WHERE TO START AI IN MANUFACTURING

The first serious generative AI use case in a regulated manufacturing site shouldn't sit on critical GMP. But it should sit right next to it.

And make sure you’re using that non-GMP use case as a place to practise the habits that matter later, when AI moves to quality-critical work.

Data access. Ownership. Version control. Model behaviour. Human review. Escalation. Evidence capture. And whether people still use the system once the project lead leaves the room.

Get this right and a non-GMP use case is no experimental sandbox. It can deliver for real work, real users, real data - at lower consequence. And you can still track the trail. What did the system access? What did it infer? What did the human accept, reject or override? What changed downstream? How was performance checked over time?

A non-GMP use case can start answering those questions immediately.

A tacit-knowledge agent. A management-reporting assistant. A finance-planning workflow. A maintenance-prioritisation tool. None of them touch batch release. Yet they can teach the site how to make AI-supported work legible, governed and repeatable.

FDA's draft guidance on AI for regulatory decision-making sets out what evidence will count. Context of use. Risk. Credibility. Performance monitoring. Life-cycle maintenance.

EMA's reflection paper points the same way. But the draft EU GMP Annex 22 draws a harder line around generative AI and large language models in critical GMP applications.

So each site’s first use case has two jobs. It solves a live operational problem. And it teaches colleagues how to build evidence.

So what's today's in-the-end-at-the-end?

The lowest-consequence project on the site may be the most valuable one you run this year. It's where you find out whether your AI habits hold up while the stakes are still low enough to get them wrong.