No. A model's certification or evaluation tells you about the model in general conditions. An agent is that model plus a brief, sources, tools, limits and an escalation route, working on a particular job. Those additions decide most of what the agent actually does. The agent needs its own evidence, earned on the cases it will handle.
Why does the model's evidence not carry over?
Think of a certified engine in a car nobody has tested. The engine may be excellent. Whether the car brakes, steers and stays on the road depends on everything around it.
An agent is similar. The model supplies language ability and reasoning. The rest of the agent decides what it reads, what it is allowed to do and when it stops. Two agents on the same model can behave very differently:
- one reads a clean policy library and cites it; the other reads stale documents and blends them;
- one has a tool that can only draft a refund; the other has a tool that can issue one;
- one is told to hand off any mention of hardship; the other is not.
None of those differences appear in a model evaluation.
What does a model certification or evaluation actually show?
Useful things, at a general level. It may show how the model performs on public benchmarks, how it responds to known harmful prompts, or that its provider runs a management system to a recognised standard such as ISO/IEC 42001. It may come with documentation on intended use and known limits.
What it cannot show is how the model behaves inside your agent, on your cases, with your tools. That is the gap.
| Question | Answered by model evidence? | Answered by agent evidence? |
|---|---|---|
| Can the model follow instructions in general? | Yes | Yes, for this brief |
| Does it ground answers in your sources? | No | Yes |
| Does it refuse actions outside this agent's role? | No | Yes |
| Does it escalate your hard cases to a person? | No | Yes |
| Does it stay within its permitted depth for each tool? | No | Yes |
| Does its behaviour hold steady across runs on your cases? | No | Yes |
Where do agent-specific failures come from?
Mostly from the parts the model does not control. An agent fails the grounded gate because a source is missing or contradictory. It fails refusal because its brief is vague about what is out of scope. It fails depth because a tool was given broader permissions than its job needed. It fails escalation because the brief never defined what a hard case looks like.
A better model can reduce some of these. It cannot fix a missing source or a tool with the wrong limit. That is why agent certification, explained in whether an AI agent can be certified, tests the whole agent.
An illustrative example
Take a wealth manager building two agents on the same well-evaluated model. One answers clients' questions about their portfolio statements. The other prepares suitability review packs for advisers.
The statement agent passes every gate on the firm's past client queries. The suitability agent fails grounding: it reads client fact-finds that are often out of date, and it fills gaps with plausible assumptions about risk appetite. Same model, same evaluation, opposite results. The firm certifies the statement agent and keeps the suitability agent at read-only until the fact-find source is fixed and the agent is told to flag gaps rather than fill them.
What about when the model is updated?
A model update changes the engine, and the agent's evidence was earned with the old one. Even if the new model scores better on general benchmarks, the agent should be rerun on its own cases before its certificate is relied on again. Behaviour can also shift gradually without any update, which is why detecting drift in a certified AI agent matters after certification.
How should you use model evidence, then?
As one input when choosing a model, and as background for your own testing. Read it to understand known weaknesses, so you can include cases that probe them. Keep it in your file as part of the agent's documentation. Do not cite it as the reason the agent may act.
When an auditor asks why an agent is permitted to do something, the answer should point to the agent's own certificate and the results behind it; see how to show an auditor that an agent is certified. The full reasoning is in what it means to certify an agent.
Where Pryme fits
Attest certifies the agent, not the model: it replays the whole agent, with its brief, sources, tools and limits, against the customer's own cases and scores seven gates. A model update means a fresh run before the certificate counts again.
Questions people ask
Should we choose the model with the best evaluation scores?
Use evaluations to shortlist, then test candidate models inside your agent on your cases. The best general model is not always the best for a narrow job.
Does a certified model reduce how much agent testing we need?
It may mean fewer basic failures, so testing goes faster. It does not reduce what needs to be tested, because the agent-level questions remain.
Can we switch models without recertifying?
No. Switching the model changes the agent. Rerun on the same case set and compare; if results hold, recertification can be quick.