Break every answer into its factual claims and trace each to a source the agent was given. Then make it harder: remove the needed fact, give it conflicting sources, have the customer assert something false, and stretch the conversation. A grounded agent says it does not know when sources are silent, and prefers the stronger source when they conflict.
What does "stays grounded" add to "is grounded"?
A grounded answer is one whose claims all trace to the sources. Staying grounded is about whether that holds under conditions that tempt the agent to fill gaps: missing information, conflicting information, a confident customer, a long conversation. Most agents are grounded when the answer is sitting in the source. The failures appear when it is not.
The claim-by-claim checking method itself is covered in how to test whether an AI agent's answers are grounded in sources. This article is about the conditions to test it under.
What are the practical steps?
- List the sources the agent may use for the capability, with their versions.
- Build baseline cases where the answer is plainly in the sources.
- Build gap cases where the needed fact is absent. The right answer is to say so, and hand off if needed.
- Build conflict cases where two sources disagree, such as an old policy summary and the current policy.
- Build assertion cases where the customer states something the sources contradict.
- Build long cases, such as a conversation of many turns or a long document, where early facts must be carried accurately to the end.
- Run repeatedly and break each answer into claims.
- Trace each claim to a source and version, and mark it supported, unsupported or contradicted.
Which conditions catch agents out?
| Condition | Grounded behaviour | Common failure |
|---|---|---|
| Fact missing from sources | Says it does not know, offers a route | Supplies a plausible figure or rule |
| Two sources conflict | Uses the stronger or current source and notes the conflict | Blends them, or picks the more convenient one |
| Customer asserts a false fact | Checks the source and politely corrects | Accepts the customer's version |
| Long conversation | Keeps early facts accurate | Mis-states a figure given many turns earlier |
| Source paraphrased | Keeps the meaning, including conditions | Drops an exception or qualifier |
| Calculation needed | Uses a tool or source value | Calculates in its head and gets it wrong |
The calculation row is easy to miss. An agent that adds up figures itself, rather than reading a total from the system of record, produces claims that are not in any source.
How should conflicts between sources be resolved?
By where each fact came from. A fact issued by the system of record itself, such as the core ledger or the policy library, should outrank one a named party answers for, which outranks one a customer declared, which outranks one that arrived from outside, such as an emailed document or web page. Tests should check that the agent applies this order, and that it says which source it used when they disagree.
A customer saying "my limit is 5,000" is a declared fact. If the system of record says 3,000, the agent should go with the system and tell the customer what it shows.
An illustrative example
Take a travel insurer's agent that answers policyholders' questions about cover. Baseline cases pass easily.
In gap cases, a customer asks whether a specific adventure activity is covered, and the wording does not mention it. On the first run the agent says it is "likely covered under general leisure activities". That is an unsupported claim and fails. In conflict cases, the agent is given the current wording and an older product summary, which had a higher baggage limit. It quotes the summary. Fail. After the brief is revised to state that only the current wording counts and that silence means referral, both groups pass. Assertion cases, where the customer insists their policy includes cancellation cover when it does not, pass from the start.
How does grounding relate to other gates?
When an agent cannot ground an answer, it should usually refuse or escalate, which is why grounding failures often appear alongside those. Score them separately; see how to test whether an AI agent refuses correctly and how to test whether it escalates correctly. Scope is distinct again: an answer can be grounded and still outside the agent's role, covered in how to test whether an AI agent stays in scope.
Where Pryme fits
The Pryme Context Protocol stamps every fact with its source and provenance class, sealed, signed, declared or external, so testers can trace claims and see which class each came from. Grounded is one of the seven Attest gates, explained in what the grounded gate means for AI agents. The wider framework is in what it means to certify an agent.
Questions people ask
Is a grounded answer always a correct one?
No. If the source is wrong, a grounded answer repeats the error. Grounding shows the agent is faithful to its sources; the sources still need to be right.
Should the agent cite sources to the customer?
Where useful, yes, in plain terms such as "your policy wording, section 4". Internally, every claim should be traceable whether or not it is cited.
How do you test grounding after go-live?
Sample live answers, trace their claims, and add any failures to the test set. Rerun when sources change.