AI on an industrial site: three use cases that survive a quality review

On the sites we assess, most industrial AI pilots stop in the quality department rather than in the model. In a regulated plant, nobody in quality will ask first whether the model is accurate. They will ask what happens to the batch record when the model is wrong, and who answers for it. A pilot that cannot reply to that does not reach production, however good its metrics looked.

Three that do

Technical copilots come first. An engineer asks a question, gets an answer with the standard and the clause behind it, and makes the call. The AI supplies information; a qualified person takes the decision.

Predictive maintenance on non-GxP equipment is the second: compressors, HVAC, utilities. The value is real, and there is no batch record and no validation cliff to climb.

The third is document preparation. Draft protocols, draft specifications, draft deviation summaries, all reviewed and signed by a person. What keeps it viable is the word draft.

What kills the others

Anything that writes directly into a validated system, adjusts a setpoint that affects product quality, or produces a record nobody can reconstruct. Such tools might well work. The trouble is that you cannot demonstrate it, and on a regulated site what you cannot demonstrate does not exist.

Put AI where a wrong answer costs a discussion rather than a batch, and earn the right to move closer.

That is also how you build the trust inside the plant to go further later, with the evidence in hand.