An AI assistant for field service will be asked, sooner or later, how to bypass an interlock. Not maliciously: a customer's line is down, the interlock is "obviously" the problem, and the technician wants to keep production running until the part arrives. The assistant's answer in that moment is the most important thing about it.
Two failure modes
The first is a helpful assistant that explains how. The second, subtler, is an assistant that summarises the safety sequence into "ensure the machine is isolated" and moves on to the interesting part — and a technician who, under time pressure, takes the summary as permission to skip.
Three design decisions
Verbatim and first. Isolation, lockout/tagout, pressure release and PPE steps are identified during index review and locked. In every checklist they appear first, in the manual's words and order, including bulletin additions. The technician cannot collapse them. The agent cannot paraphrase them.
Refusal is not configurable. A request to defeat an interlock, guard, limit or safety function gets the manual's position cited and the escalation path. There is no admin setting that changes this. We decided that a feature flag for safety is itself a safety defect.
No improvisation. When a symptom is not covered by the manual, the agent escalates with the evidence package rather than reasoning its way to a plausible procedure. A plausible procedure on a machine with stored energy is how people get hurt.
The cost
Sometimes the right answer really is "there is a documented override for this interlock, for this situation, with these conditions" — and it is in the manual, and the agent will return it with the page because it is a documented procedure, not a bypass. The line is: if the manual says it, it is a step; if the manual does not, it is a refusal. That costs a little flexibility. It buys a product that an OEM's safety officer can approve.