68% of construction executives have spent $250K+ on AI, and most can't measure the return. Here's the scorecard to build first

Construction firms are spending real money on AI without a shared way to tell whether it paid off. A BRG survey of more than 100 construction and development executives found 68% have spent at least $250,000 on AI in the past 24 months, yet only 15% say AI is fully integrated across their business. The fix for a contractor is cheap: write down a baseline number before any pilot starts.

What did the BRG survey find?

BRG published the research in mid-October 2026 under the title "AI in Construction: Measurable Results, Unclear Benchmarks." The main figures, as reported in BRG's release and a news write-up of it:

BRG also reports that gains show up more often as effectiveness than as efficiency, and that double-digit cost savings are common across project phases. But no single yardstick has emerged to capture the full value.

Does this contradict the "AI replaces jobs" story?

For this sample, yes. More firms added people than cut them. That fits what shows up on jobsites: AI tools draft, summarize and check, and a person still signs the submittal, the RFI response and the pay application.

One caution. This is a survey of executives, and BRG's respondents skew toward firms big enough to spend six or seven figures. Intuit's QuickBooks report, using a different sample of mostly small businesses, found only 7% of US construction firms currently pay for AI tools. Both can be true. If you run a 30-person sub, the BRG spending numbers describe someone else's budget.

Why can't firms measure the return?

Because most pilots start without a baseline. A PM tries a tool for spec review, it "feels faster," and three months later nobody remembers how long the old process took. Effectiveness gains, such as fewer missed scope items, are even harder to see after the fact.

What should a GC or sub do before the next pilot?

Pick one task and one number, and record it before you turn the tool on:

1. Submittal review: average hours per submittal and number of resubmittals.
2. RFIs: days from question to response, and share returned for missing information.
3. Estimating: takeoff hours per bid and number of addenda items missed.
4. Daily logs: minutes per log per superintendent, and share completed same day.

Then set a review date at 60 to 90 days, name one owner, and decide in advance what result means "keep paying" and what means "cancel." Without that, the license renews by default.

What about the contract and liability worries?

Two in five respondents flagged unclear treatment of AI in contracts, and more than four in ten flagged liability for AI-generated errors. Neither is solved by buying software. Ask your attorney whether your subcontract and owner agreements say who is responsible for work product drafted with AI, and keep a human reviewer's initials on anything that goes out under your name.

The takeaway

Don't read the 68% as proof AI pays. Read it as proof that spending is ahead of measuring. Before the next tool trial, write one baseline number on one page, and check it on a calendar date. That is the only ROI study most contractors will ever need.