The FDA Didn't Ban AI. It Reminded Us Who Is Accountable.

AI is moving quickly into regulated knowledge work. That is not surprising. The work is complex, document-heavy, repetitive in places, and often painful in exactly the ways AI can help with.

But the real question is not whether experts should use AI. They will.

The question is whether the systems around AI preserve accountability, evidence, review, and human judgment.

That is why the recent FDA warning letter to Purolea Cosmetics Lab caught my attention.

I am not a regulatory affairs expert, and I do not pretend to be one.

That matters. The fastest way to get this conversation wrong is to act like knowing AI means you suddenly know the domain.

I am not that person.

What I do understand is what happens when important work depends on data quality, access control, privacy, audit trails, versioning, data residency, and systems that need to be explainable after the fact. I also understand the temptation to use AI to make hard judgment look easy.

That is the part of this warning letter I think matters well beyond one facility or one enforcement action.

It is tempting to read the Purolea letter as the FDA coming for AI. I do not think that is what happened.

This was not a clean operation that got tripped up by AI. The facility had sanitation issues, missing microbiological testing, inadequate process validation, weak Quality Unit oversight, and unapproved products marketed for shingles and genital herpes relief.

The AI issue mattered because it fit the pattern: weak review, weak controls, weak accountability.

Purolea told FDA investigators it had used AI agents to help create drug product specifications, procedures, and master production or control records. FDA's response was pretty simple: AI can help create regulated documents, but the Quality Unit still has to review them for accuracy and Current Good Manufacturing Practice compliance.

The point FDA made is the one everyone using AI in serious work should sit with:

That is not anti-AI.

That is accountability.

A tool can help create the work. A human still has to own it.

None of this is new. A calibrated balance still needs review. So does a validated system. So does a polished document. AI does not get a special exemption because the output sounds confident.

The tool changed. The accountability did not.

In serious work, architecture is not just where the data lives. It is who had access, what version they saw, what changed, what evidence was used, who approved it, and whether the whole thing can be reconstructed later.

"The AI said so" is not an audit trail.

That is true in drug manufacturing. It is true in regulatory submissions. It is true in finance, legal, healthcare, security, privacy, data engineering, and plenty of other places where getting the answer wrong has real consequences.

In drug submissions, an unflagged inconsistency can become an FDA Information Request, extra review cycles, or avoidable delay. The risk is not that AI exists in the workflow. The risk is that AI creates or changes important work inside a system that cannot explain itself.

The mistake is not using AI.

The mistake is using AI in a system that cannot explain itself.

For AI tooling in regulated or high-stakes work, the bar should be practical. The system should know what source content it used. It should preserve the prompt, rule, or model that produced the output. It should distinguish deterministic checks from model-based reasoning. It should show the evidence. It should let the human expert accept, reject, or override, and record the rationale.

The record of what happened should be append-only.

Editable history is not history.

And the system should never silently modify controlled work while pretending that is productivity.

The human still owns the decision.

That is not a limitation on AI. It is what makes AI usable in serious work. The best use of AI in expert workflows is not to replace expertise. It is to make expert review sharper, faster, and more consistent.

That is the version of AI I trust.

Not "generate me the final answer."

More like: "Show me what I may have missed, explain why it matters, connect it to the source, and make it easy for the right person to decide."

That is the standard serious AI tools in regulated work should meet.

This is the standard we are building toward at Chrona Bio: better systems for expert work, with the human still clearly in control.

The serious future of AI at work is not experts handing accountability to a model. It is experts using better systems. And still being experts.