Your acceptable use policy stops at the prompt

Ask a regulated company whether it governs AI use, and you will usually be handed one document. The acceptable use policy. It is the oldest rulebook in the building. It already covers laptops, email and personal devices. Some time in the last two years, a section on AI tools was added to it. That section is real work, and it does something useful. It also stops where every version of it stops. At the moment the user types.

What an acceptable use policy controls

TechTarget defines an acceptable use policy as a document that sets out the rules a user must agree to before using a company network, the internet or other computing resources. Wikipedia calls it a formal set of guidelines. It states the terms of access, and the behavior that is allowed, limited or banned. Both point the same way. The policy is addressed to a person and to a resource.

What the clauses actually say

Look at what the clauses say. Do not break the law. Do not disrupt network security. Do not send spam. Do not take a vendor's intellectual property. Report any attempt to break into your account. Accept the stated discipline if you violate the policy. The usual coverage areas follow the same logic: code of conduct, social media, internet use, cybersecurity, non-employee users, access to confidential data and personal devices.

The AI section has the same shape

TechTarget's own guide to an AI acceptable use policy lists eight essential elements. Purpose and scope. User responsibilities. Prohibited uses. Data governance. Security requirements. Compliance and legal obligations. Reporting and consequences. A review and update process.

Data governance there means rules for data access, sharing and processing. Not one of the eight asks whether the content the approved assistant reads was approved by anyone. None asks whether it is the current version. None asks what happens when two documents in the same system disagree.

The sentence that never gets written

A well drafted AI clause says two things. Do not paste confidential data into a public tool. Use only the tools on the approved list. Both are good rules. Both are about the input, and about the identity of the tool.

The missing rule is about the output. It would read something like this. An answer used in a regulated decision must trace to an approved entry, with a named approver, a date and a source. Almost no acceptable use policy says that. The reason is not careless drafting.

Why the gap is structural

An acceptable use policy is a conduct instrument. It binds people, and it is enforced through discipline. It also sits in a library of policies built the same way. The SANS and Cybersecurity Risk Foundation template library is a good sample. Its policies are named for assets and activities: access management, data inventory, software, cloud service providers, privileged accounts, privacy, perimeter networks.

Built around assets and actors

NIST SP 800-53 Rev. 5 works on the same axis. It is a catalog of security and privacy controls for information systems and organizations. Controls protect systems, data and people. A library built that way has nowhere to put a rule about whether a paragraph in a document is still true. Knowledge is not an asset class, and it is not an actor. So it falls between the shelves.

What regulators ask for instead

The EU AI Act is specific about what the deployer of a high-risk system owes. Article 26(1) asks for technical and organisational measures so the system is used in line with its instructions for use. Article 26(2) asks that human oversight go to natural persons with the necessary competence, training and authority. Article 26(6) asks that logs be kept for at least six months. Article 4 asks providers and deployers to support the AI literacy of their staff.

The one clause that touches content

Article 26(4) comes closest. Where the deployer controls the input data, it must ensure that data is relevant and sufficiently representative in view of the intended purpose. Read that carefully. Relevant and representative is a fit test. It is not a test of whether the document was approved, whether it is the current version, or whether another document says the opposite.

Where the obligation stays

No tool confers compliance with the Act, and no vendor should say otherwise. What the text does do is tell you what a reviewer will look for. FINRA takes a similar line in Regulatory Notice 24-09. Its rules are meant to be technology neutral, and they keep applying when a firm uses generative AI. The obligation stays with the firm. It does not move to the tool, and it does not move to the policy that describes the tool.

The failure a conduct policy cannot catch

There is good evidence that a sanctioned tool can be confidently wrong. A preregistered study of leading AI legal research products found that Lexis+ AI, Westlaw AI-Assisted Research and Ask Practical Law AI each produced hallucinations between 17 and 33 percent of the time. The vendors had described their systems as eliminating or avoiding hallucinations.

Now test one of those answers against an acceptable use policy. Approved tool. Authorized user. No confidential data sent anywhere it should not go. No prohibited use. The answer passes every clause, and it is still wrong. A conduct policy cannot see that, because nothing about the user's conduct was at fault.

What a knowledge clause looks like

The fix is not a longer acceptable use policy. That document is doing its job. The missing piece is a second artifact that governs the knowledge itself, and a system that can hold it.

What the second artifact has to do

Cognatum is a knowledge base built for that job. Every entry carries a named approver, a date and a source, so an answer can be traced rather than trusted. Where two documents disagree, it presents both and shows where the disagreement sits. It does not pick a winner. When a source changes, it notifies the people who own what depended on it. It does not quietly rewrite anything, and it is never current by itself.

The Cognatum Knowledge Loop runs eight steps across four phases. AI runs seven of them: the finding, the gathering, the cleaning and the surfacing. A named person approves. That gate is the step that stays human, and it is where authority actually attaches.

A test you can run this week

Open your acceptable use policy and read the AI section. Count the clauses that would change if every document your assistant can reach were six months out of date. For most companies the count is zero. That is the gap, and it is not a drafting problem. It is a missing system.

Your company's knowledge isn't missing. It's unusable. Cognatum changes that. You can see how the Loop works, step by step, on the Cognatum Knowledge Loop page.