AI is genuinely useful in EHS today for one thing: carrying volume. Reading, sorting, drafting, cross-referencing, and surfacing patterns across documents and incident data at a speed no human can match. It is not useful for judgment — deciding what applies to your operation, what a finding means, and who is accountable for the exposure. We run AI across our own delivery and our own firm, and the governing rule hasn’t changed: AI carries the volume, a CSP or CHMM owns the judgment and the liability.
What can AI actually do in EHS right now?
Here is what we run in production, not what we could theoretically build.
Gap assessment at speed. We ingest a client’s written programs, OSHA 300 and 300A logs, incident and near-miss reports, training records, the SDS inventory, and an operational profile. The system maps NAICS code and process inventory to applicable regulatory requirements, identifies missing or outdated programs, reads the injury history for patterns, and scores gaps by severity, likelihood, and citation exposure.
What used to take three weeks of a senior person’s time takes days. That’s the honest claim. Not “AI does the assessment” — AI does the reading, the cross-referencing, and the first-pass scoring, and a certified professional reviews every applicability determination before it reaches a client. The regulatory mapping is where a language model is most confident and least trustworthy, which is exactly why a human owns that step.
Grounded document assistants. An assistant that answers questions against a client’s own program documents — what our lockout procedure says about group lockout, where the hot work permit lives, what training this job classification requires. It doesn’t consult general knowledge about safety. It reads your documents. That constraint is what makes it usable.
Leading-indicator surfacing. Incident and near-miss narratives are the richest unstructured data most companies own and the least used. Free text describing a near-miss on second shift is worth more than the category someone selected from a dropdown. AI reads narratives at volume and surfaces recurring conditions the categorization scheme was never designed to capture.
Document and training library generation. Fast first drafts, always reviewed, never shipped raw.
What does AI do for the business, not just the safety function?
This matters to an operations reader for a specific reason: the question isn’t whether a vendor can demo something. It’s whether the people advising you actually run this way themselves.
We do. A twice-daily triage agent reads email and Slack, creates tasks, drafts replies, books calendar events, and escalates delegated work that’s gone quiet. A 6 AM weekday brief reads project management, calendar, email, and the accounting system and produces a one-screen picture of the day. The evening before an external meeting, a one-pager is built from calendar, prior email history, and public sources.
We audited 24 internal workflows and sorted them into three tiers: assisted work with no code, no-code platform automation, and coded agents. The most useful output of that audit was negative — most things people call “agents” aren’t agents. Deterministic, rule-based work belongs in native automation, where it is cheaper, faster, and doesn’t hallucinate. Reaching for a model to do work a trigger can do is how companies build fragile, expensive systems that impress in a demo and fail in production.
The other firm rule: agents research, enrich, score, draft, and triage. A human sends every message. No client, broker, or candidate ever receives a communication that a person didn’t read first.
Why do most EHS technology investments fail?
Not because the technology doesn’t work. Because the preconditions weren’t there.
We use a four-part EHS AI Readiness Model. All four have to be present. Missing any one turns the spend into waste.
Data quality. If your incident reports are three-word entries chosen from a dropdown, there’s nothing to read. If your training records live in four systems and a spreadsheet, no model reconciles them for you. AI amplifies the quality of your data. If the data is thin, you get confident, well-formatted thin.
Process standardization. Nine sites with nine different incident taxonomies produce nine incompatible datasets. A model can’t compare Site 3 to Site 7 if they don’t mean the same thing by “near miss.” Standard work has to precede the technology. This is the step everyone skips because it’s the boring one.
Leadership literacy. Someone in the executive room has to understand well enough to ask hard questions of a vendor. Where does the model get regulatory content? What happens when it’s wrong? Who reviews output before it becomes a client-facing or auditor-facing document? Buyers who can’t ask these get sold the demo.
Governance. Who reviews what, on what cadence, and who is accountable when the output is wrong. This is not a policy document. It is a named person and a step in the workflow.
Score yourself one to five on each. Below a three on any dimension, fix that before buying anything.
Where does AI break in EHS?
Specifically, from use.
Regulatory applicability. A model will produce a citation-formatted requirement that reads perfectly and is wrong — wrong subpart, superseded revision, a general industry standard applied to a construction scope, or a state-plan jurisdiction it didn’t account for. It will do this with complete confidence and clean formatting. One wrong CFR cite in a client deliverable destroys credibility that took years to build. This is why applicability determinations get human review without exception, and why I flag anything uncertain rather than publishing it.
Anything with a signature or a liability attached. Program approvals, exposure assessments, respiratory protection determinations, PSM element documentation. Certified professionals sign these because someone has to be accountable. Not a workflow problem — an accountability problem, and it doesn’t have a technical solution.
Computer vision without a use decision. Camera systems reliably detect the things they were trained to detect. What breaks is organizational: nobody decided in advance what happens when the system flags a behavior. Absent that decision, you’ve installed surveillance, your frontline knows it, and you’ve spent capital to damage the trust your reporting culture runs on.
Wearables without an adoption plan. The hardware works. Sustained wear rates are the variable nobody underwrites, and a device that’s in a locker generates no data.
Predictive models on thin incident data. A single site with a handful of recordables a year does not have the event volume to train a meaningful predictive model. Vendors will sell it anyway. Ask what data volume the model needs and compare it honestly to what you have.
What should an operations executive do about this?
Assume any vendor claim about AI is describing the demo environment, and ask what happens in yours.
Ask three questions on every demo. What data does this need from us, and do we have it in that form today? Who reviews the output before it’s relied on? What does this cost to run at our volume, not at the pilot site?
Fix the data before buying the tool. A standardized incident taxonomy across your sites is worth more than any platform purchase and costs a fraction of it.
And test whether your advisors use this themselves. At FractionalEHS the answer is our own back office, our own gap assessments, and our own go-to-market. Not because it’s impressive — because a firm advising you on technology adoption while running on email and spreadsheets is telling you something.
