A few days ago I wrote about what nobody tells you about AI validation in GMP manufacturing (https://instantgmp.com/ai-in-gmp-manufacturing/), and the core argument was simple: AI is expanding into regulated manufacturing faster than most quality systems are prepared to handle it, and the FDA’s existing regulations apply fully to whatever AI produces. I did not have to wait long to see a real-world test case.

On April 2, 2026, the FDA issued a warning letter to Purolea Cosmetics Lab, a pharmaceutical manufacturer in Livonia, Michigan, and it is worth reading the original document on the FDA’s website rather than relying on summaries. This is not just a story about AI. It is a story about what happens when a company mistakes a tool for a system.

AI Is Spreading Faster Than the Internet Did

I have been in this industry long enough to remember when web-based software was considered radical. We were early adopters at InstantGMP, and it took years before our customer base became comfortable with the idea that their manufacturing data lived in a browser instead of a server down the hall. When we made that case to customers in the early days, we encountered resistance at every turn, the kind that comes not from ignorance but from deeply held habits: the system had always worked a certain way, the data had always lived in a certain place, and the people responsible for it were understandably cautious about changing either.

AI is getting adopted faster, much faster, and the rate of acceptance across industry is accelerating at a pace that makes the internet adoption curve look gradual by comparison. That speed has real consequences for regulated manufacturers, because fast adoption means fast mistakes. When companies adopted web-based systems over the course of years, there was time to learn, adjust, and build the right controls around new technology. With AI, some manufacturers are deploying tools in weeks and assuming the tool handles compliance because it sounds confident and produces polished output. That is exactly the kind of assumption that produced the Purolea warning letter.

What AI Actually Does, and What It Does Not Do

Most people who deploy AI in a manufacturing context do not fully understand what they are working with. An AI system draws on what it has learned from the internet and reprocesses that information to address a specific question, which makes it extraordinarily useful for a wide range of tasks. But it may not have correct information to start with, and when it lacks adequate information, it will sometimes invent a plausible-sounding answer to complete its assigned task rather than admit the gap. This behavior, known as hallucination, is well documented in every large language model in use today, and it is a known risk that has to be managed, not ignored.

The second limitation is equally important and gets far less attention. AI only addresses the task it is given. It answers the question you ask, and it does not spontaneously scan your entire operation to flag what you forgot to ask about. An experienced GMP consultant walking into a facility for the first time sees things that were never on the agenda: a component qualification process that relies too heavily on supplier documentation, a microbiological testing gap that nobody has thought to question in years, a batch record review step that has quietly become a formality. AI, handed a narrow task, is blind to everything outside that task unless it has access to your complete documentation and is specifically directed to look for gaps across the whole system. That distinction matters enormously, and the Purolea warning letter illustrates it precisely.

What the Warning Letter Actually Says

The FDA inspected Purolea’s facility in Livonia, Michigan on October 28 through 30, 2025, and the warning letter that followed covers considerably more ground than the AI finding that has received the most attention. Reading the original document is instructive, because the AI issue sits in the middle of a much longer list of failures.

The inspection found insects, filth, leaves, and clutter in several areas of the facility, and the docking bay door, when open, directly exposed the manufacturing area to the outside environment. These are not regulatory abstractions or documentation deficiencies; they are basic sanitation failures of the kind that would concern any quality professional within the first hour of a facility walkthrough.

On the analytical side, the firm released finished homeopathic drug products without microbiological testing, and they qualified incoming components by relying entirely on their suppliers’ certificates of analysis without independently verifying those suppliers’ test results. Under 21 CFR 211.84, that independent verification is not optional. Accepting a COA at face value, without establishing the reliability of the supplier’s testing at appropriate intervals, is a violation, and it is the kind of gap that an experienced GMP consultant would identify on the first day of a review. An AI agent asked only to draft specifications or write procedures would be completely blind to it unless it had access to the entire procurement and testing record and was specifically directed to evaluate component qualification.

The Quality Unit Failure at the Center of It All

The third major finding in the warning letter is the one that connects everything else. Under 21 CFR 211.22, the quality unit carries the ultimate responsibility for ensuring that drug products manufactured are in compliance with cGMP, and at Purolea the quality unit failed on three specific counts: procedures were not established or followed as required by 21 CFR 211.22(d), batch records were not reviewed before drug product release as required by 21 CFR 211.22(a), and adequate production and process controls were not established as required by 21 CFR 211.100(a).

Those are not AI failures. They are management failures, and they suggest that AI may have created a false sense of security at this company, a sense that because documents had been generated and procedures had been written, the quality system was functioning. It was not. AI can produce a document that looks complete and reads as authoritative without any of the underlying substance being in place, and without a functioning quality unit to review what AI produces, there is no check on that gap.

Using AI to take shortcuts rather than using it to reinforce and strengthen your quality system is exactly what appears to have happened at Purolea. InstantGMP’s AI Insights feature could have been used to identify these gaps before any product was released, but only if management had been asking the right questions and acting on what it found.

What Intelligent Use of AI Looks Like in a GMP Environment

The Purolea case is useful precisely because it points toward what AI should be doing in a regulated manufacturing environment, not what it was doing there. AI is genuinely powerful for gap analysis, but only when it has access to your complete quality system and is directed to look across it comprehensively. At InstantGMP, we have built a fully integrated global manufacturing platform that puts compliance data, electronic batch records, document management, inventory, and quality management in one system with built-in workflows that reinforce FDA requirements. Because all of that data lives in one place, the AI Insights feature can review it the way an experienced consultant would, find the weaknesses, and surface them before an FDA inspector does, drawing on actual production records and regulatory requirements rather than general knowledge from the internet.

The critical point is that AI is a tool, not a quality system. It can identify a gap and recommend a corrective action, but it cannot implement the solution, and it cannot replace the judgment of a qualified quality unit. Ultimately it is up to management to ensure that what AI finds actually gets fixed. That accountability does not transfer to the software, and the FDA will not accept the software as the responsible party when something goes wrong. As one of our team members put it plainly: the FDA inspector is not going to cite your AI. They are going to cite you.

The Right Question Before Your Next Inspection

The Purolea situation involved violations that went well beyond the AI issue, from physical contamination in the facility to drug products marketed for serious conditions like shingles and genital herpes without approval. Not every company working through its AI adoption is in that position, and most are not.

But the underlying dynamic is one I am seeing more broadly as AI adoption accelerates. In an environment where the technology moves this fast, it is easy to mistake confidence for competence. AI produces polished, authoritative-sounding output, and it does so quickly, which can create the impression that the work is done when in fact the quality oversight has not even started. The right question before your next inspection is not whether you are using AI, but whether your quality system can govern what AI produces, whether the quality unit is still doing its job, and whether the documents AI helped create have been reviewed by someone qualified to review them.

We are adapting AI as a tool to improve our electronic software solutions and our quality management systems, and we are learning as we go, as every company using this technology seriously should be. The Purolea warning letter is a useful reminder that the FDA is watching that learning process, and that the consequences of getting it wrong are not theoretical.

The full warning letter is available on the FDA’s website. I would encourage any quality professional to read the original rather than relying on a summary, because the details matter, and most of the details have nothing to do with AI.