Tracks for each role, modules built on banking's own risks, a practicum on every queue where an agent holds actions, sign-off by a named senior manager, and renewal that keeps pace with the agents. A bank also needs internal trainers, because it has too many reviewers across branches, operations and contact centres to certify any other way.
What tracks should it have?
Start from who decides what. A bank's programme usually needs six tracks:
- Board and executive committee. The vocabulary, what is withheld at the platform floor, how depth is earned, and the questions to ask about evidence.
- First-line reviewers. Payments operations, KYC analysts, credit support, complaints handlers, branch staff who approve held account changes.
- Team leads and operations managers. Calibration, cover, drift, and what to do when an agent's certificate lapses and it drops to depth 1.
- Builders. The people who design agents, write their briefs and read certification results.
- Second line and model risk. Testing controls from the filed verdicts, and judging whether an agent's authority rests on sound evidence.
- Internal audit. Sampling the record and checking that each approver was certified at the time.
What bank-specific modules should it cover?
| Module | What reviewers must be able to do |
|---|---|
| Payments and provenance | Decline any held payment supported only by an external instruction; apply payee name verification results |
| Credit decisions | State specific reasons from the agent's draft; recognise when a case must stay with an underwriter |
| KYC and financial crime | Separate signed registry records from declared customer statements; know which alerts the agent may close and which stay with an analyst |
| Complaints and customer outcomes | Judge whether a draft response gives a good outcome, not just a correct one |
| Incidents and resilience | Know the sequence when an agent's run clears no depth, an outage hits, or a red line is missed |
Credit deserves its own module. In the EU, creditworthiness assessment and credit scoring of natural persons are high-risk under the AI Act; in the US, ECOA and Regulation B require specific reasons when credit is denied. Reviewers on credit queues need to be fluent in giving reasons, not just in approving.
How should practice on real decisions work?
A practicum on each queue where an agent holds actions, run during the agent's shadow period. A payments analyst's practicum is on held repairs and returns; a KYC analyst's on held periodic reviews. Skills transfer only partly between queues, so a reviewer moving from payments to KYC should complete a short practicum on the new queue before approving there.
Calibration runs across sites. A bank with several operations centres should calibrate them on the same cases, because consistent outcomes for customers are a conduct expectation as well as a quality one.
Who signs off?
A named senior manager for each area, on evidence. In the UK, the Senior Managers and Certification Regime makes that person accountable, so they should see practicum results, calibration records and the register of certificates before approving any capability at depth 3 or 4. Model risk is supervised in most banking markets, so second line should confirm that each agent's depth rests on certification evidence. Audit then tests that the register matched reality on the day decisions were made.
How do you renew at scale?
With internal trainers and a staggered calendar. Certifying thousands of reviewers in one quarter means renewing them all in the same quarter a year later. Spread cohorts across the year, and train team leads to run calibration and support practicums so that renewal does not depend on a central team. How to run a train-the-trainer programme for AI agents covers building that capacity.
Update modules should follow agent changes: whenever an agent gains a capability or steps up in depth, the reviewers on that queue take a short update before the change.
An example
Take a hypothetical retail bank with 400 first-line reviewers across payments, KYC and complaints. The board and executive committee take the vocabulary in month one. Forty team leads are certified first and trained as internal coaches. They then supervise practicums for reviewers in cohorts of 50, queue by queue, during each agent's shadow period. Second line traces 100 filed verdicts before the first depth 3 capability. Renewals are spread across twelve months. The AI agent training plan template for banks gives the layout.
Where the Academy fits
Pryme Intelligence Academy provides the tracks: Foundations for the board and wider staff, Operator for reviewers and team leads, Builder for designers, Architect for people leading regulated multi-agent deployments. Certificates can be checked at source, and a lapsed or revoked one shows as such, which suits audit sampling.
Questions people ask
How does a bank's programme compare with a utility's or a wealth manager's?
Same structure, different weight. Utilities centre on vulnerable customers, as in what a programme for utilities should include; wealth managers on suitability and mandates, as in the programme for wealth managers.
Should branch staff be in the programme?
Yes, if they approve held actions such as account changes. Otherwise they need the vocabulary.
What do banks most often leave out?
Renewal planning and internal trainers. More in what banks get wrong about AI agent training and the guide to training teams for AI agents.