System of record
Your data retention policy stops at the record
Retention schedules sort records into classes, put a clock on each one and say how it is destroyed. An AI answer drawn from those records has no class, no clock and no disposal method. Two failures follow, in opposite directions.
Cognatum Team · Oct 9, 2026 · 6 min read
Every data retention policy does three things. It sorts records into classes. It puts a clock on each class. It says how a record is destroyed when the clock runs out. Most regulated companies run that control well.
Cognatum governs the entry
source · version · approver · permissions
Then somebody asks the assistant a question, and it answers from one of those records. The answer is a new thing in the world. It has no class, no clock and no disposal method. Nobody wrote it into the schedule, because the schedule was written for records.
What the policy actually covers
The guides agree on this. BigID says a policy sets out what data is kept, how long it is kept, where it sits, and when and how it is deleted. Drata lists six parts. Data classes and scope. How long each class is kept. Where it sits and who can reach it. How it is destroyed. What pauses the clock, such as a legal hold. And who owns the policy and reviews it.
Read those six again. Every one of them names the record as the thing being governed.
The tooling has the same shape
Microsoft's own documentation for Purview retention describes two actions: retain content so it cannot be permanently deleted, and delete content permanently. The NIST Guidelines for Media Sanitization define sanitization as a process that renders access to target data on the media infeasible. ISO 15489-1:2016 scopes records management to the creation, capture and management of records, their metadata and their systems.
Records, media, content. The unit is always the stored object.
Two failures, pointing in opposite directions
A statement with no retention class fails in both directions at once, and the two failures look nothing alike.
The answer outlives the record
You dispose of the source on schedule. The statement drawn from it is still in a slide deck, a wiki page, an email thread and a retrieval index. GDPR Article 5 asks that personal data be kept no longer than is necessary for the purpose, and that it be limited to what is necessary. The record is gone, so the schedule reports success. The content of the record is still in circulation, in a form the schedule never counted.
Worse, the answer now has no source to check it against. It will keep being given.
The proof goes out with the record
Run it the other way. A decision was made last March on the strength of an answer. The record behind that answer has since reached the end of its retention period and been destroyed, correctly. Now someone asks why the company said what it said.
What the logs hold
There is nothing to reconstruct it from. The EU AI Act asks deployers of high-risk systems to keep the logs those systems generate automatically for at least six months. That is a retention rule for the system's logs, not for the knowledge. A log can show that a question was asked and an answer came back. It cannot show that the document behind the answer was approved, current, or the only one on the subject.
Deletion does not travel downstream
This is not a new observation in machine learning. A survey of machine unlearning opens by noting that removing data from back-end databases is straightforward, but is not sufficient in an AI context, because models often remember the old data. That survey is about training data rather than retrieval, and the mechanics differ. The shape of the problem is the same. A delete on the store does not reach what was already derived from it.
A leak produced on purpose
Drata names a closer version of it among the practical challenges: deletion breaks down when files get renamed, copied to personal drives, or moved outside retention workflows. An answer is that same leak, produced on purpose, thousands of times a day.
The index is a copy nobody classified
There is a quieter version of this. A retrieval layer holds its own copy of the text it was pointed at, usually broken into chunks. That copy is derived from records, and it sits in a store that was set up by an engineering team rather than classified by a records team.
Ask who owns its retention class. In most companies the honest answer is that nobody has written one, because the index was built after the policy was last reviewed. It is worth asking before a regulator does.
What a retention schedule would have to add
Three things would have to be written down, and none of them appear in a standard policy. First, which statements in circulation depend on which records. Second, what happens to those statements when the record is destroyed. Third, what evidence survives the destruction, so a past answer can still be explained without the record.
The first one is the hard one. You cannot list what depends on a record unless something kept the dependency when the answer was made. Nothing in the retention stack does that, because the retention stack was built one object at a time.
Where the dependency has to live
This is the part Cognatum was built for, and it is worth saying plainly what it does and does not do. Knowledge stays where it lives, in the repositories each department already uses. Cognatum sits above them and governs the entries that people and AI systems answer from.
The Clean step compares each entry against its sources. When a source changes or goes away, every entry that depends on it is flagged and routed to a person. It notifies. It does not quietly rewrite the entry, and the last approved version stays live until somebody signs off on the next one.
What survives a retirement
When an entry is retired, it stops being served on every channel at once, and its full record, including the approval history, is retained. An answer given while that entry was live stays traceable afterwards. That is the third line above, and it is the one that makes a disposal defensible rather than merely complete.
The Cognatum Knowledge Loop uses a retention question as its own worked example, because this is where the mess shows up first. One deck says five years. One policy PDF says six. A chat thread says it depends and links nothing.
One thing to take to your next policy review
Pick a record class your schedule disposes of on a short clock. Ask your assistant a question that class answers. Then ask it for the approved entry, the named approver and the date.
If it answers the question but cannot produce those three, you have just watched a statement with no retention class leave the building. Your schedule will never find it, because your schedule is looking for records.
Sources
- Data Retention Policy: Definition, Examples & Best Practices (bigid.com)
- What Is a Data Retention Policy? Best Practices + Template (drata.com)
- Art. 5 GDPR, Principles relating to processing of personal data (gdpr-info.eu)
- Learn about retention policies and retention labels (learn.microsoft.com)
- NIST SP 800-88 Rev. 1, Guidelines for Media Sanitization (csrc.nist.gov)
- ISO 15489-1:2016, Records management, Concepts and principles (iso.org)
- EU AI Act Article 26, Obligations of deployers of high-risk AI systems (ai-act-service-desk.ec.europa.eu)
- A Survey of Machine Unlearning (arxiv.org)
- The Cognatum Knowledge Loop (cognatum.ai)
Common questions
Questions this raises.
Is an AI-generated answer a record under our retention schedule?
In most schedules it is not addressed at all. Retention classes are written for objects a records team can enumerate: files, messages, database rows, media. An answer assembled on request is not enumerated anywhere, so it inherits no class and no clock. That is a gap in the schedule rather than a decision to exclude it.
If we delete the source document, does the answer go with it?
No. Deletion acts on the stored object. Statements already drawn from it survive in decks, wiki pages, email threads, support macros and retrieval indexes. Unless something recorded which statements depended on that document, there is no list to work from and no way to recall them.
Should we retain AI prompts and outputs?
The EU AI Act asks deployers of high-risk systems to keep the logs those systems generate automatically for at least six months. Logs are a sensible minimum, and they answer a different question. They show that an exchange happened. They do not show whether the document behind the answer was approved, current, or the only one on the subject.
How is this different from a legal hold?
A legal hold suspends disposal so that material is preserved. This runs the other way: disposal happens on schedule, and the derived statement is the thing that fails to be disposed of. The two problems share a cause, which is that neither preservation nor destruction reaches answers.
Does Cognatum retain data we are required to delete?
No. Knowledge stays in the systems that hold it, and a retired entry stops being served on every channel at once. What is retained is the governance record: who approved an entry, when, from what source and at which version. An answer given while the entry was live stays explainable after the entry is gone.