AI Needs Human Oversight To Better GMP Documentation
By Katie Anderson, Chief Editor, Pharmaceutical Online
As pharmaceutical manufacturers continue moving from paper-based records to digital documentation systems, artificial intelligence is increasingly entering the conversation. AI has clear potential to summarize logs, identify recurring deviations, support faster audit response, and help quality teams make sense of large volumes of GMP data. But the central message from Stephanie Wimberly (APTTMHY) and Piyush Modi (Amneal) in a recent Pharmaceutical Online Live discussion was that AI should support quality decision-making, not replace the people accountable for it.
“Digital documentation is a good step, but you can’t forget about the human oversight,” said Modi. Even as organizations modernize their documentation environments, he said, human reviewers must remain responsible for evaluating records, identifying issues, and catching problems that an automated system may miss.
AI As A Decision Support Tool
When asked where AI fits in GMP documentation, Wimberly described it as a tool for validation and verification support, not a substitute for quality review. “It should only add or enhance the speed of the process in which you review, but you do need that human component to be able to verify and double check that the actual AI model is performing accurately and correctly.”
Industry Comfort With AI Is Still Limited
Audience polling during the session suggested that many companies are still early in their AI journey. Only 16% of respondents said their organization is using AI or automation in GMP documentation with formal governance and human review. Another 28% said they are exploring use cases but have not yet implemented them, while 28% are piloting limited, low-risk applications. Nearly a quarter said they are not currently considering AI in GMP documentation.
Both panelists said they were surprised by the relatively low adoption rate, particularly given the time savings AI can offer. Modi noted that AI can help teams move from manually reviewing hundreds or thousands of pages to identifying trends in minutes or hours. Wimberly suggested that budget constraints and reluctance to change may be slowing adoption, but warned that financial hesitation should not undermine product quality or an organization’s quality culture.
Guardrails Before Implementation
The panelists stressed that companies should conduct due diligence before selecting or deploying AI tools. That means understanding how the platform will operate within the organization’s systems, ensuring the tool is properly validated, and confirming that it aligns with applicable regulatory expectations.
Validation was the first guardrail Modi identified. Like any computerized system introduced into a GMP environment, an AI tool must be assessed, documented, and controlled. Users also need training so they understand how the tool supports routine work and where its limitations are. Governance matters as well, according to Modi. “Control should be limited, and it should be with the supervisor and the key people who are regularly dealing with the AI system.”
Wimberly added that organizations should research which tools best fit their quality objectives, validate them appropriately, and make sure they are compliant with the regulatory framework that applies to their operations.
Human Review Must Be Governed, Too
Human oversight is not simply a phrase to place in a policy. It must be operationalized through SOPs, training, role clarity, and documented review. Modi said reviewers should understand the AI system, follow standard operating procedures, and be able to show during inspection that the tool was properly validated, implemented, and used by trained personnel.
This expectation is consistent with current regulatory signals. FDA’s draft guidance on AI used to support regulatory decision-making emphasizes establishing model credibility for a specific context of use through a risk-based framework. More recently, industry coverage of an FDA warning letter involving AI-generated GMP documentation underscored that AI outputs used in quality systems must be reviewed and approved by appropriate human quality personnel before use.
Leadership Cannot Step Away From Quality Risk
An audience question raised a concern many organizations will need to confront: can heavy dependence on AI-based compliance monitoring create a false sense of control, causing leaders to become less engaged in understanding actual quality risks?
The panelists’ answer was that leadership engagement remains essential. Modi said AI can support quality risk management, risk controls, CAPA systems, and trending, but leaders must understand how the system is used and where human intervention remains part of the process. Wimberly added that leaders should be trained on explainability, traceability, audit trails, and the expectations for human review so they can help ensure the system is performing correctly.
Accountability Remains With People
The discussion ended with a practical accountability question, “If AI overlooks a critical data integrity issue during a regulatory inspection, who is responsible?” The answer from both panelists was that people have to be held accountable, not AI.
For pharma companies, the opportunity is significant. AI can help uncover trends faster, reduce the burden of manual review, and support more efficient inspection responses, but the compliance challenge is just as significant. Companies must define intended use, validate tools, train users, control access, document review, and ensure the quality unit retains authority over GMP decisions.
The takeaway is not that AI has no place in GMP documentation. Rather, it is that AI’s value depends on governance. Used appropriately, AI may help quality teams work faster and see patterns sooner. Used without oversight, it can create new inspection risks. In GMP environments, the human reviewer is not optional — the human reviewer is the control.