The honest answer is that AI is currently useful for the administration of safety rather than for safety itself. It drafts, it summarises, it spots repetition across records. Controlling a hazard is a physical act, and the legal duty to see it done sits on named people.
Key takeaways
There are two conversations about AI and construction safety, and they barely overlap. One is about cameras and sensors watching a site. The other is about the paperwork a safety programme generates — job safety analyses, toolbox talks, inspection forms, training records, incident reports. The second is where AI is genuinely useful today, and it is the one nobody markets.
Start with why that ordering is right. The hazards on a Canadian construction site are known, and the requirements addressing them are prescriptive. Ontario states that falls are the top cause of critical injuries and deaths of workers at construction sites, and the provincial guidance on construction hazards sets out specifics: guardrails must be used unless it is not reasonably possible to install them, with thresholds including a fall of more than 3 metres, or more than 1.2 metres where the work area is used as a path for a wheelbarrow or similar equipment, or a fall into operating machinery, water or other liquids. A guardrail system is also required where a worker is exposed to a fall of 2.4 metres or more and has access to the open side of a floor, bridge, roof, scaffold platform, runway or ramp.
Those numbers are the point. They are not general principles a model can reason its way to — they are regulation text, they differ by province, and paraphrasing them is how a site ends up out of compliance with a document that reads well.
A job safety analysis breaks a job into steps, identifies the hazards at each step, and determines controls. CCOHS sets out the basic steps: select the job to be analysed, break it into a sequence of steps, identify potential hazards, determine preventive measures, and communicate the results. It also notes that each analysis needs revision whenever equipment, materials, processes or the environment change — which is the reason most JSAs go stale.
Feed a model the task breakdown and it drafts a credible first pass at steps and candidate hazards, including the ones people forget because they are routine. That draft then goes to the people CCOHS says should be doing the analysis — a small team of experienced workers and supervisors, with health and safety committee involvement, who often have a legal obligation to participate. The model wrote the dull majority of it; the crew corrected what was wrong or missing, which was always where the value sat.
The failure mode of toolbox talks is genericness — a printed sheet about ladder safety on a day nobody is on a ladder. Given the week’s schedule and the hazards from the relevant JSAs, a model will draft a short, specific talk for tomorrow’s activity. WorkSafeBC publishes toolbox meeting guides and construction resources that are a better starting corpus than anything a model invents.
CCOHS lists workplace inspections, job safety analyses, incident investigations, documents and records, and hazard reporting by employees among the ways hazards are identified. Most contractors have years of these on file and have never read them as a set. A model can: summarise every inspection finding from the last two quarters, group them, and show that the same three deficiencies recur on the same two sites. That is hazard identification from data you already paid to collect.
CCOHS is clear that the emphasis in an incident investigation is finding the root cause so the event does not recur, that the purpose is to find facts leading to corrective actions rather than to find fault, and that jurisdictional requirements — such as involving worker members of the health and safety committee — must be met. A model can structure and cross-reference what the team gathered. It must never be the thing that determines the cause.
The line, stated once and plainly
AI drafts the document. A competent person checks it against the regulation that applies in your province and signs it.
A model may not select a control measure. CCOHS frames control as its own step following risk analysis and evaluation — see hazard control — and the choice among elimination, engineering controls, administrative controls and PPE is a determination with legal weight.
Nothing produced by software discharges a duty. If a model tells you a control is adequate and it is not, the consequences attach to people.
This is worth being exact about, because "we have a system for that" is not a defence. Ontario’s constructor guideline explains that the intent of the Occupational Health and Safety Act is to have one person with overall authority for health and safety on a project — the constructor — who has the greatest degree of control over health and safety for the whole project, is ultimately responsible for the health and safety of all workers on it, and must ensure that all employers and workers on the project comply with the Act and its regulations.
Beneath that, the Act assigns duties to employers, supervisors, owners, suppliers and workers individually; the province’s guide to Part III of the OHSA sets out those general duties, and its overview notes that employers have the greatest responsibilities for health and safety in the workplace. Ontario also runs a dedicated Construction Health and Safety Program. Personal exposure runs further than most owners assume — see director liability for workplace safety violations.
A mid-sized general contractor with four active sites uses AI on three safety workflows and nowhere else.
First, JSAs. The site supervisor dictates the sequence for a new activity — setting roof trusses on a two-storey build. The draft comes back with steps and candidate hazards. The crew review it at the morning meeting, delete two irrelevant items, add three the draft missed because they concern this site’s access, and the supervisor signs it. Fifteen minutes instead of an afternoon that would probably have ended in reusing last job’s document.
Second, toolbox talks: each Friday, next week’s schedule produces five short drafts tied to the actual activities, reviewed in ten minutes. Third, the quarterly read-across: two quarters of inspection records are summarised into recurring findings, and one result is uncomfortable and useful — housekeeping deficiencies cluster on one site and one crew, which turns into a supervision conversation rather than another notice.
Note what is absent. No model decides whether fall protection is required. Nobody prompts for a guardrail specification. The regulation text is read as written, because it is written in numbers.
It can help you find the provision, and it should be treated as unreliable on the content. Requirements are provincial and prescriptive — Ontario’s fall protection thresholds are set out in the province’s own guidance, and British Columbia runs a different regulation entirely, with WorkSafeBC noting that amendments on washroom facilities at construction sites took effect on 1 October 2024 for sites with 25 or more workers. Read the regulator’s page, not the summary.
Different question, different risks, and a privacy regime that catches people out. It is covered in AI construction site safety monitoring cameras.
A 30-minute call is enough to tell you whether AI pays for itself here.
