EXPOSED
Roughly half this job is document work AI already handles well — cross-referencing OSHA/ANSI/NFPA clauses, drafting written safety programs and JHAs, compiling incident statistics, generating training materials, and producing compliance reports. The other half requires walking the plant floor to see the unguarded nip point nobody logged, interviewing workers after an amputation, and personally signing off that a machine guarding design or confined-space plan is adequate — with real liability if it isn't. Judgment under ambiguity (is this risk acceptable, do we shut the line down) is the durable core.
Mixed — a routine tier and a judgment tier. Hazard analyses, OSHA 300 log compilation, exposure-monitoring writeups and PPE hazard assessments are template-driven enough for a model to draft, but the walkthrough that finds an ungrounded panel or a modified guard that was never redlined, and the machine-specific risk assessment negotiated with the maintenance lead, still need a person on site — hence 12 rather than the 6 of a pure reporting role or the 16 of an inspector who is never at a desk.
Some physical or field component. You spend real time on the floor with a sound level meter, doing lockout/tagout verification, fit-testing respirators and climbing to check fall-arrest anchorages, but most weeks the majority of hours go to program documents, design reviews and CAD/spec markups in an office — that split puts it at 11 rather than the 15+ of a field inspector living in refineries and trenches.
Licensed human required and personally liable. Many of these roles run on CSP/ASP certification rather than a state licence, but where you stamp a guarding design, a confined-space entry procedure or a ventilation calculation as a PE, your seal is personally attached and OSHA citations plus tort discovery land on it — 11 reflects that the licensed-and-liable version is common but not universal.
Some relationship component. Line supervisors and workers have to trust you enough to report a near-miss instead of hiding it, and plant managers weigh your credibility when you recommend capital spend, but the deliverable is a compliance program and a signed assessment that any competent successor engineer could take over — the relationship helps the work, it isn't the product.
Exists to be accountable for ambiguous calls. Deciding a residual risk is acceptable, calling a line down mid-shift, ruling a fatality's root cause as procedural versus design, and choosing engineering controls over administrative ones are calls made with incomplete data where being wrong kills someone and shows up in litigation — 14 places it at the entry of the top band because standards like ANSI B11 and NFPA 70E still bound the decision space.
Has AI actually changed your work?