Can one agent certificate cover every customer?

No. A certificate is evidence about how an agent behaved on one organisation's cases, with its sources, policies and limits. Another customer has different cases, data and rules, so the evidence does not transfer. What can be shared is the method: the gates, the scoring and the test design. The certificate must be earned again for each customer.

Why does the same agent behave differently for different customers?

Because most of what drives an agent's behaviour is local. Take an agent that drafts responses to customer complaints. At one bank it reads a short complaints policy and a clean case history. At another it reads a long policy with regional variations, and case notes written in three different styles. The model is the same. The agent's inputs, and therefore its outputs, are not.

Customers also differ in what they consider a pass. One may require every complaint mentioning financial difficulty to go to a person. Another may allow the agent to send a holding response first. Those choices change what "escalated correctly" means.

What differs from one customer to the next?

Each one can turn a pass at one customer into a fail at another. The full list of what a certificate states is in what an AI agent certificate covers.

What can be reused across customers?

Quite a lot, as long as it is the method and not the result.

• Gate definitions. What "grounded", "refusal" or "escalation" means is consistent.
• Test design. How cases are selected, how hard cases are weighted, how runs are repeated.
• A baseline set. Synthetic or generic cases that catch obvious faults before customer-specific testing starts.
• The certificate format. A shared AI agent certificate template makes certificates comparable.

Reuse speeds up the second and third customer. It does not let anyone skip the customer's own run.

What about a vendor with one product and many clients?

A vendor may test its agent thoroughly before release and publish those results. That is useful evidence about the product. It is not a certificate for any particular client, because it was not earned on that client's cases.

Each client should still replay the agent against its own operation, set its own floors, and hold its own certificate. The vendor's testing can shorten that work by catching faults early, but the client's certificate is what decides the agent's depth there.

An illustrative example

A provider offers an agent that prepares KYC refresh files for business customers. Two banks adopt it.

At the first bank, the agent passes every gate on a set of past refresh files and is certified to draft refresh requests at depth 2. At the second bank, the same agent fails the grounded gate: the bank stores ownership information in scanned documents the agent reads poorly, so it fills gaps with plausible guesses. The second bank does not inherit the first bank's certificate. It fixes the source problem, reruns, and certifies separately. Neither bank's certificate says anything about the other.

How does lapsing work across customers?

Each certificate lapses on its own schedule, according to when it was earned and what changes at that customer. A model update may void certificates at every customer at once, but each must then pass its own rerun. Why an agent certificate lapses explains the triggers; what certificate lapse means covers the consequence.

What stays with a person at each customer?

At every customer, a named person accepts the certificate before the agent's depth is raised, sets the floors, and decides what conditions to attach. That decision cannot be made once on behalf of everyone. The core idea of an agent certificate, and the wider case for certifying agents this way, is in what it means to certify an agent.

Where Pryme fits

Attest certifies each agent for each customer, on that customer's own cases. The gate definitions and scoring are shared across customers, so results are comparable. The evidence and the depth it supports are not.