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

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.

Item 1Procore AI checks a submittal through three layers

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.

1. 1

   Multimodal index

   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.

2. 2

   Construction ontology

   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.

3. 3

   Magpie 3.0

   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.

Item 2Procore says Magpie 3.0 passed 78% of its 200-plus tasks

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%.

In plain terms, on a test built from questions that field and office staff ask, Procore's system got about four tasks in five fully right and the best general model about three in five. Procore says its internal experts judged that some of the hard and expert tasks would take a seasoned professional hours to answer.

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.

“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.

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.

Item 3Measure review days on 30 to 50 closed submittals before the pilot

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:

1. Pull the last 30 to 50 closed submittals on a comparable job and record the received and returned dates for each.
2. Count the revise-and-resubmit responses, and mark the ones whose reason came from a spec section outside the package's own division.
3. 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.

Item 4Magpie 3.0 and AI Go are in open beta in October 2026

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.

DataEvery figure on this page comes from Procore