Banks should define AI literacy by line of defence. Front-line staff must read an agent's work and release held actions correctly. Risk and compliance must be able to challenge how an agent earned its authority. Internal audit must be able to reconstruct any decision from the record. Each level needs different skills, and each must be evidenced.
What does the first line need?
The first line works beside agents every day: in payments operations, lending, onboarding, servicing and collections. Literacy here is practical:
- Depth. Each capability has its own depth: 1 read, 2 draft, 3 routine work within bounds, 4 money or hard-to-undo actions, 5 coordinating other agents. Staff must know what the agent in front of them may do.
- Verdicts. Allowed, held for a named person, refused as outside bounds, or withheld at the platform floor. Releasing a held action makes the decision yours.
- Provenance. A balance sealed by the core banking ledger outranks a customer's declared income, and an external email cannot authorise a payment.
- Expiry. Agent certificates lapse after 90 days, and a run that clears no depth drops the agent to depth 1 the same day.
What does the second line need?
Risk and compliance need to challenge, which requires understanding how authority is set. Depth is earned, not configured: the role sets a ceiling, certification evidence on the bank's own cases sets the depth, and the lower wins. Certification replays the agent against real cases and scores every answer against a floor, across separate gates such as grounded, refuses, escalates and stable.
A literate risk officer can read that evidence and ask the right questions. Were the cases representative? Did the agent refuse what it should have refused? Why does it hold depth 4 for one capability and depth 2 for another? Model risk is supervised in most banking markets, and this is where agent oversight meets the existing model risk framework. Risk literacy is the core skill: connecting what an agent can do to likelihood, impact and control.
What does internal audit need?
Audit needs to reconstruct. Every verdict is filed in an append-only record, and the minimum chain records which of the five bounds decided: role scope, certificate, sources, the bank's own limits or platform guardrails. A literate auditor can pick any payment or decision, trace it to the facts used and their provenance, the bound that allowed or held it, and the person who released it, then check that person's credential on that date.
How does literacy differ across the bank?
| Group | Literacy means | Evidence |
|---|---|---|
| Operations and approvers | Reading verdicts, releasing held actions correctly | A current approver credential |
| Risk and compliance | Challenging certification evidence and depth | Training record and challenge log |
| Internal audit | Reconstructing decisions from the record | Audit working papers |
| Senior management | Knowing what is delegated, to which agent, until when | Briefing record and test result |
| All other staff | Recognising agents and their limits | A literacy badge |
What literacy do banks not need?
Not everyone needs prompt literacy. Prompting is a builder's skill. Approvers who spend their time crafting better prompts are doing the wrong job; theirs is to judge the output against the evidence.
How do banks evidence it?
Keep training records per person, per capability, with results and expiry, and tie approval rights to a current credential. The record of agent verdicts names who approved what, and credentials that can be checked at source let a supervisor confirm standing without asking for files. Compare how universities and utilities approach the same question with lighter supervision. The full model is in AI literacy and certification for people who work with agents.
Where the Academy fits
Pryme Intelligence Academy maps onto the lines. Foundations gives everyone the core concepts. Operator certifies approvers through an exam and a two-week practicum. Builder certifies those who design agents, with a capstone agent certified in Attest to depth 2 or higher. Certificates are issued through VerifyDoc.ai and can be checked at source.
Questions people ask
Should internal auditors be certified on agents?
They need enough literacy to reconstruct decisions and challenge controls. Certification as an approver is not required unless they also release actions.
Does literacy change when an agent moves to depth 4?
Yes. Approvers for depth 4 capabilities handle money and hard-to-undo actions, so their training and testing should reflect those cases.
How does this fit with existing model risk training?
It extends it. Model risk training covers validation and use; agent literacy adds depth, verdicts, expiry and the record of each action.