Cognatum

Standards

Nonconformance: you can't quarantine an answer

Nonconformance management rests on one unstated assumption: the defective thing is still in your hands. An AI answer is delivered the instant it exists, so there is nothing to hold, count or recall.

Cognatum Team · Oct 21, 2026 · 5 min read

ISO 9001:2015 has been withdrawn. The 2026 edition now stands in its place. Certified firms are working out the switch with their auditors. Meanwhile, a slide deck in your company still cites the 2015 clause numbers. So does a wiki page. An assistant will read one of them this week. It will answer a question from it. And nothing in your quality system will treat that answer as a defect.

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

What the nonconformance process actually does

A nonconformance is any output that fails to meet requirements. That covers products, processes, services and the way people work. The steps for handling one are much the same everywhere. SimplerQMS lists eleven of them. Find it. Hold it apart. Write it up. Weigh the impact. Dig for the cause. Pick a disposition. Raise a CAPA if the pattern calls for one. Check the fix worked, then close it out.

The first move is physical

Greenlight Guru puts control first, ahead of all else. The bad product goes into a quality hold. That is a closed area only the quality team can reach. The step does not wait for a review board. Calling one takes too long. Cognidox quotes ISO 13485 on the reason. Product that fails to conform must be found and held to prevent its unintended use or delivery.

Then somebody picks a disposition

Once the thing is held, a named group decides its fate. Mitti puts segregation and disposition at the centre of the flow, and the choices are short and well drilled:

  • Scrap it.
  • Return it to the supplier.
  • Downgrade to the previous version until a fix exists.
  • Use it as is, with the justification recorded.
  • Rework it and retest.

Every one of those is an action you perform on an object you are still holding. The whole apparatus rests on one assumption nobody states, because in a factory it never needs stating. The defective thing is still in your hands.

An answer is gone before anyone can object

Now run the same workflow on a wrong answer. The assistant reads the superseded deck. It answers a question about audit requirements. The person who asked acts on it, forwards it, pastes it into a reply to a customer.

There is nothing to hold. Generation and delivery are the same event. By the time anyone could raise a report, unintended use has already happened. Step one of the process has no object to act on.

There is no quantity

A quality hold has a count. Twelve units. Three lots. One batch, quarantined on a specific date by a named person. An answer has no count you can produce.

How many people asked that question last quarter? How many got the version drawn from the superseded deck rather than the current one? Nobody recorded it, so nobody knows. Containment needs a boundary. An answer has none.

Recall does not reach it either

Quality systems do have an answer for defects found after delivery. Cognidox describes it as action appropriate to the effects: recall, rework in the field, restrictions on use. The EU AI Act gives providers of high-risk systems a comparable set of instruments. Article 20 requires immediate corrective action to bring the system into conformity, to withdraw it, to disable it, or to recall it.

Read the object of that sentence. It is the system. Not the answers the system has already given, and not the documents it read before answering.

The failure rate is not hypothetical

Researchers ran the first preregistered evaluation of commercial AI legal research tools. Their finding is worth sitting with. Lexis+ AI, Westlaw AI-Assisted Research and Ask Practical Law AI each hallucinate between 17 and 33 percent of the time. All three are retrieval grounded. All three were sold on claims of eliminating or avoiding hallucinations.

Those tools serve a profession that reads citations for a living. Your assistant reads a shared drive. Put a floor like that under a process whose outputs you cannot count, cannot hold and cannot recall, and you have a class of nonconformance you will never raise.

What would have to exist first

The gap is not a missing policy. It is four missing fields. Before an answer can enter any nonconformance process, it needs:

  • An identity. Which entry it came from, at which version.
  • An owner. A named approver and the date they signed.
  • A boundary. A record of where the entry was served, so it can be reached again.
  • A trigger. Notice that the source underneath it has changed.

A retrieved passage carries text and a relevance score. It carries none of the four. You cannot hold a thing that has no identity, let alone disposition it or recall it. You cannot write a report about it either.

Where Cognatum comes in

Your company's knowledge isn't missing. It's unusable. Cognatum changes that.

Cognatum operates off its own framework, the Cognatum Knowledge Loop: eight steps in four phases, with one human gate. AI handles seven of the steps. A person holds Approve, and nothing reaches a consumer, human or machine, without passing it. Every live entry then carries the record that makes the four fields above real: the named approver, the approval date, the source, and the version.

Notify, not rewrite

When a source document changes, every entry that depends on it is flagged and routed for re-check. A revision is drafted and queued for the gate. The last approved version keeps serving, with its approval date visible, until a person signs off on the new one. Cognatum tells you something moved. It does not quietly correct itself, and it does not claim to be always current.

Retirement that reaches every channel

The archive step is the closest thing here to a recall. A retired entry stops being served everywhere at once, while its full approval history is kept. An answer given while that entry was live stays traceable afterwards. Where two approved entries disagree, the conflict is held open and routed to a person. Neither version is presented as approved while it is contested.

One test, five minutes

Pick a procedure your team superseded this year. Ask the assistant the question that procedure answers. Then ask for the entry behind the answer, the approver, and the date.

If those three come back, you have something you could raise a report about. If they do not, the answer is already out in the business, and no process you own has a way to get it back.

Common questions

Questions this raises.

Is a wrong AI answer a nonconformance under ISO 9001?

That is a question for your own quality system, not one a vendor should answer for you. The standards require nonconforming outputs to be identified and controlled so they are not used or delivered unintentionally. Whether answers produced by an assistant sit inside that scope depends on how your organization has defined its processes and outputs. Most companies have not asked the question yet.

Why can a bad AI answer not simply be contained?

Containment assumes a gap between finding the defect and the defect reaching someone. With an answer there is no gap. It is generated and delivered in the same moment. There is also no quantity to quarantine and no record of where it went, so there is no boundary to draw around it.

Does retrieval-augmented generation fix this?

It lowers the error rate. It does not sign anything. The first preregistered evaluation of commercial legal AI tools found that retrieval-grounded products still hallucinated between 17 and 33 percent of the time. Retrieval tells you where text came from. It does not tell you who approved that text, at which version, or whether it has since been superseded.

What record would make an AI answer auditable?

Four things: the specific entry and version the answer drew from, the named person who approved it, the date of that approval, and the source it derives from. Point-in-time provenance matters as much as the fields themselves, because a reviewer asks what the knowledge said on the day the answer was given, not what it says now.

Does Cognatum make us compliant with the EU AI Act or ISO 30401?

No. Software does not confer compliance with any framework. Cognatum produces the evidence you supply toward your own obligations: approval records with approver, timestamp and source, version history, deployment records and retirement records. The obligation stays with your organization.

Knowledge governed. Intelligence everywhere.

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