How to answer staff policy questions with an AI agent

Explaining a policy is agent work; granting a permission is not. It answers staff questions from the current approved policies, quoting the passage and version, at depth 3 once certified. Where the answer is "you need approval", it drafts the request for the compliance team to decide. Whistleblowing questions go straight to the confidential channel.

What is the process, and who owns it?

Staff ask compliance the same questions every week. Can I accept tickets from a supplier? Do I need to pre-clear a share purchase? What do I do if a customer offers a cash tip? How long do we keep call recordings? The answers are in policy documents, but those documents are long, and staff often ask a person instead. That person's time is the scarce resource.

The compliance function owns the policies and the answers. Each policy has an owner who keeps it current. The agent works as the first point of contact. It is not a substitute for compliance advice on hard cases, and it is not a channel for granting permission. That line follows from our position on AI compliance agents: second line, never the last word.

What are the steps, and at what depth?

1. Agent, depth 1: reads the question and identifies which policy or policies apply.
2. Agent, depth 1: retrieves the current version. Only approved, in-force policies are sources. Drafts and superseded versions are excluded.
3. Agent, depth 3: answers what the policy says. In plain words, quoting the passage, naming the policy and version. Once certified, these answers go directly to the member of staff.
4. Agent, depth 1: spots requests for permission. "Can I accept this?" is asking for a decision. If the policy requires approval, the agent says so and offers to draft the request.
5. Agent, depth 2: drafts the approval request. With the facts the staff member gave and the policy passage. Compliance decides.
6. Agent, depth 3: records declarations, such as a gift register entry, where the policy only requires the gift to be declared and the staff member asks for it to be logged.
7. Agent, depth 1: routes sensitive questions. Whistleblowing, harassment and suspected misconduct go straight to the confidential channel, with no detail kept beyond what that policy allows.
8. Person, compliance: answers what the policy does not cover, and decides every approval or exception.

Which controls apply, and what evidence does each leave?

Answers, not approvals. The agent's role has no capability to grant permission. A sentence like "you may accept the gift" is outside its scope, and certification tests for it. Evidence: scope-gate results and the filed answers.

Quoted passage and version. Every answer names its source. If the policy changes, you can find every answer given under the old version. Evidence: the citation on each filed answer.

Approved sources only. The policy library is controlled by the policy owners. When a policy is replaced, the old version stops being a source that day.

Confidential routing. Questions about raising concerns are routed without being stored in the general record beyond the routing itself. The whistleblowing policy decides what may be kept.

Policy gaps reported. Questions the policies do not answer are logged for policy owners. Evidence: the gap log and the owner's response.

Staff using the agent should know the rules for using it too; the acceptable use policy template for staff using AI agents covers that.

What should you measure?

• Questions answered with a citation, against those routed to compliance.
• Answers later found wrong, by policy. Often a sign the policy itself is unclear.
• Approval requests drafted, and how quickly compliance decides them.
• Policy gaps logged and closed. A useful by-product: the agent shows policy owners where their documents fail staff.
• Repeat questions on one rule. Many questions about the same passage usually mean the rule needs rewriting, not just explaining.
• Staff treating answers as permission. Found through spot checks on registers and approvals.

Where does it usually go wrong?

The answer sounds like a yes. "The policy allows gifts below the limit, so that should be fine" reads as permission. The agent should state the rule and, where a judgement is needed, say who makes it.

Stale policies. An old version left in the library gives a confident, cited, wrong answer. Version control in the source library matters more than anything in the agent.

Different policies with the same word. Staff policies are not insurance policies. An agent that also serves customers must keep those sources apart; answering insurance policy questions is a separate capability with separate sources, as are customer queries such as questions about a customer's bill.

The wrong entity's policy. Groups often have different policies for different subsidiaries or countries. The agent must know which policy applies to the person asking, from their employment record, and say which one it used. A correct answer from the wrong entity's policy is still wrong.

Sensitive questions logged in full. A staff member asking how to report their manager deserves the confidential channel, not an entry in a general log.