Provision Launches Scope Agent with 97% Accuracy on Automated Construction Scope Extraction

Provision, the AI construction technology company, today announced the general availability of Scope Agent, its AI system for automated scope-of-work generation from drawings and specifications. Alongside the launch, Provision released a benchmark report documenting Scope Agent’s performance against Claude, the general-purpose AI model from Anthropic, on plumbing scope extraction from a live construction project.

Missing or ambiguous scope is one of the largest hidden costs in construction. The Construction Industry Institute (CII), based at the University of Texas at Austin, has found that rework costs run from 2% to 20% of a project’s total contract amount, with direct costs averaging roughly 5% of total construction cost. CII estimates rework represents approximately $15 billion annually in direct costs for the U.S. construction industry alone. Design errors and omissions — the category scope gaps fall under — are consistently identified as a leading cause.

On a large commercial project, a single missed scope item can expose a general contractor to hundreds of thousands of dollars in unplanned cost. The problem compounds across a portfolio.

“This is one of the highest-risk activities if you’re a bidding GC,” said Michael DeMarie, President of ProWest Constructors, a California general contractor and Scope Agent customer. “If you don’t get it right, it stays with you from the day you start the job to the day you finish. If you want arguments from the day you start beyond the day you finish, then don’t scope very well.”

On Provision’s internal validation dataset, measured against human-prepared scope sheets on real projects from general contractors across North America, Scope Agent delivered:

• 97.01% match accuracy on scope item extraction
• 96.02% subcontractor classification accuracy (Top 2)

For reference, Provision ran the same exercise with a professional estimator working to a fixed time budget. The human baseline came in at 91.3% accuracy — and took roughly four days per project. Scope Agent returns a first pass in a fraction of that time.

Provision’s new benchmark report, Plumbing Scope QA: AI Scope Agent vs. Claude, compares both systems on the same set of construction drawings from a live hospital project. Reviewers validated every extracted line item against the source drawings.

• Scope Agent extracted 145 plumbing scope items at 91.7% verified accuracy. Claude extracted 96 items at 72.9% accuracy.
• Scope Agent captured 51% more scope items than Claude, materially reducing downstream scope-gap risk.
• Scope Agent’s accuracy was 18.8 percentage points higher than Claude’s on the same dataset.
• Claude’s dominant failure mode was fabrication. 18 of its 26 flagged rows involved invented drawing marks, manufacturer model numbers, or schedule items that do not appear in the source drawings. Scope Agent produced zero fabricated items.

Provision attributes the gap to three design decisions:

• Drawing element identification. Scope Agent decomposes drawings element by element rather than reading them as flat pages, allowing it to find information at far higher fidelity.
• Deep linking. Scope Agent builds a graph of relationships between drawing elements, surfacing conflicts and scope errors rather than silently absorbing them into a line item.
• Estimator intelligence. Estimating business logic is built into the system, so the output mirrors how an estimator actually reasons about a package — reducing errors and the configuration required to produce a usable scope.

The full report is available at getprovision.co.

“For the teams we work with, the question was never whether AI could read a drawing. It was whether the output could be trusted on a real bid,” said Luigi La Corte, CEO and founder of Provision. “A 97% accurate scope sheet with full traceability is a document you can put in front of a subcontractor. A 73% accurate scope sheet with fabricated model numbers is a liability. That distinction is what Scope Agent was built to solve.”

“Scoping has always had elements of both art and science, so every estimator ends up doing it their own way,”said Michael DeMarie, President of ProWest Constructors, a California general contractor and Scope Agent customer. “A product like this gives us a protocol, and it gives us consistency. I’ve seen roughly a 50% savings in the time from award to all the subcontracts being written. On an average $25 million project, going from award to every subcontract issued in 60 days — that’s a really good target for us.”

The first project ProWest ran through Scope Agent was a county job in Santa Barbara that a senior estimator had already begun scoping by hand.

“It saved him three or four days on a project that size, and we noticed four or five items we had missed — we had misinterpreted some of the sheet metal requirements,” said DeMarie. “It was impressive how quickly it did it and how much information was there. It was good enough for me to say, okay, we’re going to incorporate this.”

Scope Agent has been used on live tenders with several of the largest general contractors in Canada, including EllisDon on the New Brunswick Museum project, where the output was used to issue a formal addendum modifying subcontractor scopes of work. That is the first documented case of AI-generated scope output flowing directly into a bid package at a top-tier general contractor.

“The value of Scope Agent is that it catches what a tired estimator will miss at 11 p.m. the night before a bid closes,” said La Corte. “Our customers are not looking to replace their senior estimators. They are looking to give them a tool that removes the fear of a missed scope item.”

Scope Agent is available today for general contractors, construction managers, and trade contractors. Teams can book a demo at https://provision.com/request-a-demo