Cloud became part of Forma, and its products took Forma names. Autodesk said no data migration was required and that licensing, permissions, folder structures and integrations stayed as they were.
On August 20, 2026, an Autodesk News Q&A on the connected client said more work will happen directly in Forma over time.
On September 15, 2026, the AU 2026 FAQ answered the replacement question yes, and Autodesk said Civil 3D will become the next connected client.

Autodesk Construction Cloud product names since March 24, 2026
Old name	New name
Autodesk Construction Cloud	Autodesk Forma
Autodesk Docs	Forma Data Management
Autodesk BIM Collaborate Pro	Forma Design Collaboration
Autodesk Build	Forma Build
Autodesk Takeoff	Forma Takeoff
Autodesk Estimate	Forma Estimate
Preconstruction Bundle	Forma for Preconstruction
Model Management Bundle	Forma for Model Management
Construction Operations Bundle	Forma for Construction Operations
Autodesk Construction Cloud Connect (ACC Connect)	Forma Construction Connect

Source: autodesk.com

Forma Data Management, formerly Autodesk Docs, is now what Autodesk calls the common data environment for all of Forma. A common data environment is the shared repository where every company on a job finds the current drawings, models and approvals. The AI agents Autodesk announced for construction teams at the same event are covered in our September 25 post.

What it does

Today the connected client is a design-side feature. According to Autodesk’s Revit 2027 help pages and its August Q&A, an architect working in a Revit cloud model can reference other design models in a Forma scenario, add terrain, buildings and parcel data from the Forma Data Marketplace, and run Forma analyses such as wind without leaving Revit. Autodesk says this cuts the import and export steps between Forma and Revit.

For a contractor, the AU 2026 changes that touch coordination and document control sit in Forma Data Management and Model Coordination. Today a GC often holds a file until it clears review, then copies it to another folder with different permissions so subcontractors can see it. Autodesk says Smart Controls change who can see a file when its approval status changes, so the file stays in one place. In Model Coordination, Autodesk says teams can view point clouds with the federated model and run clash checks against a saved search set. It also says schedule activities can now be linked to model elements for a 4D simulation in Forma.

Civil 3D comes next. Autodesk says it will start as a tech preview that references Forma and Revit designs and shares one geolocation engine with them, which is meant to reduce the manual coordinate transforms that leave civil and building models misaligned.

Evidence so far

These claims come from Autodesk:

Connected workflows reduce the information teams rebuild as a project moves between phases.
Smart Controls remove the duplicate files created by copying between folders.
The shared geolocation engine reduces manual transforms and model misalignment.

Autodesk’s article describes the Kraft Group using Forma Data Management and Forma construction tools at Gillette Stadium. Measured results for these workflows, from Autodesk or from independent testers, are not yet public.

Autodesk’s own help pages set limits on the Revit tech preview. Autodesk recommends against using its features on production projects. Once a project administrator adds a project to the preview, the project cannot be removed from it. Processing the project’s Revit cloud models can take up to 30 minutes, and Autodesk lists slower Synchronize with Central on large, complex models as a known issue.

How to measure it on your projects

Track the handoff count per coordination cycle. That is the number of times someone exports, re-uploads or copies a model or sheet set between the design team’s issue and the coordination meeting, including folder copies made only to change who can see a file. This is the manual work Autodesk says Forma removes, and you can count it today on any platform.

Capture the baseline on one active job before changing anything:

For a month of coordination meetings, have the VDC coordinator log each export, upload and folder copy, with the file name and the minutes it took.
Mark the handoffs that exist only to change permissions. Smart Controls is aimed at those.
Count the clash runs that needed a model re-upload first.

Try the Revit tech preview on a copy of a project, because joining cannot be undone and Autodesk advises against production use. A first step this week is to start the handoff log at the next coordination meeting on one job.

Availability and cost

Revit as a Forma Connected Client is a tech preview for Revit 2027 and later. It works on Revit cloud models in projects stored in the United States, European Union or Australian data regions, and a project administrator turns it on one project at a time. The Autodesk pages cited here do not list a price for it.

Civil 3D will start as a tech preview, and Autodesk gave no date. Autodesk says Forma Data Management Essentials is included in standalone subscriptions to AutoCAD, Revit and Civil 3D. Several of the construction features announced at AU 2026 are in beta on separate timelines, and Autodesk points customers to its Forma product roadmap for dates.

Researchers release ConSynth-X, 34,199 synthetic jobsite images for testing site cameras in rain, fog and darkness

Sun, 27 Sep 2026 07:35:32 -0400

Four researchers at Kent State University and the University of North Texas posted a paper to arXiv on September 21, 2026, that releases ConSynth-X, a set of 34,199 construction-site images built to test jobsite cameras in bad weather and darkness. The team took 3,109 daytime photos from three public construction image collections and generated 11 altered versions of each, covering rain, snow, fog, night, precipitation at night, and distant views where workers and machines appear small. The work applies to the computer vision behind camera-based safety alerts and progress tracking, where software reads site photos or video to find workers, hard hats and equipment. The paper checks whether the altered images look like real bad-weather photos.

What it does

Most public jobsite images were shot in clear daylight, the authors say, because rain and night shifts are hard to photograph on a schedule and every image has to be labeled by hand. So a model that scores well on those images has not been tested on a worker without a hard hat before sunrise in fog. ConSynth-X gives a researcher the same scene under each condition. If accuracy drops on the fog version of a photo, the fog caused it, since the workers and the camera angle did not change.

The inputs are three existing datasets, and every altered image keeps the labels of the photo it came from. A bounding box is the rectangle drawn around each object a model is supposed to find, and ConstructionSite 10k supplies boxes for excavators, rebar and workers in white hard hats, along with safety-rule annotations. Because the labels carry over, nobody had to relabel 34,199 pictures.

Where the ConSynth-X images come from
Source dataset	Labels carried over	Source images	Scenes kept	ConSynth-X images
ConstructionSite 10k	Boxes for excavators, rebar and workers in white hard hats; captions; safety-rule annotations	10,013	1,586	17,446
SODA	Boxes for 15 classes of workers, materials, equipment and temporary site elements	19,846	608	6,688
SODA-KTSH	Five written captions per image	9,988	915	10,065
Total			3,109	34,199

Source: arxiv.org

Each condition came out of its own pipeline:

Rain and snow. An image-editing model rewrote the sky and surfaces from a text prompt, and a particle overlay then drew streaks or flakes in near, middle and far layers.
Fog. A depth model estimated how far each pixel sits from the camera, and a standard atmospheric formula faded distant objects more than near ones. Visibility was set at 750 to 1,000 meters for light fog, 500 to 750 for moderate and 300 to 500 for heavy.
Night. A day-to-night translation model relit each scene.
Rain or snow at night. The weather edit ran first, then the night translation, then the overlay.
Distant views. Each photo was shrunk inside a larger canvas, and an image model filled the border with ground and sky, so every object takes up less of the frame.

The 11 conditions, 3,109 images each
Condition group	Condition	How it was made	Passed 0.85 similarity gate	Images
Precipitation	Light rain	Image edit from text prompt plus rain-streak overlay	Yes	3,109
Precipitation	Heavy rain	Light-rain image plus denser streaks and haze	No	3,109
Precipitation	Light snow	Image edit from text prompt plus snowflake overlay	Yes	3,109
Precipitation	Heavy snow	Image edit from text prompt plus snowflake overlay	Yes	3,109
Fog	Light fog (750 to 1000 m visibility)	Depth estimate plus atmospheric scattering formula	Yes	3,109
Fog	Moderate fog (500 to 750 m visibility)	Depth estimate plus atmospheric scattering formula	Yes	3,109
Fog	Heavy fog (300 to 500 m visibility)	Depth estimate plus atmospheric scattering formula	Yes	3,109
Low illumination	Nighttime	Day-to-night translation model	Yes	3,109
Adverse weather at night	Rain at night	Weather edit then night translation then rain overlay	No	3,109
Adverse weather at night	Snow at night	Weather edit then night translation then snow overlay	No	3,109
Small-object or long-distance view	Small-object view	Photo shrunk inside a larger canvas with the border filled in by an image model	Yes	3,109

Source: arxiv.org

The authors compare their release with five earlier construction image sets. The largest, MOCS, has 41,668 real photos. ConSynth-X is the only one in their comparison with each synthetic image tied to a specific source photo.

ConSynth-X next to other construction image datasets, as the authors compare them
Dataset	Tasks	Adverse conditions	Paired with source	Type	Images
ConSynth-X	Object detection, VQA, image captioning, visual grounding	11 subsets across five condition groups	Yes	Synthetic	34,199
ExtCon	Object detection	Five adverse conditions	No	Synthetic	506
SODA	Object detection	Not explicitly categorized	No	Real	19,846
ACID	Object detection, instance segmentation, image captioning	Not explicitly categorized	No	Real	10,000
MOCS	Object detection, instance segmentation	Not explicitly categorized	No	Real	41,668
ConstructionSite 10k	Visual grounding, VQA, image captioning	Illumination and scale/viewpoint	No	Real	10,013

Source: arxiv.org

Evidence so far

The validation answers two questions. The first is whether each altered image still shows the original scene, and the second is whether it looks more like a real bad-weather photo than the original did. Detection accuracy on the new images, which would show how much a site camera loses in fog, is not part of this paper.

For the first question, the authors ran every image through DINOv3, an image model that turns a picture into a list of numbers describing what it shows. They compared each altered image with its source by cosine similarity, a score where 1.0 means the two images show the same content. A source scene stayed in the dataset only if all eight gated versions scored 0.85 or higher.

Source scenes kept at each similarity threshold
DINOv3 similarity threshold	ConstructionSite 10k scenes	SODA scenes	SODA-KTSH scenes	Total scenes kept
0.70	3,995	3,214	2,618	9,827
0.75	3,485	2,545	2,284	8,314
0.80	2,738	1,635	1,694	6,067
0.85 (used)	1,586	608	915	3,109
0.90	406	65	217	688

Source: arxiv.org

At 0.85, 3,109 scenes survived. At 0.90, only 688 would have. The heavy rain, rain at night and snow at night versions were kept without the gate, and their similarity scores ship in the metadata so users can filter them.

For the second question, the authors measured distance to real adverse-weather photos from three public sets, ACDC, WeatherBench and WeatherNet. They added 199 construction-site rain photos retrieved through the Openverse API (a search service for openly licensed images). On the ConstructionSite 10k subset, 81.85% of the altered images sat closer to the real-weather photos than their daytime sources did. Nine of the ten conditions they checked moved in that direction. Rain at night moved the other way.

The table below uses FID, or Fréchet Inception Distance, a score of how different two sets of images look to a standard image-recognition network. Lower means more alike. Each row compares the original daytime photos and the converted photos against the same ACDC reference set.

FID against real ACDC weather photos, ConstructionSite 10k subset (lower is closer)
Condition	Baseline FID (original daytime photos)	FID after conversion	Change
Nighttime	282.2	191.4	-90.8
Heavy snow	217.4	192.0	-25.4
Light snow	217.4	199.6	-17.8
Light rain	199.3	192.2	-7.1
Light fog	220.1	214.4	-5.7
Moderate fog	220.1	214.4	-5.7
Heavy fog	220.1	217.1	-3.0
Rain at night	199.3	203.8	4.5
Snow at night	217.4	226.9	9.5
Heavy rain	199.3	237.6	38.3

Source: arxiv.org

Night conversion made the biggest difference. FID fell from 282.2 for the daytime originals to 191.4. Heavy rain went the wrong way, rising from 199.3 to 237.6, and the authors call heavy rain and precipitation at night inconsistent across reference sets. Apart from the 199 Openverse images, the reference photos come from road and weather collections outside construction.

Two parts of the release disagree with each other, and both matter if you plan to use it:

The paper says the 0.85 gate applied to eight subsets. The GitHub README says nine gated conditions.
The paper says the dataset is deposited on Hugging Face. The README says the full dataset will not be redistributed through GitHub or Hugging Face, and that the Hugging Face page holds metadata and any future access instructions. The Hugging Face listing does show Parquet files for all 11 conditions, behind a gate that requires manual approval.

How to measure it on your projects

The paper’s design, the same scene under different conditions, works as a test on your own footage. If you pilot an AI camera for PPE or exclusion-zone alerts, track detection recall by condition. Recall here is the share of real events a person finds on review that the system also flagged, split into clear day, rain, fog and after dark.

Capture the baseline before relying on the alerts:

Pick one fixed camera on an active job, and record the weather and light for each shift in the daily log.
Each week, have the safety manager review a set amount of footage from a clear shift and the same amount from a rain, fog or night shift, and log every PPE miss and zone entry.
Match the logged events against the system’s alerts. Recall for a condition is flagged events divided by logged events.

If recall after dark runs well below the clear-day figure, those are the hours when a person still has to watch the feed. A first step this week is to add a weather-and-light field to the daily log for the area that camera covers.

Availability and cost

The dataset is free under the Creative Commons Attribution-NonCommercial 4.0 license, and the source datasets’ own terms still apply. The license excludes commercial use, so a vendor would need separate permission to train a paid product on it. The generation code is on GitHub under the Apache 2.0 license, and users have to download the third-party model checkpoints from their own publishers. The work was funded in part by the Farris Family Innovation Award, with computing time from the Ohio Supercomputer Center.

STACK IQ lets estimators work STACK takeoffs and estimates from Claude or ChatGPT

Sat, 26 Sep 2026 07:34:52 -0400

STACK Construction Technologies has introduced STACK IQ, a connection that lets an estimator run tasks in STACK’s takeoff and estimating software by typing requests into Claude or ChatGPT. For Construction Pros listed the release on September 22, 2026. It is aimed at the preconstruction work that follows a finished estimate, such as proposals, schedules of values, vendor leveling sheets and takeoff libraries.

What it does

After an estimate is finished in STACK, the estimator has to carry the same numbers into a proposal and a schedule of values for the owner. For Construction Pros reports that STACK built the tool after customers asked for help with repetitive work of this kind, such as reformatting proposals and building takeoff libraries. With STACK IQ, the estimator asks an AI assistant for those documents in plain English, and the assistant pulls the figures from the STACK project.

The link runs on the Model Context Protocol, an open standard that lets an AI assistant read from and act in another company’s software. The estimator adds STACK from the connector directory in Claude or the plugin directory in ChatGPT, signs in to STACK, and picks which capabilities to allow. STACK says any assistant that accepts custom MCP connectors can connect, including Copilot, and its FAQ says Gemini support is following quickly.

STACK’s prompt library walks through the order of work. The estimator asks for a project by name, and the assistant asks which one if more than one matches. The estimator can then ask it to flag missing unit prices and mistagged items, and to build a proposal broken out by CSI division. The library also has prompts for a schedule of values on the AIA G703 form, a vendor leveling sheet with freight, sales tax and lead time, and a job cost budget with direct costs split from indirects. STACK says some actions wait for the user’s approval before anything changes in STACK.

What STACK IQ can do today, according to STACK's prompt library
Task	Status per STACK
Read your project list and takeoff details	Available
Edit takeoff colors in bulk or by tag or one at a time	Available
Update a project's name or status or bid date	Available
Copy takeoffs into a new or existing library	Available
Read estimate data	Available
Build proposals and schedules of values and vendor sheets	Available
Flag stale snapshots and missing prices and mis-tags	Available
Bulk-edit projects	Not yet available
Delete takeoffs or estimates or projects	Not yet available
Read or write items and assemblies	Not yet available
Upload plans or read plan sheets and graphics	Not yet available
Audit measurements for outliers	Not yet available
Update or write to the estimate	Not yet available

Source: 21266854.fs1.hubspotusercontent-na1.net

STACK IQ cannot yet read plan sheets or audit measurements for outliers, so the estimator still measures the takeoff in STACK. It also cannot read or write items and assemblies, and it cannot write to the estimate. The library adds two smaller cautions. Tags do not carry over when takeoffs are copied into a library, and the assistant works from a snapshot of STACK, so it may miss something created seconds earlier.

Evidence so far

The evidence so far is three short customer endorsements on STACK’s product page, from Gulf Coast Pavers, Turner Brothers and Hemington Landscape Services. Evan Borseti of Turner Brothers says the firm produces estimates, proposals and budgets from one prompt and has made its document formatting identical from job to job. Time savings, error rates and pilot results for STACK IQ are not yet public.

These claims come from STACK:

Tasks that used to take an afternoon now take a sentence.
Pricing can come from your last 50 projects, down to the line item.
STACK IQ lets a veteran estimator do the work of three people.

Read the second claim against the table above. STACK lists reading items and assemblies as not yet available, and how the tool prices from project history without them could not be verified.

How to measure it on your projects

Track document prep hours per bid. Count the estimator hours from the moment the estimate revision is locked to the moment the proposal, schedule of values and any vendor leveling sheet are ready to send. Leave takeoff hours out, since STACK IQ does not measure quantities today.

Capture the baseline before the pilot:

Pick the last 10 bids that went out from STACK.
For each one, ask the estimator to log or reconstruct the hours spent on the proposal, the SOV and the leveling sheet.
Note every correction made after the documents went out, such as a wrong unit price or a missed exclusion.

During the pilot, have the same estimators build those documents through STACK IQ for the next 10 bids and log the same hours. Count post-send corrections too, and compare both figures with the baseline. Record the time spent checking the assistant’s output as prep time, because STACK tells users to review every change before they approve it.

A first step for this week is to have one estimator time the proposal and SOV on the next bid that leaves your office.

Availability and cost

STACK says STACK IQ is included in all paid STACK subscriptions at no extra charge. The user needs an active STACK subscription and access to ChatGPT or Claude, and nothing has to be installed. STACK also runs a user group on its support site, STACK IQ Insiders, where it posts example prompts and known issues.

Autodesk puts a Submittals Evaluation Agent and a Bid Proposal Agent into beta in Forma

Fri, 25 Sep 2026 07:38:04 -0400

Autodesk announced a set of AI agents for construction teams at Autodesk University 2026 in Las Vegas, starting with a September 15 press release. An agent here means AI software that carries out a multistep task inside the application and hands a draft to a person for approval. The Submittals Evaluation Agent checks what a subcontractor submitted against the spec. The Bid Proposal Agent pulls pricing out of subcontractor proposals. Both are in beta inside Forma, Autodesk’s cloud platform for architecture, engineering and construction, which includes Forma Build and Forma Bid. A Daily Log Agent for superintendents is listed as coming soon.

What it does

Start with submittals. Today a submittal manager or project engineer opens each product data sheet and shop drawing, finds the spec section it answers to, and checks the product against it before the package goes to the architect. Autodesk says the Submittals Evaluation Agent compares the submittal attachments against spec requirements and lists potential compliance concerns next to the source evidence. The submittal manager reads the flags and still decides what goes out.

The Bid Proposal Agent is for estimators. It reads subcontractor proposal PDFs and extracts the cost breakdown, the inclusions and the exclusions into structured bid data in Forma Bid, the bid management tool Autodesk is building out of BuildingConnected. Each extracted item stays linked to the PDF it came from, so the estimator can check a number against the proposal and correct it before leveling bids.

The Daily Log Agent is for the superintendent who coordinates trades all day and documents it afterward. The super narrates the day into the Forma mobile app, and Autodesk says the agent transcribes the recording and turns it into a structured daily log.

Construction agents Autodesk announced at AU 2026
Agent	Who uses it	What goes in	What comes out	Status
Submittals Evaluation Agent	Submittal manager or project engineer	Submittal attachments and spec requirements	Potential compliance concerns shown with source evidence	Beta
Bid Proposal Agent	Estimator	Subcontractor proposal PDFs	Cost breakdowns and inclusions and exclusions as structured bid data in Forma Bid	Beta
Document Agent	Anyone on the project team	A plain-language question about project files	Links to matching files	Beta
Sheets Agent	Anyone reading drawings	A plain-language question about the drawing set	The relevant sheet or section	Beta
Daily Log Agent	Superintendent	A spoken update in the Forma mobile app	A structured daily log	Coming soon

Source: autodesk.com

Two drawing tools came with the agents. Drawing Change Analysis, in beta, shows what changed between two revisions of a sheet and where. Drawing Compliance Review compares drawings against rules a team sets for constructability and documentation standards, and Autodesk plans it for early next year.

Evidence so far

The evidence so far is Autodesk’s own description of the agents, which are in beta. Accuracy rates, time savings and results from contractor pilots are not yet public. The customer example in Autodesk’s AEC announcement post, the Kraft Group at Gillette Stadium, is about Forma Data Management.

These claims come from Autodesk:

The Submittals Evaluation Agent helps submittal managers catch issues earlier.
The Bid Proposal Agent cuts the time estimators spend reviewing and retyping sub proposals.
The Daily Log Agent reduces end-of-day re-entry for superintendents.

The status of the Daily Log Agent is unclear. The construction blog labels it “Coming Soon,” while an Autodesk News post by executive vice president Amy Bunszel describes it as available on mobile. Its release date could not be verified.

How to measure it on your projects

Track recall on spec deviations, meaning the share of real spec deviations on a job that the agent flagged. A real deviation is one the architect cited when returning a submittal, or one your own reviewer caught before transmittal. Divide the deviations the agent flagged by all real deviations.

Capture the baseline before the pilot starts:

Pull the submittal log for one active job and take the last 50 closed submittals.
Mark each one the architect returned as revise and resubmit or rejected, and note whether the reason cites a spec requirement.
Record the calendar days from receipt of the sub’s package to transmittal to the architect.

During the pilot, run every new submittal through the agent before your reviewer sees it, and log what the agent flagged. After 50 submittals, compare the return rate for spec reasons and the days to transmittal against the baseline, and compute recall from the flags. If recall stays low, keep your reviewer’s full check in place. Count the false flags as well, since each one takes reviewer time to clear.

A first step you can take this week is to pull that log on one job and tag each returned submittal among the last 50 by its reason.

Availability and cost

The Submittals Evaluation Agent, the Bid Proposal Agent, the Document Agent and the Sheets Agent are in beta. Autodesk says availability varies by subscription and that its AI features are open only to value-based subscription types, which include bundles and enterprise business agreements. The Forma Bid beta is open to customers with a Forma for Preconstruction or BuildingConnected Pro subscription, according to Autodesk News. The announcements do not list a separate price for the agents. Autodesk tracks rollout dates on its Forma Product Roadmap.

A broader next-generation Autodesk Assistant, which would work across Autodesk products and projects, was also previewed at AU. The press release says it is not yet generally available and is planned to begin rolling out in 2027.

Suffolk and MIT model 17% to 20% AI cost savings on one multifamily project

Thu, 24 Sep 2026 19:45:00 -0400

Suffolk, the MIT Center for Real Estate and the MIT Media Lab City Science group published “Construction in the Age of AI: An Industry White Paper and Research Roadmap” on September 16, 2026. Its headline figure comes from a model Suffolk built on one multifamily project in San Francisco. By that model, AI applied across six parts of project delivery would have cut total cost by 17% to 20% and total schedule by 22% to 25%. The workflows in scope run from permit applications and drawing review through schedules, RFIs, submittals and procurement. The authors describe their evidence as “suggestive rather than definitive,” and the methodology appendix says it cannot yet support causal conclusions.

What it does

The paper sorts AI uses into six domains, which the authors call levers: design automation, offsite manufacturing, permitting, scheduling, skilled labor and subcontracting, and supply chain and procurement. For each one it collects reported savings from a literature review of 178 sources, plus case studies, expert interviews and survey responses. A roundtable held on March 11, 2026, with more than 50 industry participants, settled the list. The acknowledgments name participants from Procore, Autodesk, Oracle and Bluebeam.

The percentages in the table below are the maximums the authors found for each domain. Most come from early-stage pilots, and the sources include industries next to construction. A footnote in the paper says to read them as directional, and warns that they are not benchmarks validated on construction projects. The design automation figures also apply to a narrower base than the others, as the last column shows.

Maximum reported gains by AI domain, from the white paper's literature review
Domain	Example AI application	Time reduction up to (%)	Cost savings up to (%)	What the percentages apply to
Design automation	Generative design and BIM AI	39%	21%	Design cycle time; design and engineering cost
Offsite manufacturing	AI-optimized factory prefabrication	32%	14%	Project timeline; total cost
Permitting	Automated code-compliance review and application preparation	24%	8%	Project time; cost
Scheduling	ML-driven sequencing and optimization	18%	11%	Project timeline; total cost
Skilled labor and subcontracting	AI trade development, coordination and performance management	17%	13%	Project time; total cost
Supply chain and procurement	Predictive lead time, delivery optimization and process automation	15%	10%	Total project time; total cost

Source: suffolk.com

Suffolk then took one completed job and estimated what the six levers would have done to it together. Section 5 of the paper describes a 180,000-square-foot multifamily development finished in 2024. Table 1 breaks the result down by phase.

Suffolk's model of AI savings on the sample multifamily project, by phase
Phase	Baseline cost ($M)	Modeled cost savings ($M)	Cost savings (%)	Baseline time (months)	Modeled time savings (months)	Time savings (%)
Acquisition	27	0		3	0
Entitlement	1.4	0.05	4%	11	2	19%
Design	6	1.2	20%	24	7	30%
Preconstruction	3	0.3	10%	9	1	14%
Construction	128	28	22%	30	7	14%
Finance	7	1.2	17%
Cost escalation	7	1.2	17%
Total	179	32	18%	51	11	22%

Source: suffolk.com

Start at the bottom row. The model takes $32 million off a $179 million baseline, which is 18%, and 11 months off a 51-month timeline, which is 22%. Most of the dollars come from construction, where $28 million of the $32 million sits against a $128 million phase. Design has the largest schedule cut, 7 of 24 months. The model leaves acquisition untouched. The months do not add to the total because phases overlap, and the paper notes this under its table.

The authors then run the savings through to developer returns. By Suffolk’s internal model, unlevered IRR, the project’s return before any debt, rises about 5 to 6 percentage points, for example from 15-20% to 20-25%. Yield on cost rises 1 to 2 points.

The readers who get the most from the paper are preconstruction leads, VDC and scheduling managers, and CFOs who have to justify an AI budget. The domain sections name tools already sold to contractors. They cite ALICE Technologies and nPlan for schedule optimization, OpenSpace for tracking installed work against the plan, and TrunkTools’ TrunkRFI agent for flagging unnecessary RFIs and drafting needed ones. Procore’s in-platform agents come up for RFI deflection and drafting, and LightTable and BuildCheck for drawing review. (The paper identifies nPlan as a Suffolk Technologies portfolio company.)

Evidence so far

The 17% to 20% savings come from the model, and the appendix calls the paper a methodology statement. It lists what a field test would need: project-level data covering several hundred projects, and comparisons between AI-assisted and conventionally managed jobs that control for complexity, geography, delivery method and owner type.

The per-domain figures come from other parties, and the paper attributes them. Volumetric Building Companies reports up to 47% schedule savings from offsite manufacturing. The paper cites a Deloitte estimate that AI-assisted estimating cuts budget and timeline deviations by 10% to 20% and engineering hours by 10% to 30%. In its passage on ALICE Technologies, the paper says AI planning tools cut costs by up to 30% on one large solar farm, and it does not name the project or its baseline.

Suffolk calls its model a first pass, and some figures in the paper disagree with each other. If you plan to quote the paper in a budget request, check these first:

Baseline cost. The footnote under the summary table anchors survey responses to a $180 million project. Figure 5 labels the sample project $150 million. Table 1 totals $179 million.
Duration. Section 5 gives 4.5 years from initial design to closeout. Table 1 gives 51 months, which is 4.25 years.
Scheduling savings. The summary table gives up to 11% cost savings for scheduling, and the scheduling section gives 8%. A footnote there says the model uses a 15% schedule reduction, where the literature reports up to 18%.

For a contractor choosing where to start, the expert ranking says more than the model. Participants scored each lever by Borda count, a ranking method that awards points by where an item falls on each voter’s list. Design automation collected the most points, 206, but 89 of them came from the design group voting for its own lever. With each group’s votes for its own lever removed, scheduling ranked first with 163 points. The skilled labor and subcontracting group spread its points almost evenly, between 19 and 26 per lever, which the authors read as a sign that field productivity depends on work done upstream.

Borda points by expert group (rows) for each AI lever (columns)
Expert group	Design automation	Offsite manufacturing	Supply chain	Skilled labor and subcontracting	Schedule optimization	Permitting
Design automation	89	51	45	33	69	41
Offsite manufacturing	20	29	18	14	15	9
Supply chain	27	16	16	20	28	12
Skilled labor and subcontracting	24	19	26	23	23	20
Schedule optimization	26	12	28	16	28	10
Permitting	20	14	16	20	28	37

Source: suffolk.com

How to measure it on your projects

So the schedule is where to start measuring. Track forecast finish slip, defined as the calendar days between the baseline substantial completion date and the forecast date in each monthly schedule update. If an AI scheduling or progress-tracking tool helps, slip on the pilot job should grow more slowly than on comparable jobs without it.

Capture the baseline before the pilot starts. Pull the monthly updates from three to five completed jobs of similar type and size, and record the slip at each update and at completion. The white paper asks owners to fund pilots with outcome measurement written into the project brief from day one, and a contractor’s own trial needs the same setup.

A first step on one job: export every monthly update from the P6 or Microsoft Project file for your most recently completed project, and chart the forecast completion date by month in Excel or Power BI.

Availability and cost

The white paper is free, as a PDF linked from Suffolk’s news page. The MIT research team is designing data collection instruments for the next phase. Firms that want to contribute project data, take part in surveys or join the research advisory panel can register through the MIT Center for Real Estate. The paper gives no date for that phase.

The Impact of AI on Construction: Pros and Cons Revealed

Tue, 09 Jul 2024 17:32:01 +0000

As artificial intelligence continues to transform industries across the board, the construction sector is no exception. We’ve been exploring AI’s potential to revolutionize our workflows. But like any powerful tool, AI comes with both benefits and challenges. Let’s break down some key pros and cons we’ve encountered:

Pros

Document Intelligence

By leveraging AI with Retrieval-Augmented Generation (RAG) on our SharePoint sites, we’ve unlocked a treasure trove of institutional knowledge. AI can now quickly sift through years of project reports, safety guidelines, and best practices to provide relevant insights on demand.

Predictive Maintenance:

AI algorithms play a crucial role in predictive maintenance, enabling us to proactively identify potential issues with machinery and equipment. By analyzing historical data and real-time performance metrics, these advanced algorithms can predict equipment failures before they occur, thereby minimizing unplanned downtime and optimizing operational efficiency. This proactive approach not only saves time and resources but also enhances safety by preventing potential hazards on job sites. By leveraging the power of AI, we can ensure that our equipment remains reliable and safe, ultimately contributing to a more productive and secure working environment.

Design Optimization

Generative AI plays a pivotal role in assisting our architects and engineers by leveraging advanced algorithms and machine learning techniques to explore innovative design solutions. By harnessing the power of Generative AI, we are able to optimize designs not only for cost, sustainability, and structural integrity but also for energy efficiency, aesthetic appeal, and user experience. This cutting-edge technology empowers us to push the boundaries of creativity and efficiency, ultimately leading to the development of forward-thinking and sustainable architectural and engineering solutions.

Cons:

Data Hallucinations

We’ve encountered instances where AI confidently presents false information as fact. This “hallucination” problem requires constant human oversight to catch and correct. It’s crucial for AI developers to implement robust fact-checking mechanisms to ensure that the information generated is accurate and reliable. Additionally, ongoing research and development in the field of AI ethics and accountability are essential to address these challenges. As AI technology continues to advance, the need for human involvement in validating the output becomes increasingly evident. Finding the right balance between autonomous AI capabilities and human intervention is a key area of focus for ensuring the responsible and effective deployment of AI systems.

Inconsistent Calculations

While AI excels at processing vast amounts of data, we’ve found its mathematical accuracy can be surprisingly unreliable for critical structural calculations. This has prompted further research into enhancing the accuracy of AI algorithms for specialized tasks such as structural engineering. The complexity and nuances of structural calculations require a more robust and nuanced approach, considering not only the mathematical precision but also the real-world implications of the results. As AI continues to evolve, it is crucial to address these limitations and refine the algorithms to ensure their reliability in critical scenarios.

Over-reliance Risks

There’s a danger of team members becoming too dependent on AI, potentially eroding crucial problem-solving skills and domain expertise. As AI technology continues to advance and automate various tasks, there is a risk of employees relying excessively on AI for decision-making, analysis, and creative problem-solving. This over-reliance on AI may lead to a decline in critical thinking abilities, innovation, and hands-on experience within specific domains. It is essential for organizations to foster a balance between leveraging AI tools for efficiency and preserving the cognitive capabilities and specialized knowledge of their teams.

Implementation Challenges

Integrating AI systems with our existing software stack and training our workforce to use them effectively has been a significant undertaking. Not only have we had to ensure seamless compatibility between the new AI components and our current infrastructure, but we have also dedicated substantial resources to educate and empower our employees to leverage these advanced tools. This comprehensive effort has involved restructuring internal processes, providing specialized training programs, and fostering a culture of continuous learning and adaptation. As we continue on this journey, we are seeing the transformative impact of AI on our operations, decision-making processes, and overall business performance.

The Path Forward

Despite these challenges, we believe AI’s potential in construction is immense. The key lies in thoughtful implementation:

Establish clear processes for human verification of AI outputs
Invest in ongoing AI literacy training for all employees
Collaborate with AI providers to address industry-specific needs
Use AI as a complement to, not a replacement for, human expertise

By embracing AI’s strengths while mitigating its weaknesses, we can build safer, more efficient, and more innovative projects for our clients.

In an upcoming post, we’re going to go through the exact strategies we took to overcome the challenges presented here. Stay tuned.

What has your experience been with AI in construction? I’d love to hear your thoughts in the comments.

Share of project complete	PPC managed traditionally (%)	PPC managed with Buildots (%)
35%	27	52
40%	27	57
45%	27	71
50%	27	37
55%	27	32
60%	31	57
65%	22	72
70%	27	81
75%	27	90
80%	22	96

Share of project complete	PPC managed traditionally (%)	PPC managed with Buildots (%)
35%	27	52
40%	27	57
45%	27	71
50%	27	37
55%	27	32
60%	31	57
65%	22	72
70%	27	81
75%	27	90
80%	22	96

Share of project complete	PPC managed traditionally (%)	PPC managed with Buildots (%)
35%	27	52
40%	27	57
45%	27	71
50%	27	37
55%	27	32
60%	31	57
65%	22	72
70%	27	81
75%	27	90
80%	22	96