It’s Past Time to Reality Check AI Usage in Manufacturing

On April 2, 2026, the U.S. Food & Drug Administration (FDA) sent a warning letter to Purolea Cosmetics Lab. Purolea Cosmetics Lab operates as a contract manufacturer that specializes in producing oil-based skin and healthcare products, primarily with natural or organic compounds. The FDA wasn’t taking issue with the compounds. It did, however, take issue with the fact that Purolea had been caught using Artificial Intelligence (AI) agents to create drug product specifications, procedures, and master control/production records without any human review that ensured the AI-created content was accurate, let alone Current Good Manufacturing Practice (CGMP) regulations-compliant.

The temptation to use AI to automate is not just present, it has been a focal point of messaging campaigns aimed at manufacturers. Organizations such as Tulip Interfaces, Siemens, and Infor spent Hannover Messe showcasing the vision of a factory floor where an AI model could just plug in operational information and immediately produce positive decisions and results.

The messaging is that AI data can be trusted and AI usage is revolutionary. It urges manufacturers to imagine the possibilities of advanced AI agents operating autonomously across numerous aspects of the production workflow, plugging knowledge gaps and reducing time to value. The idea is beyond tempting and sounds like the realization of years of digital transformation messaging. The automated, super-fast future is here and can be used now.

This, unfortunately, will likely not be the only time that the FDA or other regulatory body reprimands a company over the excessive and ill-advised use of AI. What happened with Purolea Cosmetics Lab is a warning story, but not one against AI usage in general, or even AI usage in manufacturing. It is, however, a reality check that AI by itself is not enough, no matter how advanced the program. Human control and authority cannot and should not be usurped—and if companies try to take shortcuts, there will be consequences.

As AI usage in manufacturing moves from pilots and experimentation to mainstream adoption, the conversation needs to shift away from the potential to the reality, taking into account crucial considerations such as account governance and quality control. Heavily regulated industries such as pharma need to know they can always test the AI solution, always monitor and control exactly where and how it is used. Data must be understood and acceptance criteria need to be set well in advance of integrating any agentic platform into a production workflow.

In the European Union (EU), this is essentially spelled out in Annex 22, a draft guidance document created by the European Medicines Agency (EMA) to regulate AI in Good Manufacturing Practice (GMP) settings. It dictates that organizations describe, in detail, the intended use of the AI application, including requiring such data as inputs, variation parameters, limitations, outlines of potential erroneous or biased inputs, test data, and documented Subject Matter Expert (SME) approval and awareness of the solution’s intended usage.

Regulated environments do not just suggest a need for control, they mandate it. This is crucial to remember because many AI tools are designed with Machine Learning (ML) capabilities. The software will naturally look for an optimum method to accomplish its directed task. Yet, AI is not aware of its environment or its duties, not in the way that people are, so it can create new types of vulnerabilities for manufacturers that do not use it properly.

Successful AI integration in regulated industries such as pharma require close partnerships, not just between the manufacturer and the AI solution provider, but internally. All staff need to be well aware of the process at hand, understand its needed steps, and have full confidence that they are working alongside an AI solution that carries the needed context to execute not just efficiently but correctly, taking into account aspects such as security, quality control, and the appropriate governance measures. As embarrassing as the Purolea Cosmetics Lab incident is, the reality is that it could have been much worse without the FDA’s oversight. Manufacturers need to imagine a world where AI is treated like a tool to help improve job performance, and less like a colleague who says it’s okay to skip process validation.