TL;DR: AI is genuinely good at noticing things on a jobsite. A peer-reviewed study got 96% accuracy detecting missing hard hats and vests. What almost nobody can show you is evidence that noticing reduces injuries, and falls still account for over a third of construction deaths. The strongest case for AI here is not safety at all. It is documentation.
The one-sentence answer: AI reliably detects conditions in images and reliably assembles records, and the distance between detection and prevention is where most safety-AI marketing lives.
Safety is the easiest thing to sell a contractor and the hardest thing to prove.
Every vendor in this category shows the same demo: a camera spots a worker without a hard hat, a box appears around them, an alert fires. It works. It is also the beginning of the argument rather than the end of it.
So this post separates three things that get sold together and perform very differently: safety monitoring, quality control, and documentation. One of them is a clearly good buy today, and it is not the one with the compelling demo.
What can AI actually detect on a jobsite?
Personal protective equipment, with accuracy that has been peer-reviewed rather than merely claimed.
Researchers at the Indian Institute of Technology Bombay published a study in Frontiers in Built Environment in September 2020 testing YOLOv3 deep learning on construction site imagery. Across three datasets it reached 96.27%, 96.51% and 96.92% accuracy, with F1 scores of 0.96 to 0.97, classifying images as safe, not safe, no hard hat, or no jacket. The dataset comprised 2,509 images and 4,132 data points.
That is a 2020 result and should be read as a legacy benchmark; detection models have improved since. The reason it remains the useful citation is that its stated limitations are structural rather than temporary, and they are the limitations vendors do not mention.
What breaks it?
Three things the authors named, and they matter more than the accuracy figure.
Processing ran at roughly two frames per second. Occlusion degraded performance, meaning a worker partly hidden behind material or another worker. And colour mismatching caused problems, particularly for hard hats.
Sit with the frame rate. A fall to a lower level takes about a second. At two frames per second the system may capture the event and cannot possibly intervene in it. That is not a criticism of the research, which was measuring detection. It is a caution about the inferential leap from "the model can tell whether someone is wearing a hard hat" to "the model makes your site safer."
Occlusion is the harder problem in practice. Construction sites are visually cluttered by definition. The moments of highest risk are frequently the moments of highest clutter.
Does detecting hazards actually prevent injuries?
Nobody has shown convincingly that it does, and that gap should shape what you pay.
Detection accuracy is well studied. Independent, peer-reviewed, longitudinal work linking computer vision monitoring to reduced injury rates or claim severity is scarce. Vendor case studies exist and report large improvements; they are vendor case studies, usually without control groups, and frequently confounded by the fact that a company installing safety cameras is a company already paying attention to safety.
This is not an argument against the technology. It is an argument against paying for outcomes on the strength of demos.
The scale of what would need improving is worth stating precisely. CPWR, the Center for Construction Research and Training, drawing on the Bureau of Labor Statistics Census of Fatal Occupational Injuries, reports that falls alone account for over a third of construction deaths in an average year, and that the Focus Four categories together cause almost two thirds. Nationally, OSHA records 5,283 fatal work injuries across all industries in 2023, a rate of 3.5 per 100,000 full-time equivalent workers.
One methodological caution if you go looking at this data yourself. CPWR notes that changes introduced with OIICS 3.0 mean pre-2023 Focus Four estimates are not comparable with post-2023 estimates. Anyone showing you a smooth multi-year safety trend line across that boundary is showing you an artefact.
What does AI do for quality control?
It finds the deviation you would have found eventually, earlier.
The practical applications are narrower than the pitch. Comparing installed conditions against a drawing to flag what does not match. Reviewing photo sets for missing documentation before a pour. Checking that a checklist was actually completed rather than clicked through.
That last one is unglamorous and it is where we have put our own effort. Our field platform treats inspection checklists as structured records with their own threat model, because the integrity question is not whether AI can read a checklist. It is whether the checklist reflects what happened. We also wrote threat models for clock-in geofencing and for the time-of-check-to-time-of-use gap in clock-in, because a record that says a person was somewhere at a time is only a safety record if it cannot be trivially falsified.
None of that is AI. It is the substrate AI needs. A model summarising unreliable records produces confident unreliable summaries.
Why is documentation the strongest case?
Because the output is verifiable and the failure is visible.
If AI drafts your daily report from field notes and photos and gets it wrong, you can see that it is wrong. The source material is right there. Compare that with a safety system that fails to flag a hazard, where the failure is silent and you learn about it from an incident.
Verifiability is the entire reason documentation is the better first purchase. It is the class of AI work where a human check is cheap and immediate.
The concrete wins are dull and real: assembling a daily report from notes and photos, finding the photograph from four months ago that shows the condition before it was covered, organising inspection evidence so it can be produced when someone asks, and turning a long thread into something a project manager reads in a minute.
Contractors appear to recognise this. In the AGC and NCCER 2025 Workforce Survey of roughly 1,400 firms, 44% expected robotics and AI to improve the quality of construction jobs and make workers safer and more productive, against 45% expecting benefit from automating manual, error-prone tasks. The expectation is assistance, not autonomy.
What should you ask a safety AI vendor?
Four questions, and the last one is the one they will struggle with.
What is the frame rate and what happens during occlusion. Ask for the number, not a reassurance.
What is your false positive rate in a cluttered environment. High false positives are how these systems die: alerts get ignored, then muted, then the subscription gets cancelled.
Who receives the alert and what are they expected to do in the next sixty seconds. An alert with no defined response is a log entry.
And: can you show me independent evidence, not your own case study, that this reduces incidents. Most cannot. A vendor who says so plainly is more trustworthy than one who produces a confounded testimonial.
What is a sensible first purchase here?
Documentation, then quality control, then safety monitoring last.
That order inverts how the category is sold, deliberately. Documentation has verifiable output and immediate time savings. Quality control has verifiable output on a slower cycle. Safety monitoring has the strongest emotional pull, the weakest independent evidence, and the highest chance of being switched off within a year because of alert fatigue.
If your motivation is genuinely safety rather than paperwork, the honest advice is that the highest-return spending in construction safety remains unglamorous: fall protection equipment, competent-person training, and enough schedule that people are not rushing. AI has not displaced any of that and is not close to it.
Frequently Asked Questions
Can AI monitoring create liability if it detects a hazard nobody acted on?
This is a genuine question for your counsel rather than a settled one, and it is the reason to define the response protocol before deployment. A system generating a record that a hazard was identified, with no record of anyone responding, produces documented awareness without documented action. Decide who owns alerts and how responses are logged before the first camera goes up.
Do workers accept camera-based monitoring?
Variably, and it depends almost entirely on whether the stated purpose is safety or surveillance. Systems framed and operated as PPE compliance tools tend to be tolerated. Systems that also produce productivity metrics tend to be resented and worked around, which degrades the safety function too.
Does this work in poor lighting or bad weather?
Worse, and the research is explicit that lighting and occlusion degrade performance. Anyone selling into night work or enclosed spaces should be asked for performance figures under those specific conditions rather than general accuracy claims.
Will AI documentation satisfy an OSHA inspector or an insurer?
The record needs to be accurate and attributable, and how it was assembled matters less than whether it holds up. The risk with AI-drafted records is unreviewed output containing plausible errors. Keep human sign-off on anything that could become evidence.
Is there a version of this that works for a small contractor?
Yes, on the documentation end. Photo organisation and report drafting scale down to a handful of crews and do not require integration. Camera-based site monitoring generally does not scale down, because the fixed cost of installation and the ongoing cost of someone triaging alerts do not fall with company size.
If your interest here is fewer incidents, start by asking any vendor for independent evidence rather than a case study. If your interest is getting evenings back from paperwork, that is a smaller and much better-supported purchase.
See ClickWerxs AI services or get in touch. Related: what AI actually does for a construction business and how to evaluate an AI vendor.
Sources
- Delhi, Sankarlal and Thomas (Indian Institute of Technology Bombay), "Detection of Personal Protective Equipment (PPE) Compliance on Construction Site Using Computer Vision Based Deep Learning Techniques," Frontiers in Built Environment, 24 September 2020 — YOLOv3 with transfer learning; 96.27% validation, 96.51% test and 96.92% novel-dataset accuracy with F1 scores of 0.96 to 0.97; 2,509 images and 4,132 data points; stated limitations of approximately two frames per second processing, occlusion, and colour mismatching particularly for hard hats. Legacy benchmark: detection performance has advanced since publication; the stated limitations are structural. frontiersin.org
- CPWR — The Center for Construction Research and Training, Construction Chart Book, Focus Four injuries, drawing on the Bureau of Labor Statistics Census of Fatal Occupational Injuries 2011–2024 — falls alone account for over one third of construction deaths in an average year; Focus Four categories together cause almost two thirds of construction fatalities; OIICS 3.0 changes mean pre-2023 estimates are not comparable with post-2023 estimates. cpwr.com
- Occupational Safety and Health Administration, Commonly Used Statistics — 5,283 fatal work injuries in 2023 across all industries, a rate of 3.5 per 100,000 full-time equivalent workers. osha.gov
- Associated General Contractors of America and NCCER, 2025 Workforce Survey Analysis — approximately 1,400 firms; 44% expect robotics and AI to improve job quality and make workers safer and more productive; 45% expect benefit from automating manual, error-prone tasks. agc.org
- ClickWerxs field operations platform — inspection checklists implemented as structured records with a dedicated threat model; separate threat models for clock-in geofencing and for the time-of-check-to-time-of-use gap in clock-in; safety module and end-to-end safety test suite present as of 17 June 2026. First-party operator data.
Third-party figures are attributed with dates and sample sizes; the PPE detection study is labelled as a legacy benchmark. Nothing here is safety, legal, or compliance advice. Fatality and injury data are population statistics and say nothing about risk on any specific project. Consult a qualified safety professional and your counsel before relying on any monitoring system as part of a compliance programme. ClickWerxs sells AI implementation services and earns revenue from those engagements. This post reflects operator opinion.
Kaleb Dickhaut — Founder, ClickWerxs. Kaleb built ClickWerxs from the ground up, from payment processing ISO to the Command Center platform to the AI SEO methodology the blog runs on. He has onboarded hundreds of small businesses onto payment and CRM systems. linkedin.com/in/kaleb-dickhaut