AI governance
Data classification sorts by harm, not accuracy
Most regulated companies can produce a data classification policy in a minute. Four tiers, named owners, handling rules. Every tier answers one question: what happens if this leaks. None asks whether the content is right.
Cognatum Team · Oct 12, 2026 · 6 min read
Ask a regulated company whether it governs its own information. A data classification policy is usually the first thing produced. It arrives fast. Four tiers, sometimes five. Public, internal, confidential, restricted. Named owners, named stewards, handling rules for each level. For many firms it is the fullest map of their own content that exists.
Cognatum governs the entry
source · version · approver · permissions
It is also built from a single question. What happens if this gets out.
What the tiers actually measure
Pick any published scheme and the ladder has the same shape. Fortinet's glossary sorts data into high, medium and low tiers. Social Security numbers sit at the top. Public web pages sit at the bottom. State and college policies run the same ladder under other names. The dividing line is always harm on exposure.
Who owns the question nobody answers
The roles follow the same logic. An owner approves the tiers. A steward applies them. A custodian builds the safeguards. A user follows the handling rules. Fortinet's own list of things to settle when writing the policy asks who is in charge of data integrity and accuracy. The scheme names an owner for that question. Then it never answers it.
One question, asked of every document
That is a real control and it works. A file marked restricted gets encryption, access limits and audit logs. A file marked public gets none of that. The scheme does just what it was built to do.
What it never does is ask a second question. Is this document right.
Integrity means something narrower than it sounds
The obvious objection is that classification already covers correctness, because integrity is one of the three security objectives. FIPS 199 categorizes federal information by confidentiality, integrity and availability. ISO/IEC 27001 rests on the same triad.
Read the FIPS wording closely. It concerns compromise through unauthorized access, use, disclosure, disruption, modification or destruction. Integrity there means nobody altered the file without permission. A procedure approved in 2019, untouched since, and quietly superseded last March has perfect integrity. Nobody tampered with it. It is simply wrong.
Accuracy appears once, in a narrow place
There is one place where a legal duty of accuracy attaches to content. GDPR Article 5 requires that personal data be accurate and, where necessary, kept up to date, and that reasonable steps be taken to erase or rectify inaccurate data.
Note the scope. It covers personal data about a person. It says nothing about the retention schedule, the pricing policy or the validated method your staff answer questions from. For everything that is not personal data, no comparable duty of currency sits anywhere in the classification stack.
The label travels, the approval does not
Here is the part worth sitting with. The machinery to carry a property from a source document into an AI answer already exists, and it already works.
Microsoft describes a sensitivity label as a stamp stored in a file's metadata, one that stays with the content no matter where you save or store it. Copilot and its agents read those labels. When a conversation draws on items of different sensitivity, the user sees the label with the highest priority, which is normally the most restrictive one. Where a label applies encryption, the assistant returns content from an item only if the user holds the right to copy from it.
Two axes travel, and neither is approval
So the pipeline does inherit properties from its sources. It inherits sensitivity. Microsoft's documentation notes that an item can carry a sensitivity label and a retention label at the same time. Those are the two axes that travel. Neither is approval state, version or effective date.
An answer assembled from four documents needs more than the strictest of four labels. It needs the approval state of each source, the effective date of each, and a flag where two of them disagree. None of that exists as a field, so none of it can be inherited.
Why this lands harder on an answer
A person reading a document picks up context the label never supplied. They see the revision number in the footer, the date on the cover page, the name in the approval block. They notice that nobody has touched the file in four years.
An answer strips all of that away. The retrieval layer reads across tiers in one motion. A confidential procedure from 2019 and an internal memo from last month arrive in the same paragraph, in the same confident register. The label on the output tells the reader how carefully to handle it. It tells them nothing about whether to believe it.
Good sources are not the same as a right answer
The first preregistered evaluation of commercial AI legal research tools found that products from LexisNexis and Thomson Reuters each hallucinated between 17 and 33 percent of the time, despite vendor claims of eliminating hallucinations. Those tools retrieve from curated, high quality legal sources. Retrieval from good material is not the same thing as a correct answer.
What the regulator asks, and what it does not
The EU AI Act comes closest to placing an obligation on the content a deployed system works from. Article 26 requires deployers of high risk systems to follow the instructions for use, assign human oversight to people with the necessary competence, training and authority, and keep logs. Where the deployer controls the input data, that data must be relevant and sufficiently representative in view of the intended purpose.
A fitness test, not an approval test
That is a fitness test. It is not an approval test, a currency test or a conflict test. NIST frames its own data classification work around knowing what data you have, what its characteristics are, and what security and privacy requirements it has to meet. Security and privacy. The same two axes again.
A second axis, and who holds the pen
Cognatum is a knowledge base organized around the axis the classification stack leaves out. Knowledge is captured, structured, cleaned, enriched and then approved, and the approval is a named person rather than a model. AI does the finding, gathering and surfacing. It routes the item to a human, who decides what is true and can rework the wording before signing off.
What an entry carries
Entries then carry an approver, a date and a source, and the system keeps track of where the knowledge stood when an answer was given. Where two documents say contradictory things, both are surfaced along with the point of disagreement rather than one being picked quietly. When a source changes, Cognatum notifies the people whose answers depend on it. It does not rewrite them on its own.
None of this replaces a classification policy. Sensitivity and accuracy are different questions and both need answering. The point is narrower. Most companies have built a careful answer to the first one and nothing at all for the second.
A test you can run this week
Find a document your scheme classifies as confidential or restricted, one that decisions rest on. Ask the assistant your company has sanctioned a question that document answers.
Read what comes back. Then ask a second question. Which approved entry did that come from, who approved it, and on what date. If the first answer arrives and the second one does not, your content is classified by sensitivity and by nothing else.
Sources
- What Is Data Classification? (fortinet.com)
- Data Classification Practices (nccoe.nist.gov)
- FIPS 199, Standards for Security Categorization of Federal Information and Information Systems (csrc.nist.gov)
- ISO/IEC 27001:2022, Information security management systems (iso.org)
- Learn about sensitivity labels (learn.microsoft.com)
- Art. 5 GDPR, Principles relating to processing of personal data (gdpr-info.eu)
- Article 26, Obligations of deployers of high-risk AI systems (europa.eu)
- Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools (arxiv.org)
- The Cognatum Knowledge Loop (cognatum.ai)
Common questions
Questions this raises.
What is a data classification policy?
It is the document that sorts company information into tiers by sensitivity, usually something like public, internal, confidential and restricted, and sets handling rules for each tier. It also names who classifies data and who is accountable for it. Every tier is defined by the harm that would follow if the content were exposed.
Does data classification tell you whether a document is accurate?
No. Integrity, in the security sense that FIPS 199 and ISO/IEC 27001 use, means the content has not been altered without authorization. A document can be perfectly intact, correctly labeled and three versions out of date all at once. The one place a legal duty of accuracy attaches is GDPR Article 5, and that is scoped to personal data.
Do sensitivity labels carry over into AI answers?
Yes, and that is the interesting part. Microsoft 365 Copilot and its agents read sensitivity labels and, in a conversation that draws on several items, show the label with the highest priority. So the pipeline already inherits a property from its sources. It inherits sensitivity, and alongside it retention. It does not inherit approval state, version or effective date, because those are not fields anywhere in the chain.
Is this just access control for AI assistants?
No. Access control answers who may see a document. This is a different question: whether the document the assistant just answered from was approved, is current, and does not contradict another document in the same library. An assistant can be perfectly scoped and still give a confident answer from superseded material.
Do we have to replace our data classification policy?
No. Sensitivity and accuracy are separate axes and both matter. The gap is that most organizations have a detailed, owned, audited answer for the sensitivity axis and no equivalent for the other one. The work is adding the second axis, which means an approver, a date, a source and a status on each entry, plus a way to surface conflicts and flag sources that have moved.