"It's a workshop activity": Trish Kerin on where AI helps process safety, and where it doesn't

In a Chemical Processing podcast published on 13 October 2026, process-safety specialist Trish Kerin and editor Traci Purdum worked through a question that has been circling the discipline for a while: what is AI actually good for in process safety, and where does it stop being useful. The framing is worth repeating because it is neither boosterish nor dismissive, and because it lands on the risk-assessment workflow directly.

A diligence partner, not the analyst

Kerin’s useful cases sit upstream of and around the study, not inside it. An internal, company-authorised agent can compile the inputs a hazard review depends on — incident history, trend and maintenance records, similar events elsewhere — faster than a person can, provided every output is checked. Agents can also do weak-signal detection: noticing small drifts in process parameters before they approach limits, and flagging when several drift at once. Her caution there is the old one, stated plainly — “correlation is not causation” — so a flagged pattern is the start of an investigation, not a conclusion.

Where she draws the line is the study itself. A risk assessment, she says, is “a workshop activity”: the expert team working through nodes and guide words brings experience and the ability to make connections that an AI does not have. That maps cleanly onto how HAZOP and LOPA actually generate value — the deviations a good team finds are often the ones nobody wrote down beforehand.

The cautions are the practical part

Three of Kerin’s warnings are worth carrying into any pilot. First, hallucinated sources: generative tools can invent references, so each one has to be confirmed to exist and to say what it is claimed to say. Second, codified bad process — “you can’t fix a process problem with an AI agent”; build an agent on top of a flawed workflow and it “will just codify the existing process issues,” so the process gets mapped and tested before it is automated. Third, data leakage: once plant information goes into a public generative tool, control over it can be lost, which is why she favours locked-down internal systems and restricting agent-building to a small team that applies security controls. She also notes the resource cost — AI’s energy and water use, with chemical sites potentially competing with data centres for both.

The through-line is “garbage in, garbage out,” and the discipline it implies: accurate inputs, verified outputs, and a human who stays accountable for the judgement. For teams fielding vendor pitches that promise AI-run PHAs, it is a clear-eyed reference point from a recognised practitioner.