Approval discipline above all. At depth 4 an agent can move money or take actions that are hard to undo, so the people around it need to approve without rubber-stamping, run two-person rules properly, check the provenance of anything that justifies a payment, know which actions are withheld, and respond fast when something goes wrong.
What makes depth 4 different?
Depth 3 covers routine work that can usually be corrected: a wrong booking can be rebooked, a mismatched entry re-matched. Depth 4 covers the actions where correction is expensive or impossible: releasing a payment, paying a claim, closing an account, deleting data, disbursing a loan. Mistakes here leave the organisation.
Depth 4 must be earned on the organisation's own cases through certification, capped by the role's ceiling, and certificates still lapse after 90 days. Even certified, a depth 4 agent usually acts alone only within tight limits, and anything above those limits is held for people. The skills below are mostly about what those people do.
Which skills matter most?
- Approving with attention. At depth 4, held items are high-value or unusual by definition. Approvers must read the evidence, not the summary, and write down what they checked.
- Running two-person rules. For actions that cannot be undone, two named people approve, independently. That means the second approver checks the case, not the first approver's signature. It also means neither approver is the agent that prepared the action.
- Checking provenance. Every fact is stamped sealed, signed, declared or external. An external claim, such as an emailed invoice or a bank letter, cannot authorise a payment. Approvers must spot when a payment rests on one.
- Knowing withheld actions. Some actions, such as reversing a settled payment on the organisation's own rail, are withheld at the platform floor for every agent at every depth. People must know these and use the formal process instead.
- Responding to failure. If a depth 4 action goes wrong, speed matters: recall requests, customer contact, incident logging. Teams must know the route without looking it up.
- Reading the minimum chain. Every verdict names the deciding bound among role scope, certificate, sources, the organisation's own limits and the platform guardrails. "Held: sources" on a payment means the evidence is too weak, not that the limit is too low.
What does the agent do, and what stays with people?
| The agent at depth 4 | People |
|---|---|
| Executes certified money movements within limits | Set the limits and review them |
| Prepares larger or unusual actions with evidence | Approve them, with two people where irreversible |
| Refuses actions resting on external claims | Verify through independent channels |
| Is blocked from withheld actions | Use formal processes for those outcomes |
Depth 4 builds on depth 3 supervision skills, covered in what skills people need at depth 3. The design side, including how to set limits and scenario suites for depth 4, is in what skills an AI agent builder needs.
What does good depth 4 practice look like?
Take a lender's operations team with an agent certified at depth 4 to disburse loans to verified accounts up to 20,000. Above that, two officers approve. A 45,000 business loan is ready for release. The agent prepares it with the verified account, the signed loan agreement and the approval record.
The first officer approves after checking the agreement. The second officer, rather than glancing at the first approval, opens the account verification and sees it was completed three months earlier, before the borrower's company changed directors. She asks for re-verification. The account is still correct, and the release goes out the next morning. The delay was a day; the check was real. That is what the second approval is for, as argued in two people should approve what cannot be undone.
Which controls support people at depth 4?
Tight limits with clear ownership. Two-person rules for irreversible actions, with independence between approvers. Provenance visible on every held item. A practised recovery route. Regular review of approvals to check they are real. Teams can record who holds which depth 4 skills in the skills matrix template. Organisations sometimes tie depth 4 approval rights to a current operator certificate, such as one from Pryme Intelligence Academy. The wider context is in the skills that matter when agents act alone and the role map in AI agent skills by role.
Questions people ask
Should every depth 4 action need two people?
No. Within certified limits, the agent can act alone; that is what depth 4 means. Two-person rules apply above limits and to irreversible actions the organisation designates.
Can the second approver be junior to the first?
Yes, if they are trained and independent. What matters is that they check the case rather than defer to the first approver.
What if approvers become a bottleneck?
Look at why so much is held. Often the limits or the sources need adjusting, through the risk owner, rather than approvers moving faster.