Can banks use AI agents before staff are trained?

It depends on what the agent does. A bank can run agents at depth 1 and 2, reading policies and drafting case summaries and complaint responses while staff train. Anything that moves money, changes an account or decides on a customer must wait until the people overseeing it are trained. In banking, that waiting line is also a supervisory expectation.

Why are banks stricter than most?

Banks operate under named accountability, model risk supervision and conduct rules. In the UK, for example, the Senior Managers and Certification Regime ties outcomes to named individuals, and PRA SS1/23 sets model risk management principles; in the US, the interagency guidance SR 11-7 plays a similar role. Supervisors in most markets expect a bank to show that the people overseeing automated decisions understand them.

An untrained reviewer approving agent actions is not oversight. A supervisor will ask how the reviewer was qualified, and "they learned on the job" is a weak answer.

Where can agents start safely?

1. Internal knowledge. Answering staff questions from the current policy library, with citations (depth 1 and 2). Staff check answers against the policy, which is itself a form of training.
2. Complaint drafts. Summarising the case history and drafting a response for a complaints handler (depth 2).
3. Case preparation. Assembling the file for a credit, KYC or fraud case, with the provenance of each fact (depth 1 and 2).
4. Customer correspondence drafts. Letters and messages a person reviews and sends (depth 2).
5. Reconciliation breaks. Listing and explaining breaks for operations staff (depth 1 and 2).

What must wait for trained reviewers?

• Depth 3 routine actions on accounts, such as fee refunds or limit changes.
• Depth 4 actions: payments, transfers, anything hard to undo. These are held for a named approver even after training.
• Credit decisions about individuals, which the EU AI Act lists as high-risk.
• Account closures, exits and anything touching vulnerable customers.

Some actions never reach any agent. Reversing a settled payment on the bank's own rail, for example, can be withheld at the platform floor for every agent at every depth.

What does a sensible start look like?

Take a regional bank's complaints team of 40. The bank starts an agent drafting complaint responses at depth 2. Handlers review every draft. In the first weeks, the team lead notices two handlers approving drafts almost unchanged, including one that offered redress below the bank's own policy. The lead pulls those handlers into a short calibration session and reinforces the point that the draft is a proposal.

Meanwhile, every handler completes foundational training, and the ten who handle the most complex complaints work towards an operator certification. The bank's model risk team reviews the agent's certification results on the bank's own past complaints. Only then does the bank consider depth 3 for a narrow class: small fee refunds below a set value. Banks make predictable mistakes in this sequence, covered in what banks get wrong about AI agent training, and a planning tool is the AI agent training plan template for banks.

Which controls keep early use within appetite?

• The bank's own limits cap every agent at depth 2 until a named senior manager approves more, on evidence.
• Every verdict is filed in an append-only record, so a supervisor can see who approved what.
• Sampling of approved drafts, watching for rubber-stamping.
• Certificates on agents that lapse after 90 days, and a failed run that drops the agent to depth 1 the same day.
• A record of who on the staff is trained, to what level, and when that lapses.

The same reasoning applies outside banking; see whether utilities can use agents before staff are trained and whether wealth managers can. In payments, the gap between agents and trained reviewers is the binding constraint, as argued in payments will not scale AI agents without trained reviewers. Banks wanting a certified operator standard can look to Pryme Intelligence Academy's Operator track. The topic guide on training teams for AI agents covers sequencing in full.