“A frontier model with no context is a brilliant new hire on day one. Smart, and useless, until you onboard them.”
Shreyas Becker is Co-Founder and COO of Katalyze AI, which builds AI for process development and manufacturing. Before Katalyze, he worked at Sanofi, where he saw how hard it is to build AI that understands manufacturing science, even at a company ahead of the curve.
In April 2026, FDA issued what is widely reported as its first cGMP warning letter citing overreliance on AI. The firm had used AI agents to write specifications, procedures, and master production records without adequate human review. In Europe, draft GMP Annex 22 on AI remains a consultation draft, yet it is already shaping how manufacturers plan. For anyone bringing AI into regulated work, the question is no longer whether to use it. It is how to make its output defensible. Shreyas argues most teams are working on the wrong layer of that problem.
The Problem is Missing Context
Most AI pilots in manufacturing start with the data. Shreyas says that is the wrong starting point, because the meaning of the data doesn’t live in the data.
“A temperature reading of 37.2 tells you nothing on its own,” he says. “You need the batch, the process step, the equipment, the setpoint, the material lots in play, and the SOP or development protocol version before that number means anything.”
That context is scattered across MES, LIMS, historians, ERP, ELNs, and development reports. A lot of it still sits in PDFs, paper records, and CoAs. “Each team stitched together just enough context to answer today’s question, then threw it away,” he says.
The data shape works against AI too. In development, a team might have a few dozen runs, each with thousands of parameters. Shreyas calls this “small-n, wide-p”: the opposite of what modern AI was built for. Add inconsistent equipment naming and units across sites, plus reasoning buried in free-text deviation reports, and a generic model has little to work with.
He saw this even at Sanofi, which he describes as one of the leaders in AI adoption. The gap wasn’t appetite or capability. The context a model needs “exists as scattered fragments across systems and people.”
Why the Model Matters Less Than What Sits Under It
Teams often spend months choosing a model. Shreyas thinks that choice is converging, because frontier labs are shipping roughly interchangeable reasoning at the top end.
What separates expert-grade output from confident nonsense is context. “Same model, wrong context, you get a fluent wrong answer,” he says. Put the same model on top of the right run, equipment and material history and “the twelve relevant prior deviations,” and “you get something a senior scientist would sign off on.”
He calls the discipline context engineering: building process, product, equipment, material genealogy, regulatory and deviation history into a durable layer a model can reason over. The operative word is durable. It is not prompt wording, and not a one-off retrieval thrown away after the query.
“The binding constraint for enterprise AI in manufacturing is context, not data volume and not model capability,” he says.
What Makes an AI Recommendation Defensible Under GxP
The April warning letter showed what happens when AI output enters GMP records without human scrutiny. Shreyas distinguishes between consumer AI and GxP AI. “A consumer recommendation optimizes for being right often enough. A GxP recommendation has to be defensible every single time.”
He names three requirements:
- Provenance. The recommendation shows which data, prior runs, SOPs, and precedent deviations it used. “No black box gets to influence a batch disposition.”
- A human decision-maker. The system proposes, a qualified person decides, and attribution stays with that person.
- Conservatism about uncertainty. In GxP, he says, “‘I don’t have enough to conclude this’ is the correct and expected answer.”
Underneath, the system has to be validated, reproducible, and change-controlled, because the recommendation itself becomes a record. That emphasis on human oversight and traceability matches the direction of draft Annex 22, which assigns key decisions to process subject matter experts rather than data science teams.
Inside a Deviation, and The Raw Material Blind Spot
Shreyas uses deviations as the clearest test. When a parameter drifts, the system’s first step is not diagnosis. “It contextualizes,” he says, pulling the batch record, equipment history, material lots, SOP version, and environmental data for that window.
Next it retrieves precedent. If a run two years ago had the same profile and traced back to a media lot, that case surfaces with its evidence attached. The system then ranks hypotheses tied to specific data and drafts the investigation’s starting point: timeline, affected batches, impact scope, and confirming tests. An investigator reviews, edits, accepts, or rejects. When the case closes, the resolution feeds back into the context layer.
That feedback loop matters most for raw materials. Shreyas declines to give a percentage, because he doesn’t believe an honest industry-wide figure exists. His view, which he flags as a belief, is that raw material variability is the most under-attributed root cause in fermentation.
Complex media such as soy peptones, yeast extracts, and trace element mixes vary lot to lot. “The CoA tells you the material is within spec. It does not tell you how that lot will actually perform in your process,” he says. Investigations check equipment, operators, and process parameters first, because those are easier to see, and the lot gets a pass. Catching the pattern requires linking performance to full material genealogy across many runs. Where that link exists, raw materials show up as a driver “more often than the records say.”
What To Do Next
- Map the context before you pick a model. For one decision type, such as deviation root cause, list where every piece of needed context lives: which systems, which PDFs, which people.
- Ask any AI vendor to show provenance. Pick one recommendation and ask which runs, SOP versions, and prior deviations it drew on. If the vendor can’t trace it, the output won’t survive an inspection.
- Test how the tool handles thin evidence. Give it a case with insufficient data. A GxP-ready system should say it can’t conclude, not produce a confident guess.
- Check where closed investigations go. If resolutions end up in inboxes and PDFs rather than a shared, queryable record, the next investigation starts from zero.
- Audit raw material attribution. Review recent fermentation deviations and ask whether anyone linked process performance to material genealogy across runs, or stopped at the CoA.
