Delegation of authority stops before the answer

Most regulated firms have a delegation of authority document. It names who may approve what. It sets a limit on each one. A director signs off to one threshold. A committee signs off above it. The matrix gets reviewed, dated and filed.

Then an AI assistant answers a question about policy. Somebody acts on it. Ask which line of that matrix covered the answer. There is no line.

What a delegation actually grants

A delegation of authority is a written document, not a habit. The University of California sets out what one must state. The source of the power. What is being handed over. And the post that gets it.

Deloitte puts it in plainer terms. A delegation hands a named person a set of tasks, plus the right to decide on them.

Two things matter there. The power is bounded. And it lands on a person.

The answer nobody authorized

A retrieval pipeline appears nowhere in that structure. Nobody wrote it a delegation. Nobody set its limit. No committee reviewed its scope.

That is not a drafting oversight. It is a category problem. A delegation runs to a position held by a person who can be asked to explain a decision. A pipeline holds no position.

What happened instead is quieter. The firm never delegated authority to the assistant. It simply began treating the assistant's output as though somebody with authority had produced it.

Accountability does not transfer

Delegation has a standard limit, and the frameworks all state it the same way. Authority can be delegated. Accountability cannot. The person who delegated stays answerable for the result.

So when an AI answer turns out to be wrong, the accountability has not moved anywhere. It still sits with whoever owns that policy area. They are answerable for a statement they never saw, drawn from a document they may not know exists.

Human oversight is not a delegation

Firms usually answer this with oversight. A person reads the answer before acting on it. That helps. It is not the same thing as delegated authority.

Delegated authority is specific and prior. It says this position may approve this class of decision, up to this limit. Oversight at the point of use is general and late. The reviewer sees an answer, not the approval history behind it.

Often the reviewer has nothing to review against. They cannot see who approved the source, which version it was, or whether a second document says the opposite.

What the AI Act asks of the overseer

The EU AI Act is specific on this for high-risk systems. Article 14 requires that the people assigned human oversight can grasp the system's limits, read its output correctly, and decide to disregard, override or reverse it.

Automation bias, named in the law

It also names the failure mode. Article 14(4)(b) asks the overseer to stay aware of the tendency to over-rely on output, which the Act calls automation bias. It flags this for systems that supply information for a decision a person then takes.

And where the Act does demand a second pair of eyes, under Article 14(5) for one narrow class of biometric systems, it asks for two natural persons with the necessary competence, training and authority. Authority is written in as a requirement, not assumed.

What the rulebooks already say

FINRA has described this exact scenario. Notice 24-09 observes that generative AI tools may let an associated person easily locate and query a firm's policies and procedures or forms.

The Notice then lists what supervisory policies should address: technology governance, model risk management, data privacy and integrity, and the reliability and accuracy of the AI model.

Every item on that list is about the system. Not one is about who approved the policy the system just quoted.

The control frameworks have the same shape

COSO's Internal Control framework is what most firms build on, and it works by assigning authority and responsibility among people. COSO issued supplemental guidance on internal control over generative AI in 2026, so the gap is recognized.

Section 404 of Sarbanes-Oxley requires management to assess the effectiveness of internal control over financial reporting, and the auditor to attest to that assessment. An assessment needs evidence. An answer with no approver and no date supplies none.

NIST's AI Risk Management Framework takes the same line. Its Govern function puts documented accountability and clear roles underneath everything else in the framework.

Citing a source is not the same as approval

The obvious objection is that the assistant quotes real documents, so the answer must be sound. The evidence says otherwise.

A preregistered study of the leading AI legal research tools found that systems built on retrieval, from LexisNexis and Thomson Reuters, still produced false information between 17% and 33% of the time. The vendors had claimed to eliminate it.

Retrieval tells you where text came from. It does not tell you whether anyone with authority approved that text, or whether the text was still current when it was read.

Put the authority back on the knowledge

The fix is not a new delegation for the pipeline. You cannot delegate authority to something that cannot be held to it.

The fix is to record the delegation where it belongs, on the knowledge itself. A governed entry carries a named approver, an approval date, and the source it came from. When an assistant answers from that entry, the chain runs from the answer back to a person who held the authority to approve it.

One step stays human

That step is Approve, step six of the Cognatum Knowledge Loop. AI runs the other seven: capture, structure, clean, enrich, improve, deploy and serve. A named person holds the sixth, and nothing goes live without that signature.

Two other behaviors matter for a delegation argument. Where two approved entries disagree, Cognatum holds both apart and routes the conflict to a named person rather than picking a winner. And when a source changes, it flags the entry and queues a revision for the gate.

It does not quietly rewrite the approved answer, and it is not current on its own. A person still decides. That is the point.

Four questions to put to your own firm

• Which position in your delegation of authority covers a statement of company policy made to a client or a regulator?
• When an AI assistant makes that statement, whose authority is it running on?
• For any answer given in the last quarter, can you produce the approver, the date, and the version of the source?
• When a policy changed last month, who was told that answers derived from the old version are now stale?

If the first question has an answer and the second does not, that is the gap. Your company's knowledge isn't missing. It's unusable. Cognatum changes that.