Pharma AI: Without Document Integrity, There Is No Defensibility

On October 6, 2026, IIoT World published the first of two parts covering audience questions left unanswered during the Industrial AI Summit 2026. The session was titled "You Closed the Sensor-to-Decision Gap. FDA Will Ask You to Explain It." The speakers were Adam Procopio (Scientific Associate VP, Merck) and Kristen Sauter (President and GM Life Sciences, Adlib Software).

A useful caveat: the session was sponsored by Adlib, and Sauter's points also promote her product, a "document accuracy layer." They should therefore be read as opinions from an interested party. The article contains no numerical data, new products or new regulations: it reflects practitioners' views, not a study.

Prevent with CDMOs, don't fix afterwards

Procopio suggests building digital data requirements into the quality and technical agreements with contract manufacturers (CDMOs): prevention costs less than remediation. However, since CDMO digital maturity varies widely, perfectly structured data cannot be expected. You need tools that ingest spreadsheets, scans and PDFs while keeping the link to the original record.

Today, in his view, AI works best at three tasks: extraction, search and linking of evidence. Gaps remain in semantic interpretation, reasoning across multiple documents, and explainable conclusions that can be defended before regulators.

Context is lost at the boundary

For Sauter, traceability added at receipt is always incomplete: when a batch record crosses a boundary, the context of its generation is lost, namely the approved version, the signatory and the source system. Quality agreements should therefore specify file formats, mandatory metadata fields, version identifiers and delivery protocols. On receipt, it is advisable to lock the version immutably and validate its structure, creating a defensible handoff record.

The most common gap is the link between what was received and what the AI actually saw.

Where transformation breaks the data

Sauter points to three typical risks when preparing documents for AI pipelines:

• Chunking: can split tables, separating headers and values into different chunks.
• Summarization: can compress evidentiary details such as specification ranges, lot numbers and reviewer qualifications.
• Embedding: ranks results by similarity, not by the authority of the document.

Hence her definition of "AI-ready," which goes beyond digitization: intact tables, labeled sections, critical fields extracted and validated, and provenance locked before any chunking.

Can you trace from an AI output to a specific chunk, then to a passage in a precise document version, and demonstrate that the passage means what the AI claimed?

If even one link is opaque, Sauter argues, the transformation has broken something.

Why it matters even outside pharma

The same pattern applies to any regulated or audited plant: batch records, MES and historian exports, and AI or LLM pipelines built on these data. Integrity must be designed into the pipeline, with versioned and verifiable transformations; adding it after the fact is not enough.

Practical takeaways

A few concrete checks, drawn from the authors' reasoning:

• define format, metadata, versions and delivery methods in supplier agreements;
• lock and validate every document on receipt, recording version and origin;
• log every transformation (parsing, chunking, summarization) in a versioned manner;
• periodically test the output → chunk → passage → document chain on real cases.

In short, the message is that the defensibility of an AI system depends less on the model and more on the discipline with which data reaches the model. It should be remembered that these are practitioners' opinions, partly tied to a vendor, but the proposed control points are reasonable and applicable to many industrial contexts.