Cognatum

Life sciences

Electronic batch records: answers have no master

An electronic batch record works because an approved master record defines the batch before it runs. Nothing defines an AI answer in advance, so there is no exception for anyone to review.

Cognatum Team · Oct 27, 2026 · 5 min read

An electronic batch record works because something was approved before the batch started. That something is the master record. It fixes the steps, the materials, the limits, the checks and the signatures. The executed record is then measured against it. Nothing plays that role for an answer your AI assistant gives.

Served to one approved entry
AI assistants & agents
Proposal tools
Internal search & chat
Customer portals
Compliance & audit

Cognatum governs the entry

source · version · approver · permissions

This is not a complaint about EBR software. It is a point about what the control rests on. Review by exception needs a defined normal. A batch has one. An answer does not.

Two records, not one

Qualityze draws the line plainly. The master batch record is the template. It holds the formula and the process steps for a product. The executed batch record is the finished record of one run. It carries the real data, and it follows the master. Microsoft's own documentation splits it the same way. The master manufacturing record sets out the methods that might be used. It also sets out how each trait might be measured. The batch production record holds the method actually used. It holds the measured value too.

The master cannot be edited quietly

The controls on that template are strict. In Dynamics 365, turning on electronic batch records blocks edits to approved bills of materials, formulas and routes. Approval must be removed first. After the edit, approval is needed again before the record can be used in production. Now compare a procedure on a shared drive. Anyone with write access can change it at lunchtime. Nothing downstream notices.

Why review by exception works

Siemens describes the mechanism in one sentence. As the system creates the electronic batch record, it enforces the master batch record. Because the master is enforced, a departure is detectable at the moment it happens. Seal lists what then reaches quality: changed data, failures, alarms, overrides, missing evidence, open samples and deviations, brought to QA while the batch is still running. MasterControl makes a similar claim about speed, saying real-time visibility lets a deviation be identified and resolved quickly, which accelerates batch release.

An exception is a comparison

Strip the tooling away and review by exception is one operation. Compare what happened against what was approved to happen. Flag the difference. Everything else passes without a person reading it. The saving comes entirely from the comparison being possible. Take away the approved side and there is no exception. There is only output.

Four things an answer is missing

When your assistant answers a question from your own documents, what it produces is closer to an executed record than to a master. Four facts are absent.

  • An approved template. Nothing states in advance what a correct answer to that question contains.
  • A limit. There is no range, tolerance or specification to test the output against.
  • A unit. A batch has a number, a size and a boundary. An answer has none of the three.
  • A gate. Release and delivery happen in the same instant, with no review step in between.

So nothing can be flagged, because nothing is being compared. Review by exception collapses into review by nobody.

Which version was in force

Seal makes one more point worth borrowing. The latest version is not always the correct one. Its system resolves effectivity for the product, site, market and start date when the batch is created, then records which versions were selected. A batch already running keeps the version it started with. That is point-in-time provenance, written into a manufacturing system. Ask your assistant which version of a procedure it answered from last Tuesday, and see what comes back.

The guidance already asks for the knowledge

ICH Q10 names knowledge management as an enabler of the pharmaceutical quality system, alongside quality risk management. It says product and process knowledge should be managed from development through the commercial life of the product. EU GMP 1.4(ii) asks that the quality system ensure product and process knowledge is managed throughout all lifecycle stages. Both sentences are about knowledge, not about a system that stores it.

No system required, but the knowledge must be used

The ICH Q9(R1) questions and answers are blunter. There is no regulatory requirement for a formal knowledge management system. It is still expected that knowledge from different processes and systems is appropriately utilised. Knowledge quoted out of a superseded file is not being appropriately utilised. It is being used wrongly, at speed, with nobody in a position to say so.

Oversight needs something to compare

EU AI Act Article 14 requires high-risk systems to be overseen by people. Paragraph 4(b) names automation bias directly, and singles out systems used to give information or recommendations for decisions taken by people. Paragraph 4(d) says the overseer must be able to disregard, override or reverse the output. An override needs a basis. Without an approved version and a date, the reviewer is guessing against the machine.

Retrieval was already not enough

Researchers ran the first preregistered evaluation of AI legal research tools. Products from LexisNexis and Thomson Reuters hallucinated between 17 and 33 percent of the time, against vendor claims of eliminating hallucinations. Those tools retrieve from curated case law. A site's shared drive is not curated case law.

What a master record for knowledge looks like

This is the work Cognatum does. Each entry carries the source it came from, its version, the named person who approved it and the date. When a source changes materially, every entry that depends on it is flagged and a revision queues for a human. The last approved version stays live, with its approval date visible, until a person signs off on the new one. Cognatum notifies. It does not quietly rewrite, and it is never simply current.

Where two entries disagree

Two approved entries that contradict each other are held apart and routed to a named person. Neither is averaged into the other. Neither wins on recency. While the flag is open, the question comes back as a conflict rather than a confident answer. AI runs seven of the eight steps in the Cognatum Knowledge Loop. Approve is the one a person holds, and that step is where the master record gets made.

A test you can run this week

Pick a procedure your site superseded this year. Ask your assistant the question that procedure answers. Then ask it for the entry, the version, the approver and the date. A batch out of specification would have been held. This answer went out, and no exception report has it.

Your company's knowledge isn't missing. It's unusable. Cognatum changes that. The eight steps are at cognatum.ai/the-loop.

Common questions

Questions this raises.

What is an electronic batch record?

An electronic batch record is the controlled digital record of how one batch was made. It has two halves. The master record holds the approved formula, the process steps and the measurements that apply to the product. The executed record holds what actually happened: the materials consumed, the values recorded, the equipment used, the departures raised and the signatures given. The second is produced against the first.

What is review by exception?

Review by exception means quality reviews only the parts of a batch record that departed from what was approved, instead of reading every page. Changed data, failed limits, alarms, overrides, missing evidence, open samples and deviations are surfaced, often while the batch is still running. The rest passes without a person reading it, and the full record stays available.

Why can't review by exception be applied to an AI answer?

Because there is nothing to compare the answer with. An exception is the difference between what happened and what was approved to happen. For a batch, the approved side exists in advance as a master record with limits. For an answer there is no approved template, no tolerance, no bounded unit and no review step before release. With no approved side, every answer looks normal.

What should a regulated site require from an assistant that answers from its own documents?

Four fields on every answer: the entry it came from, the version of that entry, the named person who approved that version, and the approval date. Then three counts a reviewer can read about the library itself: how many entries duplicate another, how many pairs of approved entries conflict, and how many questions have no approved answer at all. Those are the facts an inspector asks for, and they can be produced on demand.

Does Cognatum update entries automatically when a source changes?

No, and that is deliberate. When a source changes materially, Cognatum flags every entry that depends on it and routes a revision to the Approve gate. The last approved version stays live, with its approval date visible, until a named person signs off. Cognatum notifies rather than rewrites, so it should never be described as always current. The approval records it keeps are evidence you supply toward frameworks such as the EU AI Act, ISO 30401 and NIST AI RMF. No software confers compliance with any of them.

Knowledge governed. Intelligence everywhere.

See it on your own content, in your own environment.