Yes, to anyone who has to check or rely on the answer. Reviewers and auditors should see the exact source, version and passage behind every claim. Customers should see sources in plain terms, such as the section of their own policy. The deciding factor is not whether a citation appears, but whether the cited passage supports what the agent said.
Why do sources matter more for an agent than for a search tool?
A search tool hands you documents and leaves the judgement to you. An agent hands you a conclusion, and often acts on it. Without its sources, the person reviewing a draft has two options: trust it or redo the work. Neither is review.
Sources also make errors findable. If an answer is wrong, the source shows whether the agent misread a correct document, read an outdated one, or had nothing to go on. Each failure has a different fix.
Who should see what?
| Audience | What to show | What to leave out |
|---|---|---|
| Reviewer approving a held action | Exact document, version, passage, and where each figure came from | Nothing material |
| Auditor or supervisor | Everything the reviewer saw, plus the record of when each source was current | Nothing material |
| Customer | The plain-language source: their contract section, the published fee schedule, their statement | Internal system names, internal guidance, other customers' data |
| Colleague asking an internal question | The policy or procedure, section and version | Sources they are not cleared to read |
The customer row needs care. Some internal documents, such as fraud rules or underwriting guidance, should never be quoted to customers. The agent can still use them; it explains the outcome in terms the customer can check, drawn from documents they are entitled to see.
What is the difference between citing a source and being grounded in it?
An agent can attach a citation to a claim the source does not make. It may cite the right document but the wrong clause, or a passage that is close in wording but different in meaning. A citation that looks right is persuasive, which makes a wrong one worse than none.
So the control is not "show a link". It is testing that each claim follows from the cited passage. Our guide on testing whether an AI agent's answers are grounded in sources covers how to run those tests on real cases.
What does this look like in practice?
Take a home insurer's customer service team. A policyholder asks whether accidental damage to a laptop is covered. The agent answers that it is covered only if the optional accidental damage cover was added, and that it was not on this policy.
The reviewer sees three sources: the policy wording, version and section on accidental damage; the policy schedule for this customer, showing the options taken; and the product's published cover summary. The customer sees: "Your policy wording, section 3, explains that accidental damage is optional cover. Your schedule shows it was not added." The internal claims handling guide, which the agent also consulted, is not mentioned.
When the reviewer checks, she notices the wording cited is the previous version. The answer happens to be the same under the current one, but she flags it. The source register had not been updated after a wording change, which is the problem keeping an AI agent's sources up to date is meant to prevent.
Which controls make showing sources useful?
- Provenance on every fact. Each source carries where it came from and how far to trust it; see stamping sources with provenance.
- Versioned sources. The citation names the version, so a later change does not rewrite history.
- Grounding tests on real cases before the agent goes live and at each recertification.
- No source, no answer. If the agent cannot find a source for a factual claim, it says so and routes the question to a person.
- Audience rules. Which documents may be quoted to customers is listed, not left to the agent.
Where Pryme fits
In Pryme, every fact reaches the agent through the Pryme Context Protocol, stamped as sealed, signed, declared or external. Reviewers see that stamp alongside each claim. Attest's grounded gate tests whether answers follow from their sources, and the certificate records the result, which is what you show when asked how an auditor knows an agent is certified. More in our guide to agent governance.
Questions people ask
Does showing sources slow the agent down?
Not noticeably. The agent already retrieves the sources to answer. Showing them is a matter of keeping the reference attached rather than discarding it.
What if two sources disagree?
The agent should show both and say which it relied on and why, usually the higher provenance class or the newer version. If that does not settle it, the question goes to a person.
Should the agent show sources for refusals too?
Yes. A refusal is a decision, and the rule behind it is a source. Citing it makes the refusal easier to accept and easier to challenge.