AI jobsite safety tools are spreading faster than proof they prevent injuries. What a GC should measure
A 2,500-worker rollout and vendor claims of up to 48% fewer incidents sit beside research that finds little evidence of injury prevention. How to set up a safety AI pilot that produces an answer.
AI hazard-spotting tools for construction are being rolled out to thousands of workers, but the evidence that they prevent injuries is still thin. A GC or sub piloting one should measure injuries, near misses and corrected hazards, not just how often the app gets opened.
What happened?
CNBC reported on Oct. 31 that AI is moving into construction safety work: software that reviews photos and short videos from the field, flags hazards, and drafts reports. FYLD, a London-based vendor, says its case studies show incident and injury reductions of up to 48%. Amey, a UK infrastructure services firm, ran a three-month FYLD pilot and plans to roll the tool out to more than 2,500 field workers in its Highways division.
The researchers CNBC interviewed were more cautious. Waleed Umer of Northumbria University said most studies test whether AI can detect hazards, not whether it prevents harm. Billy Hare of Glasgow Caledonian University said pilots mostly measure productivity and often assume less hazard exposure means lower risk, a link he says needs more investigation.
What did the Amey pilot actually prove?
Adoption, not injury prevention. Per TokenPost's summary of the pilot:
| Measured | Result |
|---|---|
| Jobs logged | More than 500 |
| Jobs with an AI-enabled assessment | 95% |
| Users consistently active | 85% |
| Injury rates | Not reported |
Those are good numbers for a tool people will actually use. They don't show a safer jobsite. The same report notes a 2025 systematic review of workplace AI safety research that screened 1,255 articles and found only two eligible studies, rated medium and low quality.
Fewer deaths don't settle it either. Great Britain recorded 25 construction worker deaths in 2025/26, down from 35 the year before. Hare told CNBC the record low may partly reflect an economic slowdown, which makes the technology's effect hard to isolate.
What are the risks of relying on it?
The CNBC piece names three:
- False alarms. Too many flags and crews learn to ignore them.
- False comfort. A tool that checks a photo can make a super feel covered when the hazard is outside the frame.
- Review overload. If the AI generates more reports than supervisors can read, recommendations get approved without scrutiny.
How should a GC set up a safety AI pilot?
The upside is real. CNBC cites field workers spending 20–25% of site time on admin, and a report that took about half an hour drafted in two to three minutes. Time back to the super is worth something. But decide in advance what you'll do with that time and what result would count as success.
- Set a baseline. Pull 6–12 months of near-miss reports, recordables and corrective-action closeout times before turning anything on.
- Pick outcome metrics. Near misses reported, hazards corrected and how fast, first-time fixes, recordable incidents. Usage rate is a side metric.
- Use a comparison. Run the tool on some projects or crews and not others of similar type and size, if you can.
- Log false alarms. Have the super mark each flag as real, wrong or already known. If most are noise, adoption will decay.
- Keep a human signature. Supervisors still review and sign off. Umer's framing: "an extra pair of eyes, not a replacement for supervision."
- Check the contract terms. Ask who owns the site photos and video, whether workers appear in them, and how long the vendor keeps them.
What's the takeaway?
Treat vendor injury-reduction figures as claims to test on your own jobs. A pilot that tracks only logins will come back looking successful whether or not anyone is safer. Define the safety outcome first, hold the vendor to it, and keep your supervisors in the loop.
- Does AI actually reduce construction injuries?
- That is not yet established. Researchers quoted by CNBC say most studies test whether AI can detect hazards, not whether it prevents harm. Vendor case studies report reductions of up to 48%, but those are vendor-reported figures.
- What did the Amey and FYLD safety pilot measure?
- It measured tool use, not injuries. The three-month pilot logged more than 500 jobs, with an AI-enabled assessment recorded for 95% of them and 85% of users consistently active. Injury rates were not reported.
- What should a contractor track in a jobsite safety AI pilot?
- Track outcomes, not just usage: near-miss reports, corrected hazards, first-time fixes, and recordable incidents compared with a baseline period or a comparable crew without the tool. Also track false alarms, since too many train crews to ignore warnings.
- Can AI replace a safety supervisor on a jobsite?
- The researchers quoted advise against it. Northumbria University's Waleed Umer said to treat AI as an extra pair of eyes, not a replacement for supervision.