Pharma AI and ALCOA: Every Decision Needs Contemporaneous Evidence
According to a panel at the Industrial AI Summit 2026, AI agents in pharma must meet ALCOA principles just like any human decision. That requires a record created at the moment of the decision, not reconstructed afterward.
On October 28, 2026, IIoT World published a summary of a panel from the Industrial AI Summit 2026 focused on AI in pharmaceuticals. The central message, voiced by Adam Procopio (Merck), is that AI agents must meet the ALCOA data integrity principles with the same rigor required of human decisions. Kristen Sauter (Adlib Software) and Maya Schushan-Orgad (Teva) also took part; Rick Franzosa (Tech-Clarity) moderated.
A note on transparency: the article is sponsored by Adlib Software, although IIoT World describes it as editorially independent, and it discloses that AI tools were used to help summarize its content. It is best read as a record of positions voiced at a panel, not as guidance from a regulatory body.
What ALCOA means for an AI agent
ALCOA stands for Attributable, Legible, Contemporaneous, Original, Accurate. Applied to AI, the principle also covers decisions generated by an agent. For example, if an agent links a critical process parameter (temperature, pH) to a critical quality attribute (potency, purity), the provenance metadata captured at that moment must include:
- the models used
- the equipment involved
- timestamps
- the operators or robots involved
- the proven acceptable ranges within which the process was operating
According to the panel, reconstructing this information after the fact does not meet the standard. The test Procopio applies to his team is whether agents are designed to "emit evidence, not answers."
The contemporaneity test
Adlib proposes a practical test: can you show the record without re-running the model? Re-running it produces, in the panel's words, "a new opinion that happens to agree with the old one," and that fails the "C" in ALCOA. Also according to Adlib, vendors tend to polish the attributable and legible aspects because they show well in demos, while contemporaneity only surfaces as a problem during an inspection, perhaps two years later.
Agents must emit evidence, not answers.
AI notebooks and error control
The panel cites the concept of an AI notebook: prompt history, provenance of retrieved data and model version serving as a contemporaneous record. Procopio uses AI "as a consultant, not an oracle," and his team still signs every document submitted to a regulatory authority.
To limit errors, Procopio classifies the risk of each source document the agent uses: a dataset that has already passed regulatory review is low risk, while a table found only in a PowerPoint slide is high risk. He also uses Stanford's STORM method (2024), in which multiple personas question the agent, forcing it to state confidence levels, gaps and multiple viewpoints.
Why it matters for automation professionals
A concrete design requirement follows for agents in regulated plants, pharma and, by extension, other validated industries: record the evidence at the moment of the decision, including model version, inputs, equipment and ranges. The same approach, based on immutable, timestamped logs, applies to integration with MES, SCADA and historians, where much of this data already resides.
The limits should be noted, however: the article provides no implementation details (tools, storage) and no statements from regulators beyond the ALCOA principles.
Practical takeaway
When evaluating an AI agent for a validated plant, ask the vendor to show, during testing, the record of a past decision without re-running the model. If that is not possible, the system probably will not withstand the contemporaneity an inspection requires.