When an AI agent recommends a process change on a pharmaceutical production line, the recommendation itself is secondary. The defensible record of how the agent reached that recommendation is the product. At the Industrial AI Summit 2026, Adam Procopio of Merck explained how ALCOA, the data integrity standard governing every pharmaceutical record, applies to AI-generated decisions with the same force it applies to human ones.
What Does ALCOA Require from AI-Generated Decisions?
ALCOA stands for Attributable, Legible, Contemporaneous, Original, and Accurate. Every pharmaceutical manufacturing record must satisfy all five criteria. When an AI agent links a critical process parameter like temperature or pH to a critical quality attribute like potency or purity, the provenance metadata captured at the moment of that decision must include the full feature set: which models were used, what equipment was involved, timestamps, operators or robots involved in the production operation, and the proven acceptable ranges the process was operating within. Reconstructing this information after the fact does not satisfy the standard. “If it isn’t documented, it’s not sourceable. It didn’t exist, it didn’t happen.” The question Adam puts back to his team is whether agents are architected to emit evidence, not answers, because fluency-optimized output and attribution-optimized output serve different purposes.
How Do You Test Whether an AI Record Is Contemporaneous?
Adlib Software described a single test that separates a genuine record from a retroactive explanation: can you show the record without running the model again? If reproducing the explanation requires reinvoking the system, what you have is a new opinion that happens to agree with the old one. That fails the C in ALCOA. Vendors tend to build for attributable and legible because those properties are visible in a product demo. Contemporaneous only surfaces two years later during an inspection, and by then it is the only letter that matters. Adam treats AI agents the same way he treats human employees in this regard: both can produce different outputs when prompted again, but both must leave behind a defensible record at the time of the original decision. The AI notebook concept, which captures prompt history, retrieval provenance, and model version, functions as that contemporaneous record. Adam uses AI as a consultant, not an oracle. His team still signs off on every document submitted to a regulatory body.
How Should AI Agents Protect Against Their Own Errors?
Beyond documenting decisions, Adam categorizes the risk level of every source document an agent references. A dataset that has been through regulatory review and formal procedures qualifies as low-risk. A data table that exists only on a PowerPoint slide qualifies as high-risk. This classification allows a human reviewer to weigh the confidence of any AI-generated conclusion proportionally. Adam also draws on Stanford’s STORM methodology, published in 2024, which installs multiple personas to question an agent from different perspectives before accepting its output. STORM forces agents to provide confidence levels, identify gaps, and surface multiple viewpoints rather than converging on a single answer. An AI system optimized only for correct-sounding output is, in a regulated environment, a finding waiting to happen.
This article is based on a panel discussion at the Industrial AI Summit 2026 featuring Adam Procopio of Merck, Kristen Sauter of Adlib Software, and Maya Schushan-Orgad of Teva Pharmaceuticals, moderated by Rick Franzosa of Tech-Clarity. AI tools were used to help summarize and organize the content. Reviewed and edited by the IIoT World editorial team.
Editorially Independent, Sponsored by Adlib Software
