Cognatum

Standards

Work instructions: nobody is trained on an answer

Work instructions are the most tightly controlled documents in a quality system: numbered, revised, signed, and tied to a training record. AI answers do the same job with none of that.

Cognatum Team · Oct 28, 2026 · 6 min read

Your quality system has one document that tells a person exactly how to do a job. That document is the work instruction. It sits at the bottom of the document stack, and it is the most tightly controlled thing in it. It carries a number. It carries a revision. It has an owner and a signature. The person who follows it has a training record naming the version they were trained on.

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

Now someone asks the assistant instead

They type how do I do this, and an answer arrives in four seconds. The answer has no number. It has no revision, no owner and no signature. Nobody was trained on it. It is doing the job of a work instruction with none of the control that makes a work instruction trustworthy.

This is a gap in the document hierarchy, and the hierarchy is where quality systems do their work.

The document that answers how

The distinction is well settled. A process says what is done, by whom, where and when. A procedure says how to perform the work and why. A work instruction says how to accomplish one specific task, in full detail, for one role. One published comparison calls the procedure a mandatory method and the work instruction mandatory guidance.

A work instruction vendor puts it more plainly. Teams know what needs to be done, who will do it and when it is due. What they are explicitly missing is how the job needs to be performed.

So the work instruction layer is the how layer. That matters, because how is almost always what someone asks an assistant.

The template opens with a number and closes with a signature

Look at the canonical format. One widely used template runs: work instruction number, purpose, scope, records, associated documents, definitions, responsibility, tools, safety requirements, instructions, appendix, supervisor name, signature.

Why the first and last fields matter

The number makes the instruction findable, citable and supersedable. The signature makes it attributable. Everything between the two is content.

An AI answer is that content with the first and last fields deleted. It is the middle of a controlled document, delivered without the ends that control it.

Training is where the control actually bites

Document control alone is not what makes a work instruction binding. Training is. A regulated plant runs a training matrix by department, initial and job-specific training, and fresh training whenever a procedure is revised.

What an inspector is checking

When an inspector pulls training records, the account of what they are checking is specific. That the employee is authorized to work. That the employee was trained on the proper procedures. That the employee used a certain version of the procedures. That the training was completed. That competency was evaluated.

The same source lists a common deficiency: the wrong version of an SOP recorded in the training records. It lists a matching best practice: ensure there is a connection between training and revisions done on controlled documents.

No revision, no competence

Read that chain backwards and the problem is clear. Competence is evidenced against a version. A version belongs to a numbered document. Take away the number and nothing holds the chain together.

Nobody can be trained on an AI answer, because there is no stable object to be trained to. Nobody can be retrained off one either, because there is no revision to retire. The answer simply changes, and no record says when.

The four fields an answer is missing

Before an answer can enter any document control regime at all, it needs four things a work instruction already has.

Identity, meaning a number that names this answer rather than this conversation. Version, meaning which revision of the source it was built from. Authority, meaning the person who approved that wording. Currency, meaning the date on which they approved it.

Those four fields are not paperwork. They are the difference between guidance and output.

The duty already names competence

Regulators have started to write the human half of this down. Article 14 of the EU AI Act requires high-risk systems to be designed so that natural persons can effectively oversee them.

Three of its provisions are worth reading closely. Paragraph 4(b) requires the overseer to remain aware of automation bias, and it singles out systems used to provide information or recommendations for decisions taken by people. Paragraph 4(d) requires that the overseer be able to disregard, override or reverse the output. Paragraph 5 requires, for certain systems, verification by two natural persons with the necessary competence, training and authority.

Competence for the person, nothing for the content

Notice what is governed. The Act names competence, training and authority for the human reading the output. It names nothing at all for the version of the document the output was assembled from. A plant has solved the second problem for forty years and calls it a work instruction.

Grounded is not qualified

It is tempting to assume retrieval closes the gap. It does not. Researchers ran the first preregistered evaluation of commercial legal research tools and found that systems from LexisNexis and Thomson Reuters hallucinated between seventeen and thirty three percent of the time, despite vendor claims of eliminating hallucinations.

What the standards say about the knowledge itself

The pharmaceutical frameworks are blunter than people expect. ICH Q10 names knowledge management an enabler of the quality system. EU GMP 1.4(ii) asks that product and process knowledge be managed throughout all lifecycle stages. The ICH Q9(R1) questions and answers then say there is no regulatory requirement for a formal knowledge management system, but that knowledge from different processes and systems is expected to be appropriately utilised.

Appropriately utilised is the operative phrase. A work instruction nobody can find is not utilised. An answer nobody can trace is not utilised either, however fluent it reads.

A live example of why versions matter

ISO 9001:2015 has been withdrawn and replaced by ISO 9001:2026, and the standard's page now carries a question on what that means for organizations certified to the 2015 edition.

An assistant answering today from 2015 clause numbers will sound exactly as confident as one answering from the current edition. Nothing in the answer tells the reader which one it used.

What a governed answer looks like

The Cognatum Knowledge Loop runs eight steps across four phases, and seven of them are run by AI. Approve is the one that is not. A named person signs off, and the AI never decides what is true.

What the AI does is route. It finds, gathers and surfaces, then hands the item to a human who can rework the wording before approving it. Where two documents contradict each other, it presents both and shows where they disagree.

What that buys a quality team

Every approved entry carries the approver, the date and the source. Cognatum also records where the knowledge stood when an answer was given, which is point in time provenance rather than a current snapshot. When a source changes, it notifies the people who depend on it. It does not quietly rewrite anything.

None of this asks you to move your documents. Knowledge stays in the repositories each department already uses, with a bidirectional flow to the places people work. Your company's knowledge isn't missing. It's unusable. Cognatum changes that.

One test to run this week

Pick a work instruction that was revised this year. Ask your assistant the question that instruction answers. Then ask it for the entry, the revision, the approver and the date.

If it can produce all four, your knowledge is governed. If it produces a good answer and nothing else, you have a work instruction with no number, no signature and nobody trained on it.

Common questions

Questions this raises.

What is the difference between a work instruction and an SOP?

An SOP describes the sequence of a procedure, including what is done, by whom and why. A work instruction describes how one person performs one task within that procedure, in full detail. Work instructions are the most specific layer of the document stack, which is why they are also the layer a person is most likely to ask an AI assistant about.

Should an AI answer be treated as a controlled document?

An answer that tells someone how to perform a task is doing the work of a controlled document, so it needs the same four attributes: an identity, the version of the source it came from, the person who approved that wording, and the date of approval. Without those, the answer cannot be superseded, audited or trained against.

How do training records relate to AI answers?

Training records bind a named person to a named version of a procedure, which is how an inspector confirms someone was authorized to do the work. An AI answer carries no version, so there is nothing for a training record to point at and nothing to retrain against when the underlying document changes.

Does Cognatum make us compliant with ISO 9001 or the EU AI Act?

No. No software confers compliance with any framework. What governed knowledge does is supply the evidence a reviewer asks for: who approved an entry, when, from which source, and where the knowledge stood at the moment an answer was given. The compliance judgment stays with your organization and its assessors.

Do we have to move our work instructions into Cognatum?

No. Information stays where it already lives, in the departmental repositories each team uses. Cognatum maintains a bidirectional flow to those systems rather than housing the content itself, so the departmental split stays intact and nothing has to be migrated first.

Knowledge governed. Intelligence everywhere.

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