AI is coming to the GMP quality system. Understanding has to come first.
FDA has a planned guidance on AI and machine learning in pharmaceutical manufacturing. Having evaluated process analytical technology from the inside, I think the hard part is not the model. It is the signature underneath it.
FDA released its 2026 guidance agenda and one item stopped me: a planned guidance on artificial intelligence and machine learning quality considerations in pharmaceutical manufacturing.
This is specifically about AI inside the manufacturing quality system. Process monitoring. Deviation detection. Batch release decisions. Continued process verification. That is CGMP territory, and the questions it raises are completely different from earlier AI-related guidances.
Honestly, this one is personal for me.
What process analytical technology taught me
Earlier in my career I evaluated process analytical technology at FDA, before the agency moved to commodity-specific investigator assignments. Back then investigators covered multiple product areas and I had to go deep across all of them. For that assignment I took beginner and intermediate chemometrics courses in 2015, because you simply cannot evaluate what you do not understand. Near-infrared spectroscopy. Multivariate calibration. Real-time dissolution while the process is still running.
That experience taught me something I have never forgotten:
A model is only as good as the understanding behind it. You can have beautiful output and still miss everything that matters if you do not understand what the model is actually responding to.
That lesson applies directly here.
Vioela’s note: A calibration model cannot be more accurate than the laboratory method it was built against, and it knows nothing outside the range of samples it was shown. In practice these models fail quietly, when something changes that nobody thought was related: a new excipient supplier, a different particle size, a humid week. The spectrum shifts for a reason that has nothing to do with the thing being measured, and the model keeps reporting numbers with complete confidence. Good systems build in the model’s own alarm, a statistical check that says this sample does not look like anything I was trained on. The first thing to ask about any model in a GMP process is what happens when that alarm fires, and who is allowed to override it.
FDA has already met this problem once
I will say something else that I do not think gets enough attention. FDA has already wrestled with this in the device world.
Their transparency guidance for machine-learning-enabled devices acknowledges that these algorithms can exhibit a degree of opacity, and that the concepts surrounding their development are unfamiliar to users. In plain terms: a physician using an AI-assisted EKG interpretation or a continuous glucose monitor is trusting the clearance and using the output. They are not evaluating the algorithm. That is the reality of clinical practice, and FDA knows it.
Now apply that same reality to pharmaceutical manufacturing quality.
If a quality unit is signing a batch record influenced by a model they do not truly understand, that signature is not the same as owning the decision.
That is the gap this guidance needs to close.
The standard I would hold to
Not AI because it is exciting. AI because you understand it well enough to:
- validate it,
- govern the change,
- document the decision,
- and defend it during inspection.
If FDA writes this well, the interesting question becomes: what is the first thing your quality unit would need answered before allowing a model to influence a GMP decision?
I have my own answer. I suspect yours is worth writing down before somebody asks you for it in a conference room.
Vioela’s note: Three questions cut through most AI proposals that reach a quality unit. Is the model locked, or does it keep learning after validation? A model that changes itself has changed without change control. Can you produce the data it was trained on, the way you would produce raw data behind any other result? And when the model and the person disagree, whose decision goes in the batch record? Europe’s draft Annex 22 takes a hard line on the first question. For critical GMP uses it accepts only models that are fixed and that give the same output for the same input, and it keeps generative AI away from critical decisions altogether. It is still a draft, but it shows where regulators’ instincts are.
Adapted from a post first published on LinkedIn.