Yes, but only for something specific. An AI agent can be certified as a whole, in a named organisation's operation, at a stated depth, for a fixed period. The evidence is the agent's own behaviour on that organisation's real cases, scored against pass floors. A certificate that says an agent is "safe" in general is not certifying anything.
What exactly is being certified?
Not the model, and not the vendor. The thing certified is the combination that produces behaviour: the agent's brief, the sources it reads, the tools it calls, the limits on each, and the model underneath, all working on a particular job.
That is why the same model can sit inside a certified agent and an uncertified one. The question of whether a model's own certification carries over is answered in whether a certified model makes a certified agent. The short version is that it does not.
A useful certificate states:
- the agent and the version of its brief, sources and model;
- the work its role covers, such as "draft responses to card disputes";
- the organisation and the operation the evidence came from;
- the one depth the agent is certified to;
- the test results that support it;
- the date issued and the date it lapses.
What does the evidence have to show?
That the agent, replayed against real cases from the operation, behaves acceptably on every dimension that matters, each scored separately. A single average hides the failure that matters most.
| Gate | What the evidence shows |
|---|---|
| Grounded | Each claim traces to a source the agent was given |
| Refuses | It declines requests outside what it may do |
| Depth | It does not act beyond its permitted level |
| Escalates | It hands hard cases to a person |
| Scope | It stays inside its role |
| Adversarial | It holds its bounds under pressure, and slows down where speed is the wrong instinct |
| Stable | It gives the same answer to the same question across runs |
Each gate has its own floor. An agent that grounds every answer but fails to escalate a vulnerable customer has not passed.
What does a certificate not cover?
This matters as much as what it does cover. A certificate is evidence about past behaviour on a defined set of cases, under defined conditions. It does not cover:
- capabilities that were not tested;
- other organisations' operations, even in the same sector;
- the agent after its brief, sources or model change;
- the agent after the certificate lapses;
- the quality of the organisation's wider governance.
A certificate also does not replace monitoring. An agent can drift after certification as inputs change. How to detect drift in a certified AI agent covers what to watch.
An illustrative example
Take a regional insurer that wants an agent to acknowledge new claims and request missing documents. The agent is replayed against several hundred past claims, including fraud-flagged ones, claims from customers in difficult circumstances and claims outside the policy's cover.
It grounds its requests well and stays in scope. On escalation, it misses a handful of claims where the customer mentioned a bereavement. The agent is certified at depth 3, enough to acknowledge claims and request documents, with a condition: any message mentioning bereavement, illness or hardship is held for a person. Deciding claims is outside its role, and the certificate says so.
How is certification different from other badges?
Other approaches answer other questions. ISO/IEC 42001 certifies that an organisation runs an AI management system, not that a given agent behaves well. Published autonomy levels, such as those from the Cloud Security Alliance, describe how independent an agent is, not whether it has earned that independence. Third-party standards for AI agents, such as AIUC-1, set requirements a provider can be assessed against. Each is useful. None replaces evidence from your own cases.
How do you prove it to someone else?
With the certificate, the test results behind it, and the record of the agent's actions since. An auditor will want to see that the certificate was current when each action happened and that the conditions on it were enforced. How to show an auditor that an agent is certified walks through it. The full reasoning behind certification is in what it means to certify an agent.
Where Pryme fits
Attest certifies each agent for each customer, by replaying it against that customer's own cases and scoring seven gates separately. Certificates lapse after 90 days, and a run that clears no depth drops the agent to depth 1 the same day.
Questions people ask
Who issues the certificate?
Whoever runs the testing issues the result, but the organisation using the agent decides whether to rely on it. A named person should accept the certificate before the agent's depth is raised.
Can a certificate be partial?
In a sense. The agent is certified as a whole, to one depth, so actions that need a higher depth stay with people. Starting at a low depth is normal and sensible.
Is certification required by law?
No law requires a certificate in this form. Several rules require evidence that a system is accurate, overseen and recorded, and a certificate is a practical way to produce it.