Banks tend to treat AI agent training as a model risk or compliance exercise. Validators study the agent, everyone else clicks through an annual module, and maker-checker is assumed to cover the rest. The people who approve the agent's held payments and account changes get the least preparation, yet their decisions are what the supervisor will examine.
What are the most common mistakes?
| Mistake | Why it happens | What to do instead |
|---|---|---|
| Leaving it to model risk | Agents look like models, so they go to the model team | Validate the agent there; train the operators separately |
| Annual e-learning for all | It fits the mandatory training calendar | Short shared module, then assessed practice for reviewers |
| Trusting maker-checker | The control already exists | Retrain checkers for an agent as maker |
| Training front office only | Customer-facing roles get attention | Start with payments, fraud and complaints operations |
| Training once at go-live | Projects end at launch | Recertify on a cycle and when the agent changes |
| Not explaining withheld actions | Staff assume every action is configurable | Teach the platform floor and why it exists |
Three of these do most of the damage, so they are worth looking at closely.
Why isn't model risk management enough?
Banks have mature model risk frameworks. In the US, SR 11-7 sets the interagency expectations; in the UK, the PRA's SS1/23 sets model risk management principles for banks; most banking markets supervise model risk in some form. It is natural to route agents through the same process.
That process answers a real question: is this agent sound for its purpose? It does not answer the operational one: can the people who approve its held actions tell a sound draft from an unsound one, today, on this queue? Validators review the agent; they do not sit in the payments operations team at four in the afternoon when a release is held because the payee was added through an external channel. Treat validation and operator competence as two controls, owned by two teams, and do not let one be counted as the other.
Why doesn't maker-checker cover it?
Maker-checker works when the maker is a person who can be asked why they did something, and when the checker knows the maker's habits. When an agent is the maker, both conditions change.
The checker now needs to read a verdict, not a colleague's note. They need to know which of the five bounds held the action: the agent's role scope, its certificate, a source, the bank's own limit or a platform guardrail. They need to weigh provenance, because a beneficiary account confirmed on the bank's own ledger is sealed while one quoted in a customer's email is external and cannot authorise a payment. And they need to know that an agent whose certificate lapsed after 90 days, or failed a run and dropped to depth 1 the same day, is sending them routine work it handled alone yesterday.
A checker trained only for human makers will approve the agent's tidy drafts faster than a colleague's messy ones. That is the opposite of what the control needs.
Why does withheld cause so much friction?
Some actions are withheld at the platform floor for every agent at every depth: reversing a settled payment on the platform's own rail, for example. The action exists and works; it is refused on purpose, and no bank setting can switch it on.
Staff who have not been taught this treat it as a fault. They raise tickets asking for it to be configured, or they look for workarounds, such as asking the agent to draft a "correction" that amounts to the same thing. Training should cover what is withheld, why, and what the person does instead: the reversal goes through the bank's own controlled process with a named approver. Customers narrow what an agent may do through their own limits, which produce refusals or holds; only the platform floor withholds. People who understand that distinction stop fighting the system and use it.
Who should a bank train first?
The operations teams that will meet held actions in volume: payments operations, fraud operations, complaints and account servicing. They come before the relationship managers and before the wider staff, because their decisions are the ones filed against the agent's actions.
Then the named owners. Under the UK's Senior Managers and Certification Regime, named individuals answer for their areas, and the Consumer Duty expects good outcomes for retail customers. An owner who cannot explain what depth their agent works at, or who approves its held actions, will struggle to show the controls work. In EU banks, Article 4 of the AI Act adds a duty to ensure a sufficient level of AI literacy among staff operating AI systems. This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
What should a bank do instead?
Separate the two controls: model validation for the agent, assessed competence for the people. Replace the annual click-through with a short shared module on the vocabulary followed by assessed practice on real held cases for each reviewer. Retrain checkers for an agent as maker. Recertify people on a cycle and whenever an agent gains a capability or depth. A bank can adopt the practice of tying approval rights to a current certificate, depth by depth.
For the detail, see how banks should train staff before deploying AI agents, what an AI agent training programme for banks should include, and the AI agent training plan template for banks. The general order is in the guide to training teams for AI agents. Wealth businesses inside banking groups have their own version of these mistakes, covered in what wealth managers get wrong about AI agent training.
Where the Academy fits
Pryme Intelligence Academy separates the vocabulary from the skill, as a bank should. Foundations gives everyone the shared terms. Operator certifies reviewers and team leads through a written exam and two weeks of real review decisions with no missed red line, signed off by a Pryme assessor. Certificates are issued through VerifyDoc.ai, carry an expiry date and can be checked at source.
Questions people ask
Should the second line of defence be trained as reviewers?
Risk and compliance staff should be trained to test the controls and reconstruct decisions from the record, not to clear the queue. Keep them independent of the approvals they will later examine.
How should a bank train staff in branches?
Branch staff need the vocabulary so they can explain to a customer why a request is waiting or was refused. If branch roles approve held actions, they need reviewer training and assessment like operations staff.
Does training need to differ between retail and corporate banking?
The concepts are the same, but the cases differ. Corporate payments carry larger values, more complex mandates and different fraud patterns. Assess reviewers on cases from the book they will actually handle.