What AI should, and should not, do in health and safety
A position on where machine judgement helps a safety team and where it must stay out of the way. The short version: watch everything, decide nothing.
14 July 2026 · 6 min read
Every safety product now says it uses AI. Few say what for. Here is what we think the technology is good at in this field, and where the line is.
Where it helps
Reading across everything, every day. A human H&S manager with four sites cannot re-read every asset record, every risk assessment review date and every open incident daily. Software can, and should. The value is not cleverness; it is attention that never lapses.
Ranking. Not every overdue item is equal. A missed RIDDOR report is a criminal offence; a lapsed internal service is a job. A system that knows the difference and orders your morning accordingly is doing something a spreadsheet cannot.
Connecting records. A failed pre-use check on a forklift, its inspection history, the incident on the same truck last month and the risk assessment for the loading bay are one story. Surfacing that story is where machine reading earns its place.
Keeping the reference material current. Templates, thresholds and review cycles drift out of date. Maintaining them against current HSE guidance and regulation is work that should happen centrally, once, for everyone.
Where it should stay out
Deciding. Whether an incident is reportable, whether a residual risk is acceptable, whether a contractor is approved: these are judgements a named person makes and signs. Software can present the facts and the thresholds; it should not sign.
Writing your risk assessment for you. A generated assessment that nobody on the shop floor has thought about is worse than a short one they wrote themselves. Templates help people start; they should not finish.
Nagging. Alerts that fire constantly are alerts that get muted. One digest, ranked, to the person who owns the item, escalating only when something is genuinely late.
Our rule
Watch everything. Decide nothing. Make the person who does decide better informed than they would have been, and leave a trail of what was known when.
That is what the AI watch in Safegen does, and all it does. Everything it raises is a fact from your own records or the current legislation; every decision stays with a named person.
