It can, and a well-designed programme makes sure it does. Certification rests on evidence about how the agent behaves now. If a scheduled run fails a gate, or its model, sources or brief change, that evidence no longer describes the agent in production. The certificate should stop counting the same day, and the agent's authority should fall with it.
Why should certification be that fragile?
Because agents change faster than organisations do. A management system certificate can reasonably last a year because the processes it covers move slowly. An agent's behaviour can shift between one day and the next when its model provider ships an update, when a policy document it reads is replaced, or when someone adjusts its instructions.
A certificate that survived those changes would be describing an agent that no longer exists. The difference between agent certification and a management system certificate is drawn out in whether AI agent certification is the same as ISO certification.
What can end a certificate overnight?
| Trigger | Why it voids the evidence |
|---|---|
| A scheduled run fails any gate | The agent no longer meets the floor it was certified against |
| The underlying model is updated | Behaviour may change even with the same brief |
| A source is added, removed or replaced | The agent now knows different things |
| The brief or instructions are edited | The agent has been asked to behave differently |
| A tool or its limits change | The agent can now do different things |
| The certificate reaches its lapse date | Evidence is too old to rely on |
| Monitoring shows sustained drift | Live behaviour has moved away from tested behaviour |
Not every change has to void the certificate permanently. A model update that passes a fresh run on the same case set can be recertified quickly. But until it passes, the old certificate should not be relied on.
What happens to the agent when it loses certification?
It drops to read-only. It can still look things up and help people find information, but it stops drafting for others to send and stops acting on its own. Work it was completing goes back to the team that owns the process.
That is a deliberate cost. The alternative, an agent continuing to act on evidence that no longer applies, is the failure certification exists to prevent. Teams should plan for it: know who picks up the work, and how quickly.
What has to happen before it is restored?
- Find out why it failed. A failed run should produce a report that names the gate, the cases and the outputs; a failed certification run report template gives a structure.
- Fix the cause. Narrow the scope, correct the brief, replace the source, or wait for a corrected model.
- Rerun on the full case set, not just the cases that failed. A fix for one gate can break another.
- Have a named person review the results. Not the person who made the fix. Who should review an AI agent's certification results covers who that should be.
- Restore depth one capability at a time, if several were affected.
An illustrative example
Take a lender's agent certified to draft arrears letters at depth 2 and send routine payment reminders at depth 3. On a Tuesday night, the model provider releases an update. The scheduled run the next morning shows the reminders are still accurate, but the agent now omits the line about free debt advice in a share of arrears letters, which fails the adversarial gate: a customer in arrears is a state where the shorter letter is the wrong one.
By mid-morning the agent is at depth 1. Collections staff send reminders by hand for two days while the brief is tightened and the run repeated. The reminders capability passes and is restored first. The arrears letters pass on the second attempt and are restored after the review. The record shows each step.
What evidence should the episode leave?
The failed run, the report, the fix, the rerun, the reviewer's sign-off and the dates each capability was restored. Kept together, these form part of the certification evidence pack, and they are often more reassuring to an auditor than an unbroken record. They show the control works.
The floor that decides a pass is itself an important choice, explained in what a certification floor means for AI agents. The broader idea is in what it means to certify an agent.
Where Pryme fits
In Pryme, a failed Attest run drops the agent to depth 1 the same day, and every certificate lapses after 90 days whether or not anything changed. Restoring depth needs a fresh pass on the customer's own cases.
Questions people ask
Can a certificate be suspended rather than lost?
In practice, losing depth until a rerun passes works like a suspension. What matters is that the agent stops relying on the old evidence immediately.
Does every small change really need a rerun?
Any change to the model, sources, brief or tools needs at least a rerun on the existing case set. A small change that passes is cheap to confirm; a small change that breaks something is expensive to discover later.
Who gets told when a certificate is lost?
The process owner, the team that will pick up the work, and whoever oversees agent risk. The notice should say which capabilities are affected and what the agent can still do.