Cognatum

System of record

Policy attestation stops at the reader

Policy attestation binds a named person to a specific policy version on a specific date. It is the strongest proof of knowledge most companies own. The AI answering the same policy questions signed nothing.

Cognatum Team · Oct 16, 2026 · 6 min read

Every regulated company runs the same ritual. A policy gets approved. It goes out to the people it covers. Each of them clicks to confirm they have read it. The record that comes back is small and precise. It names a person, a policy version and a date.

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

That record is the strongest proof of knowledge most companies own. It is also the only one. The assistant your team now asks policy questions to has never attested to anything.

What attestation actually records

The idea is simple to state. A worker confirms, on the record, that they have read a policy and agree to follow it. VComply's definition names the four parts that make it work: clear wording, tracking of who signed and when, renewal on a cycle, and records fit for an audit.

OneTrust frames it the same way. An attestation shows that staff have read a policy and agree to abide by it. Run that by email and you lose two things at once. You lose version control. You lose any clear read on who replied.

One control binds person, version and date

KnowledgeCity frames policy management as a chain. A written policy connects to a training record. That connects to an employee attestation. That connects to an audit-ready record. Its blunt observation is that regulators ask for the chain, not the training certificate on its own.

The load-bearing link in that chain is the attestation. It is the only artifact that answers three questions at once. Which version was in force. Who acknowledged it. On what date.

Where the chain breaks

The same analysis lists the usual failures, and every one of them is a failure of binding. The version on the intranet does not match the version people signed. Nobody owns the review cycle. Signatures were gathered by department rather than by role plus version. Training went out with no reference to the policy version it was meant to enforce.

Each gap looks small on its own. Each one is enough to fail a finding.

The assistant signed nothing

Now set the assistant beside that record. Someone asks it which approval threshold applies to a new vendor. It reads the policy library and answers in a sentence. The answer is quick, confident and probably right.

Nothing in that exchange produced an attestation. There is no version binding. There is no acknowledgment. No named party accepted responsibility for the content on any date. The person who asked walks away with an answer. The company holds nothing it could hand an auditor.

What the answer carries instead

It carries fluency. It may also carry a citation to a file. A file reference is not a version, and a version is not an approval. A policy can be retrieved correctly, quoted accurately, and still be the wrong one, because the document that replaced it last March lives in a different folder.

Re-attestation reaches people, not answers

Policy change is where the cost shows up. When a policy is updated, a mature program re-triggers attestation for everyone in scope. The prior version gets archived with the date range in which it applied. That part works.

Nothing re-triggers for the answers. Every statement the assistant produced from the superseded version is still in circulation. It sits in sent mail, in meeting notes, in a client's inbox, in a slide somebody will reuse next quarter. No workflow recalls it, because no workflow knows it exists.

Azure attests better than your policy library

Microsoft makes the contrast awkward. Where an Azure Policy control cannot be tested automatically, Azure accepts a manual attestation, and that attestation is a structured record. It carries a compliance state, an owner, a comments field, an array of links to evidence, the date the evidence was assessed, and an expiry date after which the state reverts.

So a resource group gets an owner, an evidence link, an assessment date and an expiry. The answer your assistant just gave about the cybersecurity policy gets none of the four.

Grounding is not attestation

The standard objection is that retrieval already solves this. Point the model at approved documents and the answer inherits their authority.

The first preregistered evaluation of retrieval-based legal research tools found otherwise. Tools from LexisNexis and Thomson Reuters, marketed as eliminating or avoiding hallucinations, hallucinated between 17 and 33 percent of the time. Retrieval lowers the error rate. It does not sign anything.

Where the obligations are heading

The instruments keep converging on named, competent humans. ISO/IEC 27001:2022 sets the requirements for an information security management system. ISO/IEC 42001:2023 did the same for artificial intelligence.

The EU AI Act is more specific for high-risk systems. Article 26 requires deployers to assign human oversight to natural persons with the necessary competence, training and authority. It also requires them to keep the system's logs for at least six months. No software confers compliance with any of this. What software can do is supply the evidence those duties ask for.

What an attestation-grade answer carries

Four fields, and they are the four the attestation record already holds.

The entry the answer came from, not the folder it sits in.

The version of that entry, with the dates it was in force.

The named person who approved it.

The date of that approval.

Add a fifth for the AI case. Where two approved entries disagree, the answer should say so and show both, rather than choosing a winner quietly.

A test you can run this week

Take a policy your people attested to in the last cycle. Pull the attestation record. Note the version they signed and the date they signed it.

Then ask the sanctioned assistant a question that policy answers. Ask it for the entry, the approver and the date. If what comes back cannot be matched to the version in that attestation record, you are running two accounts of the same policy inside one company. Only one of them is signed.

How Cognatum closes the gap

Cognatum is a knowledge base that runs on its own framework, the Cognatum Knowledge Loop: eight steps, four phases, one human gate. AI runs seven of the eight. It captures, structures, cleans, enriches and surfaces. Approve is the step it does not run. A named person decides what is true, and can rework an entry before approving it.

Notify, not rewrite

Where two approved entries conflict, the loop presents both and shows where the disagreement sits. When a source changes, Cognatum notifies the owners of what depended on it. It does not rewrite entries by itself and it is never simply current. Every approved entry then serves people and AI systems from one place, carrying its approver, its date and its source.

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

Common questions

Questions this raises.

What is policy attestation?

Policy attestation is the process by which an employee formally acknowledges that they have read, understood and agreed to comply with a specific policy. The useful part is the record it produces, which ties one named person to one policy version on one date and can be produced on demand during an audit.

Does an attestation prove someone understood the policy?

No. It proves receipt and formal acknowledgment. Organizations that need more than that pair attestation with training assigned against the same policy version, or with a knowledge check. The attestation record is still the only artifact that pins the version and the date.

Why does an AI answer not count as an attestation?

Because nobody signs it. An attestation names a responsible party, a version and a date. A generated answer names none of the three. It may cite a file, but a file reference is not a version and a version is not an approval, so there is nothing to produce when an auditor asks who stood behind the statement.

What happens to answers produced from a policy that has since been superseded?

In most companies, nothing. Attestation workflows re-trigger for the people in scope and the old version is archived with its date range. The statements already derived from that version stay in circulation with no recall path, because no system is tracking what depended on the source.

Does retrieval from approved documents make an AI answer trustworthy?

It helps and it is not sufficient. The first preregistered study of retrieval-based legal research tools found that products marketed as hallucination-free still hallucinated between 17 and 33 percent of the time. Retrieval narrows the error rate without establishing who approved the content or which version was in force.

Knowledge governed. Intelligence everywhere.

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