Recertification is running an AI agent through certification again, on a fresh set of the customer's own cases, to renew or restore its certified depth. It happens before a certificate lapses, after something the certificate depends on changes, and after a failed run. The new evidence alone decides the agent's depth.
Why does recertification matter?
A certificate is only as current as the evidence behind it. The agent was proven on cases from the business as it was then. Since then, products may have changed, the model provider may have updated its model, and the agent's brief may have been rewritten. Recertification is how the evidence catches up.
It also prevents credit for past performance. An agent that held depth 3 for a year gets no allowance for it when recertified. If the new run supports only depth 2, depth 2 is what it gets. That keeps authority tied to what the agent can do now, which is the principle of earned depth.
What triggers recertification?
- Schedule: certificates lapse after 90 days, so a run is booked well before.
- Change: a new or revised source, a rewritten brief, a new tool connection, a model update, or a request for more depth.
- Failure: a run that clears no depth drops the agent to depth 1 the same day; after the cause is fixed, recertification restores depth.
- Drift: a drift signal on live work brings a run forward.
Recertification covers the agent as a whole. If a tariff document changes, the agent's cases are re-run, including those that rely on it, and the new certificate replaces the old one.
How does a recertification run work?
The certification set is refreshed with cases filed since the last run, so the agent meets the business as it is now. A portion of earlier cases is kept, so behaviour can be compared with the baseline. Floors are carried over unless a recorded decision changes them before the run. The agent is scored gate by gate, and a new certificate is issued at the highest depth it clears, with a new expiry.
What does that look like in an example?
An energy supplier changes its tariffs. The billing agent's tariff document is replaced, which affects two capabilities: answering tariff questions and drafting tariff switches. The agent is recertified on cases filed since the change, plus earlier cases for comparison. Tariff questions clear at depth 3. Tariff switches clear at depth 2 again, but drift flags that the agent now recommends the cheapest tariff more often without checking usage. The team adjusts the brief, reruns the agent's cases, and it clears. The agent's other capabilities were re-run alongside them.
How is recertification different from qualification?
Qualification is the whole process of fitting an agent to a customer's work: brief, sources, bounds and first certification. Recertification repeats the proof once the agent is already in service. Qualification asks whether the agent can do this job here. Recertification asks whether it still can. Both should be done with the same independence, a point taken up in who certifies AI agents?
Where Pryme fits
Attest runs recertification the same way as the first certification: replaying the agent against the customer's own cases, gate by gate, against floors. Changes to an agent's brief or sources are recorded, so the capabilities that depend on them can be identified and the agent recertified.
Questions people ask
Does recertification use the same cases as last time?
Mostly new ones, filed since the previous run, with some earlier cases kept for comparison. Reusing the same set invites an agent tuned to pass it.
Can the agent keep working during recertification?
Yes, under its current certificate, until that certificate lapses. After a run that clears no depth it works at depth 1 until recertified. The full method is in what it means to certify an agent.