Does switching model providers change what an AI agent may do?

It should not change the rules, but it does change the evidence. An agent's role scope, your limits and the platform floor stay exactly where they were. What changes is the proof that the agent performs within them. Until it has been retested on your own cases with the new model, it should drop to a lower depth.

Why don't the rules move?

Because they were never in the model. If an agent may refund up to 200 and must never change a customer's bank details, those rules sit in the agent's bounds, checked at the moment of action. A new model proposes actions; the same checks decide them.

The same is true of the actions no agent may take at all. Those are decided once, for every agent, and a provider change does not reopen them; see how to decide which actions no AI agent may take.

If your rules do live in the model, in a long system prompt the model is trusted to follow, a provider switch is dangerous. The new model may read the same words differently, and you will not know until something goes wrong.

Why does the evidence reset?

An agent's depth is earned by showing that its answers on your cases clear a floor. That result belonged to the old model. The new model may be stronger overall and still weaker at the one thing your agent does: reading a particular invoice format, recognising a vulnerable customer, or refusing a request outside scope.

So treat a provider switch like any significant change to the agent. Retest each capability and let the agent regain its depth on the results.

What happens step by step?

1. Change owner: raises the change. Through your normal change control for AI agents, naming the agent, the old and new provider, and the reason.
2. Platform: lowers depth for the switch. Capabilities above depth 2 drop to drafting until retested. People pick up the difference.
3. Process owner: reruns the tests. The same cases, the same floor, on the new model.
4. Process owner: restores the agent's depth to what the rerun clears. It may come back higher or lower.
5. Inventory owner: updates the register. The provider and model version are part of the agent's record. An agent inventory tells you what exists, not what it may do, but it must at least be accurate about what exists.

A worked example

Take a payments company whose dispute-handling agent drafts responses to card chargebacks at depth 2 and categorises incoming disputes at depth 3. The company switches providers to reduce cost.

On switch day, categorisation drops to depth 2, so analysts confirm each category. The rerun shows the new model categorises as well as the old on most dispute types but confuses two similar reason codes. Categorisation returns to depth 3 for every reason code except those two, which stay as drafts while the brief is improved. Drafting responses clears at depth 2 with no change. The refund limit, the scope and the ban on contacting cardholders directly never moved.

Does the provider itself carry risks you must manage?

Yes. For regulated financial firms in the EU, the model provider behind an agent is part of the ICT supply chain, and relying heavily on one provider raises concentration questions. Our article on DORA concentration risk and the model providers behind agents covers what supervisors may ask. Having a tested second provider is one answer, but only if the agent has been certified on it before you need it.

Which controls make the answer hold?

• Rules outside the model, so a switch cannot change them.
• Depth tied to the model version, so evidence does not carry over silently.
• A rehearsed switch. Our guide on switching model providers without starting again covers how to keep briefs and test sets portable.
• The model version in every verdict, so later reviews can tell which model acted.
• Provider changes on the vendor side detected. If a vendor updates the model under you, treat it as a switch.

The topic guide, agent governance is a bound, not a model, explains why the bound, not the model, is what you govern.

Where Pryme fits

Pryme sits between models and the business, so role scope, your limits and the platform guardrails do not change when the model does. Attest certifies each agent on your own cases, and the certificate belongs to the agent as it runs, so a model change means a new run before depth is restored.