Describe the agent in actions, not technology. Show risk and compliance the worst thing it could do, how that is prevented, which rules apply and the control for each, and evidence from your own cases. Propose a cautious start, usually drafting, and agree in advance what evidence would let it do more.
Why do risk and compliance hesitate?
Usually because they are asked to approve a capability described in general terms: "the agent will handle customer queries". That covers everything from reading a balance to waiving a fee to telling a vulnerable customer something wrong. Faced with a description that wide, the only safe answer is no.
They also know that once something is live, it is hard to pull back. An approval request that includes how authority will be reduced or withdrawn is far easier to agree than one that assumes it only grows.
What should you bring?
| Their question | What to bring |
|---|---|
| What can it actually do? | A capability list, each with the depth it needs |
| What is the worst action it could take? | The highest-impact action per capability, and whether it can be undone |
| How is that prevented? | Limits and hand-over rules, enforced outside the agent |
| Which rules apply? | An obligation map: each rule, the control, the evidence |
| Does it work on our cases? | Test results on real past cases, including refusal and escalation cases |
| How will we know if it goes wrong? | Monitoring plan and the indicators you will report |
| Who answers for it? | Named owner, approvers and accountable executive |
| How is authority withdrawn? | Stop switch, automatic drop on failed tests, recertification cycle |
The second row is the one most often missing. Risk teams think in terms of impact and reversibility, so give them that directly. "The agent can issue refunds up to 100, which can be reversed by the customer team within a day" is something they can assess. A general description of AI risk management is not.
What are the practical steps?
- Meet early. Share the capability list before building. Ask what would make them comfortable.
- Write the risk in actions. For each capability, the worst plausible action, its impact and its reversibility.
- Map obligations to controls with compliance, one line per rule.
- Test on your own cases, including cases where the right answer is to refuse or hand over. Share the results, failures included.
- Propose a cautious start. Drafting at depth 2 for anything customer-facing is the easiest first approval.
- Agree widening criteria in advance. What evidence, over what period, would justify moving a capability to depth 3. Write it into the approval.
- Agree reporting. What risk will see each month: volumes, holds, refusals, overrides, errors.
Security has its own review in parallel; see how to get IT security to approve an AI agent.
What does this look like in practice?
Take a consumer credit firm proposing an agent for its collections team. The first request, "an agent to manage arrears contact", is declined.
The second request is specific. The agent will read accounts in early arrears, draft reminder messages for collections officers to send, and propose repayment plans within the firm's existing policy for an officer to agree. It will not contact anyone flagged as vulnerable, in dispute or in a complaint; those go to specialists. The worst action is a wrongly worded message, and every message is reviewed before sending. Testing on four hundred past cases shows the agent correctly routes every vulnerable and disputed case.
Risk approves at depth 2, with a written condition: after three months, if officers' edit and rejection rates on reminder messages stay low and no routing errors occur, reminder messages for one arrears stage may move to depth 3. Compliance adds a condition that contact-hour rules are enforced at the send action.
Which controls make approval easier?
- The depth each action needs, so approval can be partial.
- Limits enforced outside the agent.
- An append-only record of every decision.
- Automatic reduction of authority on failed tests.
- Staff trained to review drafts critically; reviewers who wave things through undermine the case. How to bring staff along is in getting staff to accept AI agents.
Where Pryme fits
In Pryme, each action needs a depth and the agent's depth is earned through Attest on your own cases, certificates lapse after 90 days, and a lapse or a run that clears no depth drops the agent to depth 1 the same day. Every verdict names the bound that decided and is filed. That gives risk and compliance the reduction and evidence routes they look for. More in deploying an AI agent: train, shadow, certify, widen.
Questions people ask
Should risk and compliance sit on the project team?
They should be consulted throughout, not embedded as builders. Their independence is part of what makes their approval worth having.
What if compliance says a rule prevents the use entirely?
Ask which rule and which part of the use. Often the rule prevents one capability, and the rest can go ahead.
How long does approval usually take?
Much less when risk and compliance have seen the plan from the start. Late involvement is the main cause of delay.