Pharma AI and ALCOA: Every Decision Needs a Contemporaneous Record

On October 28, 2026, IIoT World published a summary of a panel at the Industrial AI Summit 2026 on the relationship between artificial intelligence and data integrity in pharma. The participants were Adam Procopio (Merck), Kristen Sauter (Adlib Software) and Maya Schushan-Orgad (Teva), moderated by Rick Franzosa of Tech-Clarity. The central message: decisions made by AI agents must satisfy the ALCOA principles with the same rigor as human ones.

A note on the source: the article is sponsored by Adlib Software, although it describes itself as editorially independent, and it states that AI tools helped summarize it. It is also a short summary, with no circulation data and no statements from regulatory authorities. The claims reflect the speakers' opinions, not a formal standards update.

What ALCOA means for an AI agent

ALCOA stands for Attributable, Legible, Contemporaneous, Original, Accurate. The case described is an agent that links a critical process parameter (for example temperature or pH) to a critical quality attribute (for example potency or purity). At that moment, provenance metadata must be captured, including:

• 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. Procopio sums it up:

"If it isn't documented, it's not sourceable."

The contemporaneity test

Procopio poses a design question: are agents architected to "emit evidence, not answers"? Adlib offers a practical test: can you show the record without rerunning the model? If reproducing the explanation means calling the system again, you get a new opinion that happens to agree with the previous one, and that does not pass the "C" in ALCOA.

Also according to Adlib, vendors tend to polish attributability and legibility because they show up in demos, while contemporaneity only surfaces in an inspection, perhaps two years later. The "AI notebook" concept mentioned in the panel points to one possible answer: keeping prompt history, retrieval provenance and model version as a contemporaneous record.

Merck's approach

Merck uses AI as a "consultant, not an oracle": the team still signs every document submitted to a regulatory authority. The company also assesses the risk of each source document the agent uses. A dataset that has gone through regulatory review and formal procedures is low risk; a table that exists only on a PowerPoint slide is high risk. Reviewers can then weigh the confidence of a conclusion.

Merck also draws on Stanford's STORM methodology (2024), which questions an agent through multiple personas and forces it to state confidence levels, gaps and multiple viewpoints.

Why it matters for those designing industrial systems

The topic concerns anyone building agents or recommendation systems for regulated plants: pharma and, by extension, food, medical devices and related sectors. The audit trail is not just a compliance obligation but an engineering requirement for data pipelines and historians.

Practical guidance

The panel's remarks suggest a few design criteria:

• record the model version, inputs, equipment context and operating ranges at the moment of decision
• use immutable, timestamped records
• ensure the record can be retrieved without rerunning the model
• classify the risk of the sources the agent uses and make it visible to reviewers
• keep human sign-off on documents intended for regulators

Since these are professionals' opinions and not an updated standard, it is advisable to check these criteria with your own quality function and the applicable regulatory requirements. The underlying message, though, is clear: evidence must be designed in beforehand, not reconstructed afterward.