How to let an AI agent cite the current version of a policy

Make the citation a property of the source, not a habit of the agent. Connect the agent to a versioned policy library that returns only the version in force, with its number, effective date and clause references. Require every policy-based answer to quote document, version and clause, and have the agent decline when no current version covers the question.

Why does the version matter, and who owns it?

Policies change. A staff member asking whether a payment needs a second approval, a customer asking why a request was refused, a reviewer checking a decision: each needs the answer under the policy in force at the time. An agent that cites an old version gives a wrong answer with false authority. One that cites nothing cannot be checked at all. That is the argument in a compliance agent should cite the rule.

Policy owners own content; the library owner owns versions and approvals; compliance owns the standard for how agents use policy. The technical connection to the library is covered in connecting an agent to your policy library. This playbook covers the citing behaviour that sits on top.

How do you set it up, capability by capability?

1. Person, library owner: exposes a "current version" query. Given a document identifier, the library returns the approved version in force, with version number, effective date and stable clause identifiers.
2. Person, library owner: exposes a separate "version on date" query for questions about the past, such as what applied to a decision made months ago. This is a distinct capability, granted only to agents that need it.
3. Person, compliance: writes the citation rule into the agent's brief. Every answer relying on policy names document, version and clause, and quotes the operative words.
4. Agent, depth 1: retrieves before answering. It never answers a policy question from memory or from a previous conversation; it queries the library each time.
5. Agent, depth 1: refuses when unsure. If the library returns no applicable clause, or two clauses conflict, the agent says so and routes the question to the policy owner.
6. Agent, depth 1: re-checks cached content. If it holds a copy for speed, it checks the version number before using it.
7. Person, compliance: samples answers for correct citation.

What is recorded, and which controls apply?

Citations in the record. Every answer in the agent's log carries its citation. Evidence: you can reconstruct exactly what the agent relied on, months later.

Version check at the moment of use. The agent's reliance on a clause is checked against the library at the time of the answer, not at the time it was cached. Evidence: version number and retrieval time per answer.

Refusal is an outcome. An agent that says "the policy does not cover this; I have passed it to the owner" has behaved correctly. Evidence: refusals logged and routed, which also shows gaps in the policy set.

Held when the action depends on the citation. Where an agent will take an action on the strength of a policy clause, such as approving a routine request, and the clause changed since its last certification, the action is held for a person until the agent is re-checked. Evidence: holds linked to version changes.

Revocation of stale access. If the library is migrated or restructured, old identifiers stop resolving and the agent refuses rather than guessing. Evidence: failed lookups logged.

What should you measure?

• Answers with complete citations, by sampling. The target is all of them.
• Answers citing a superseded version, which should be none.
• Quotes that do not match the cited clause. An agent can cite correctly and then paraphrase wrongly.
• Refusals per week, and what they reveal about policy gaps.
• Time from a new version to agents using it, which should be immediate.

Where does it usually go wrong?

Clause numbers that move. If the library renumbers clauses on every revision, a citation to "section 4.2" is ambiguous. Stable clause identifiers solve this.

Summaries cited instead of policies. Teams create helpful one-page summaries, and the agent cites those. Summaries are not policy. Keep them out of the agent's sources, or label them clearly.

Local addenda missed. The group policy says one thing, the country addendum another. The current-version query must return the addendum that applies to the asker's entity.

Old answers reused. An agent with conversation memory repeats an earlier answer after the policy has changed. Policy answers should always be re-retrieved.

Annual review drift. Policies past their review date are still in force but may be out of step with practice. The agent should mention when a cited policy is overdue for review; the review itself follows running the annual policy review.

The stance on compliance agents generally is in AI compliance agents.

Where Pryme fits

In Pryme, the policy library is a sealed source: each clause reaches the agent through the Pryme Context Protocol stamped with its document, version and approval, so the citation travels with the fact rather than relying on the agent to add it. Attest's grounded gate tests that answers rest on the cited text, and a change to a clause an agent depends on is flagged for re-certification.