PrymeIntelligence

Agent certification·Article

The process owner should sign off certification, not the vendor

The process owner should sign off an AI agent's certification, not the vendor: the person who answers for the process must accept the evidence.

The person who answers for a process should be the one who accepts the evidence that an agent may work in it. A vendor can run the tests, supply the tooling and explain the results, but it does not answer to your customers, auditors or regulators for the outcome. Sign-off on certification belongs with your process owner.

Why should the vendor not sign off?

There are three separate reasons, and any one of them would be enough.

The first is interest. A vendor benefits when its agent is certified and deployed widely. That does not make vendors dishonest, but it makes their sign-off the kind of evidence an auditor discounts, in the same way a firm does not let a supplier approve its own invoices.

The second is knowledge. To certify an agent is to judge it against your own work. Whether a certification set represents the work, whether the labels are right and whether the floors are high enough are questions about your process, your customers and your policies. The vendor knows its product. Your process owner knows what a wrong answer costs on a Tuesday morning when the queue is full.

The third is accountability. When something goes wrong, the questions land on the firm, and usually on a named person within it. The vendor's sign-off cannot transfer that accountability; it can only obscure who held it. The wider argument is in a vendor's certification is not yours.

Who counts as the process owner?

The person who would be asked to explain a failure in the process if a person, not an agent, had caused it. In a claims operation that is usually the head of claims; in accounts payable, the financial controller; in complaints, the head of complaints. It should be a named individual, not a committee or a team mailbox.

In the UK, the Senior Managers and Certification Regime already assigns named individual accountability for many functions in financial services, and the natural owner of an agent's process is often the person who holds that responsibility. Elsewhere the principle holds without the regime: someone must own the decision to let the agent act. The mechanics of naming that person are covered in how to assign an owner to every AI agent.

What is the process owner actually signing?

Not a technical report on the model. Sign-off is an acceptance of specific claims, each of which the owner is in a position to judge:

  1. The certification set represents the work the agent will do, including the awkward cases.
  2. The labels reflect the correct outcomes under current rules.
  3. The floor for each gate is set at a level the business is willing to accept.
  4. The results meet those floors.
  5. The depth being granted is justified by those results and no higher.

If the owner cannot affirm one of these, the sign-off should not happen. That is the control working, not a delay to be managed.

What should the vendor still provide?

The vendor has a real role. It should make the owner's decision easier and better informed, without making the decision.

The vendor provides The process owner decides
The testing tools and the run itself Which cases go into the set
Results by gate, with failures explained Whether each result is acceptable
Known limitations of the model and agent Which limitations are tolerable for this process
Its own testing on its own cases, as context What depth the agent may hold here
Technical fixes when a gate fails Whether a fix is enough to rerun

This split mirrors how firms already handle outsourced services. The provider supplies evidence and expertise; the firm keeps the decision to rely on it. It is also why certification is per customer, not per vendor: the evidence has to be produced on each customer's cases and accepted by each customer's owner.

Can an agent sign off anything itself?

No, and the reasons are worth stating plainly because the question comes up. An agent cannot accept its own certification for the same reason it cannot sign off its own work elsewhere. The discussion of whether an AI agent can sign off a set of accounts and whether an AI agent can sign a contract reaches the same conclusion: sign-off is where accountability attaches, and accountability needs a person.

The same holds for the vendor's own agents that might assist in certification, such as an agent that summarises results. They can prepare the material. The signature is a person's.

What happens when the owner will not sign?

Then the agent does not get the depth, and the reasons are recorded. The most common reasons are a set that misses an important case type, a floor the owner considers too low, or results that pass overall but fail badly on a slice the owner cares about.

Each of these is useful information. A refusal to sign often leads to a better set, a narrower scope or tighter bounds, any of which makes the eventual certificate more trustworthy. Firms should be wary of any process that makes refusal hard, such as a launch date set before the run.

How does sign-off fit with expiry?

Certificates lapse, so sign-off recurs. Each recertification needs the owner to accept the new evidence, and a change of process owner should itself prompt a review, because the new owner has not accepted anything yet. A sign-off from someone who has left the firm is not a sign-off anyone can be asked about.

Where Pryme fits

In Pryme, Attest runs certification on the customer's own cases and presents results gate by gate, but the certificate is issued only when the customer's named process owner accepts it. The acceptance, the set it covered and the depth it granted are filed in the append-only record, and lapse after 90 days.

Questions people ask

Can the process owner delegate sign-off?

They can delegate the review work, such as checking labels or reading results. The acceptance itself should stay with them, because it is their accountability being exercised.

What if the process owner lacks technical knowledge?

They do not need it for the claims they are signing. Results should be presented in terms of cases, outcomes and gates, which an experienced process owner can judge without understanding the model.

Should risk or compliance also sign?

For high-depth capabilities, a second-line review before sign-off is sensible. It adds a challenge to the evidence, but it does not replace the owner's acceptance.

Decide what your agents may do. Prove it. Enforce it.

Pryme Intelligence certifies each agent on your own cases, checks every action against five bounds at the moment it happens, and files every verdict.