Line by line, starting with the people who will approve held actions. A bank should give the board and executives the shared vocabulary, certify first-line reviewers before the agent drafts anything they rely on, train second-line risk to test the controls from the record, and prove all three before any agent reaches depth 4 and moves money.
Why does a bank need a different order?
Banks already organise control in three lines: the business that owns the risk, the risk and compliance function that challenges it, and internal audit that tests both. An agent touches all three. The first line approves its held actions. The second line sets and checks its limits. Audit will one day ask to reconstruct a decision it made.
A bank also carries named accountability. In the UK, the Senior Managers and Certification Regime puts a named individual behind each area, and that person must be able to explain what the agent may do. Model risk is supervised in most banking markets, so the people who validate models will expect to understand how an agent's authority is earned. Training is how each of these groups becomes able to do its part.
What order should training follow?
| Order | Group | What they learn | Before what |
|---|---|---|---|
| 1 | Board and executives | Depth, verdicts, what is withheld at the platform floor, the questions to ask | Approving the risk appetite for agents |
| 2 | First-line reviewers: payments operations, KYC analysts, complaints, credit support | Reading held actions, checking provenance, declining with reasons, red lines | The agent's shadow period |
| 3 | Team leads | Calibration, cover, drift, what to do when an agent's certificate lapses | Go-live |
| 4 | Builders | The agent's brief, sources and bounds, reading certification results | Certification on the bank's own cases |
| 5 | Second line and model risk | Testing controls from the filed verdicts; tracing a decision end to end | Any step to depth 3 or 4 |
| 6 | Internal audit | Sampling the record; checking that approvers were certified when they approved | The first audit cycle |
The sequence matters more than the speed. A bank that trains builders before reviewers ends up with agents designed by people who have never handled the held queue.
What must bank reviewers learn specifically?
Provenance comes first. A payment instruction that arrives by email is external information; it cannot authorise a payment however convincing it looks. A balance from the core banking ledger is sealed. A beneficial ownership record from a verified registry is signed. Reviewers who see the class of every fact catch most fraud attempts without needing to know the fraud pattern.
Next comes the difference between refused and withheld. Some actions, such as reversing a settled payment on the bank's own rail, are withheld at the platform floor for every agent at every depth. Others are refused because the bank's own limits say so. A reviewer who confuses the two will escalate the wrong things.
Finally, credit. In the EU, creditworthiness assessment and credit scoring of natural persons are listed as high-risk under the AI Act, and Article 14 expects effective human oversight. In the US, ECOA and Regulation B require specific reasons when credit is denied. Reviewers on credit queues must be able to state, from the agent's draft, why.
What should be proven before an agent moves money?
Depth 4 is where actions become hard to undo. Before any capability reaches it, ask for four things: the agent certified on the bank's own cases at that depth; reviewers' practicum results with no missed red line; a second-line test traced from the record; and a named senior manager who has signed the depth decision and can explain it. For the EU view of the same step, see the EU AI Act for banks deploying AI agents.
An example
Take a hypothetical mid-sized bank deploying a payments exceptions agent. Board and executives complete the vocabulary course in month one. Twenty-four payments analysts take reviewer certification during a three-week shadow while the agent drafts repair proposals for failed payments. Second line samples 50 filed verdicts and traces each to source. Only then does the bank let the agent repair low-value formatting failures at depth 3. Releasing repaired payments stays with analysts at depth 2 until the evidence supports more.
Utilities and wealth managers follow the same logic with shorter chains: compare how utilities should train staff and how wealth managers should train staff.
Where the Academy fits
Pryme Intelligence Academy maps onto the lines of defence: Foundations for board and executives, Operator for first-line reviewers and leads, Builder for agent designers, Architect for whoever leads regulated multi-agent deployments. Certificates can be checked at source by number or QR code, which helps when audit samples approvers.
Questions people ask
Does internal audit need certification?
Not necessarily a reviewer certificate, but auditors must be able to read the record. Many banks give auditors the vocabulary course and a guided trace of one decision.
What do banks most often get wrong?
Treating training as a launch event rather than a control with renewals. More in what banks get wrong about AI agent training.