Every AI answer is an uncontrolled copy

Look at the footer of almost any controlled procedure in a regulated company. You will find some version of the same sentence: uncontrolled when printed. It is a small piece of honesty. The document was governed while it lived in the system. The moment it left, nobody is vouching for it.

Now an AI assistant answers a question about that procedure. It reads the approved current version. It hands back a paragraph in its own words. Nobody printed anything. What came out has exactly the status of a printout.

What document control actually controls

Document control is often called a filing discipline. That undersells it. The requirements in ISO 9001 and ISO 13485 are about circulation. Approve a document before it is issued. Review and re-approve it when it changes. Keep the revision status visible. Put the current version where the work happens. Prevent the unintended use of anything obsolete. EudraLex Volume 4 Chapter 4 sets out the same expectations for drug manufacturers in more detail.

Read that list again and notice what it is about. Not storage. Distribution. A document control system earns its name because it knows where the copies went and can get them back.

The controlled copy and the uncontrolled one

Quality teams have a precise vocabulary for this. A controlled copy sits on a distribution list. When the parent document is revised, that copy is replaced or withdrawn, and somebody signs to say it happened. An uncontrolled copy is a snapshot. It was accurate the day it was made, it carries no promise beyond that day, and no process will ever come looking for it.

The distinction is decades old and very well understood by the people who maintain it. It has simply never had to cope with a system that produces uncontrolled copies on demand, thousands of times a week, in fluent prose, with no watermark on them.

Where retrieval breaks the chain

Here is the sequence in an ordinary week. Someone asks an assistant whether a given control is still in force. The assistant retrieves the approved procedure, which is genuinely the current one, and summarizes the relevant clause. The answer is correct. It then travels into an email to a client, a slide in a board pack, a reply in a support ticket, and a paragraph in a submission.

Four copies now exist outside the document control system. None of them carries a revision number. None is on a distribution list. None can be withdrawn, because nothing in the system knows they exist.

Three things the copy loses

Revision status goes first. The answer says what the procedure requires. It does not say which version it requires it from.

Approval goes second. A named person approved the procedure. Nobody approved the sentence the assistant wrote about the procedure, and the sentence is what travelled.

The recall path goes last, and it is the expensive one. When the procedure is revised, document control withdraws the obsolete version from circulation. It cannot withdraw those four answers, because as far as it is concerned they were never in circulation.

But the assistant cited its source

Citation is worth having, and it is not the same thing as control. A citation links an answer to a document. It does not link the answer to a revision, an approver, or a date, and it does not tell you whether the document still says what it said when the answer was produced.

Grounded is not governed

Citation is not a guarantee of accuracy either. A preregistered study of retrieval based legal research tools found that products from LexisNexis and Thomson Reuters hallucinated between 17 and 33 percent of the time, despite vendor claims of eliminating hallucinations. Retrieval lowers the error rate. It does not remove it, and it does nothing at all about what happens to the output afterwards.

A worked example that landed this year

ISO 9001:2015 has been withdrawn. It was replaced by ISO 9001:2026, and ISO now points organizations to the new edition. Consider every answer an assistant produced over the past few years quoting a clause of the 2015 text. Each one was faithful when it was written. Each one is now a copy of a superseded document, sitting in inboxes and decks that nobody is going to audit.

The source moved. The copies did not. No amount of prompt engineering closes that gap, because the problem is not what the assistant said. It is that nothing above the assistant kept a register of what was said and what it was said from.

Regulators are already asking a version of this

Nobody has written a rule called control of AI answers. Several existing rules already imply one. The MHRA guidance on GxP data integrity is built around a record staying attributable and readable across its whole lifecycle, and it formalizes the true copy so that a reproduction carries the same weight as the original. PIC/S PI 041-1 gives inspectorates the matching expectations.

One question in several phrasings

The EU AI Act carries the same instinct into AI itself. Article 12 requires high risk systems to record events automatically over their lifetime, so that what the system did can be reconstructed later. Different vocabularies, one question underneath all of them: can you show what the record said at the moment somebody relied on it?

What a governed answer has to carry

This is the gap the Cognatum Knowledge Loop is built around. Knowledge stays where your departments already keep it, in the repositories they already use, and Cognatum works across those systems rather than asking anyone to migrate into a new one. What sits above them is the governance those repositories never had.

An approved entry carries a named approver, a date, and the source it came from. It also carries point in time provenance, which records where the knowledge stood when a particular answer was given. That is the part a citation cannot supply, and the part an uncontrolled copy has never had.

Change is a notification, not a silent rewrite

When a source changes, Cognatum tells you, and it tells you what downstream of it is now in question. It does not quietly rewrite the answer behind your back, and it is not magically current. A person is told and a person decides. AI runs the finding, gathering and surfacing across the eight steps of the Loop. Approve is the one step it does not run.

Where two sources contradict each other, both are surfaced along with the disagreement, rather than one being picked as the winner. Deciding what is true is a human job, and the Loop keeps it that way.

A short test you can run this week

Take an answer your assistant gave a colleague last quarter, and ask four questions about it.

Which revision of which document was it based on? Who approved that revision? Has the document changed since? And if it has, who has been told that the answer is now suspect?

In most organizations the first question is answerable with effort, the second with more effort, and the last two are not answerable at all. That is not a failure of the assistant. It is a document control boundary that was drawn before anything could read a procedure and then speak about it.

Where the boundary needs to move

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

The full Loop is at cognatum.ai/the-loop. If you would rather see it against your own knowledge stack, a twenty five minute session is at cognatum.ai/schedule.