FDA’s AI Warning Is a Governance Wake-Up Call for Pharmaceutical Manufacturing

DeepHow’s PharmaCloud® brings AI into a GMP-validated environment built around approved knowledge, controlled workflows, and human accountability.

At the September PDA/FDA Joint Regulatory Conference, an FDA official described what she called the agency’s first documented warning letter citing inappropriate AI use in pharmaceutical manufacturing. BioPharm International

The underlying letter, issued to Purolea Cosmetics Lab on April 2, states that the firm used AI agents to create drug product specifications, procedures, and master production or control records without adequately reviewing those documents for accuracy and CGMP compliance. Purolea Cosmetics Lab also told investigators it did not know process validation was required because its AI agent had never told it.

FDA’s response was straightforward: AI-generated outputs and recommendations used for CGMP activities must be reviewed and cleared by an authorized human representative of the quality unit.

For manufacturers evaluating AI, you can use AI to help, but you can’t rely on it alone. Someone still has to review its work and take responsibility for the decisions made.

AI assistance needs a governed path to use

AI can help turn complex operational knowledge into information that people can understand and apply. It can accelerate the organization of expert knowledge, support instruction development, and help teams find relevant information.

But an AI-generated instruction begins as a proposed representation of a process. Before operators can use it, qualified people must confirm it accurately reflects the approved procedure, preserves critical requirements, and fits its intended use.

A manufacturer needs to be able to explain where an instruction came from, who approved it, which version an operator used, and how later revisions were managed.

PharmaCloud®: a GMP-validated foundation for operational knowledge

PharmaCloud® is DeepHow’s GMP-validated solution for bringing AI-powered operational knowledge into regulated manufacturing. It delivers the platform in a dedicated cloud environment with governance, traceability, and audit readiness at the center of how knowledge is managed and used.

Its value extends beyond creating content faster. It connects expert knowledge with approved instructions and operator training, helping manufacturers make controlled knowledge usable at the point of work.

An approved SOP defines the process with clear visuals and step-by-step guidance to help operators understand how to carry it out. Version management and approval workflows help maintain the relationship between that guidance and the procedure it represents.

Governance follows the knowledge into the operation.

PharmaCloud’s content approval workflows provide structured review and handoffs before changes go live. Version management maintains traceability across revisions, while version-specific training records help teams understand which instruction version an operator has been trained on.

These controls address a practical manufacturing challenge: an updated SOP only changes behavior when the corresponding instructions and training stay aligned.

Consider a revised equipment setup procedure. Operational questions include whether training materials reflect the revision, whether affected operators have completed the required training, and whether the guidance available during setup matches the approved method.

AI can help organize and compare content, including surfacing differences or potential gaps for review. Authorized people determine whether those differences matter and what should change. The governed workflow connects that decision to the content operators actually use.

In this system, AI helps people manage knowledge while preserving the authority of approved procedures and accountable reviewers.

The next opportunity: evidence of how work is performed

Training and accessible instructions provide an essential foundation. On critical procedures, manufacturers also need to understand whether work is performed as intended. Often that is performed by people watching other people’s work.

AI verification offers a way to examine observable actions against a defined procedure. It can help identify a missed step, an incorrect sequence, or a departure from the expected method, bringing a potential issue to an operator’s or reviewer’s attention.

For a suitable task, that could mean identifying that a required setup action was skipped before the next stage begins. It could also help a reviewer locate relevant portions of an execution recording during an investigation.

These examples illustrate the capability’s potential. They do not establish that every action is observable, that a recording proves product quality, or that an AI assessment can support a regulated decision without appropriate validation.

Validation should follow the decision and its risk.

According to BioPharm International’s conference coverage, FDA emphasized validation rigor proportionate to the decision a tool supports. Tools supporting final quality decisions require full CGMP validation; investigative aids need a defensible fit-for-use justification.

Manufacturers should therefore define the AI’s role before selecting the controls. Helping an investigator find a relevant recording differs materially from determining whether a critical manufacturing requirement was met.

The same discipline applies to changes. A model update, revised instruction, or changed operating environment may affect performance. Quality teams need a documented assessment of that impact and a controlled approval route, including revalidation where appropriate.

The warning letter makes the stakes concrete. Pharmaceutical manufacturers remain responsible for the procedures they approve and the products they manufacture.

PharmaCloud® brings AI-powered knowledge into that framework today. As verification capabilities develop, the opportunity is to extend governed support from what operators are taught to evidence of how work is performed—with validation and human accountability guiding each step.

‍