Yes, when the change alters what the agent knows or trusts. Adding, removing or replacing a source, or issuing a new version of a policy it reads, takes the agent outside its certificate. New records arriving in a certified source, such as fresh ledger transactions, do not. The line is between changing a source and using it.
Why do sources matter so much to a certificate?
An agent's answers are only as good as what it reads. Its certificate was earned with a particular set of sources, of a particular quality, and the grounded gate in particular measured whether its answers traced back to them. Change the sources and that measurement no longer applies.
Sources also shape refusals and escalations. An agent that reads a complaints policy knows which complaints must go to a person. Replace that policy with a new version that adds a category, and the agent may not escalate the new category, even though nothing else about it changed.
Which source changes void the certificate?
| Change | Voids the certificate? | Why |
|---|---|---|
| New records in an existing system of record | No | The agent was certified to read this source; its content is expected to grow |
| A policy or procedure document replaced with a new version | Yes | The rules the agent applies have changed |
| A new source added | Yes | The agent now knows things it was not tested on |
| A source removed | Yes | Answers that relied on it may now be ungrounded |
| A source's provenance changes, such as moving from a system of record to an emailed extract | Yes | The agent's trust in the facts should change |
| A source's format or structure changes | Usually yes | Retrieval and reading can fail in new ways |
| A source becomes unavailable for a period | Not voided, but the agent should hold | It cannot ground answers it cannot check |
The certificate should list its sources by name and version, so it is clear when one has changed. The fields are set out in the AI agent certificate template, and the wider question of what a certificate covers is in what an AI agent certificate covers.
Why does provenance matter, not just content?
Because trust should follow where a fact came from. A balance read from the core ledger deserves more weight than the same number typed into an email. If a source moves from one to the other, perhaps because an integration broke and someone substituted a spreadsheet, the agent is now acting on weaker facts. Its certificate assumed the stronger source.
A useful way to think about it is in classes. Facts issued by the system of record itself are strongest. Facts a named, identifiable party answers for come next. Facts a customer or person declared, unverified, come after that. Facts that arrived from outside, such as an emailed invoice or a web page, are weakest. A change that moves a source down this order should always trigger a rerun.
An illustrative example
Take an insurer's agent certified to answer brokers' questions about policy wording. It reads the current policy wordings library.
The product team issues a new household wording with a changed flood exclusion. The library is updated overnight. The agent's certificate is treated as void for household questions from that moment. A rerun on the certification set, with cases added to probe the new exclusion, shows the agent answers correctly from the new wording. The capability is recertified the next day. Meanwhile, the agent's motor answers continue, because their sources did not change.
Compare a second change in the same week: new policies are added to the policy administration system. That is new data in a certified source, and nothing needs rerunning.
How should teams handle this in practice?
- Version every source the agent reads, and record the version on the certificate.
- Link source owners to agent owners. The team that updates a policy library should know which agents read it.
- Trigger reruns automatically when a listed source changes version.
- Scope the impact. If only one capability uses the changed source, only that capability needs to drop while it is rerun.
- Add cases for the change. A rerun on old cases may not test the new content.
A model update raises the same issue from a different direction; see whether a model update voids an agent's certificate. At renewal, the certificate renewal checklist for AI agents includes a source review.
What stays with a person?
Deciding whether a change is a version change or routine data, when it is not obvious, and accepting the rerun results before depth is restored. The idea of an agent certificate depends on someone owning that judgement. The full framework is in what it means to certify an agent.
Questions people ask
Does fixing a typo in a policy document void the certificate?
Strictly, it is a new version. In practice, a quick rerun on the affected capability is cheap and removes the argument about whether the change mattered.
What if a source is updated many times a day?
Separate the parts that change constantly, which are data, from the parts that set rules, which are versioned. Certify against the rules version.
Should the agent know which version of a source it is reading?
Yes. Its answers should record the source and version used, so a reviewer can tell which rules applied to which answer.