Adding a person to an AI workflow does not automatically create effective oversight. A reviewer may have too little time, insufficient information or no authority to challenge the system. If the model produces a persuasive explanation that does not match the evidence, a nominal approval step can simply transfer the error into the bank's official record.

The FCA's October 2026 AI findings emphasise proportionate governance, accountability and continuing monitoring within existing rules. EU applicability must be assessed separately under the AI Act framework, including the bank's role, intended use and relevant application provisions. A compliance support tool is not automatically subject to the same classification as a creditworthiness system. The controls below are implementation proposals, not a claim that one legal category covers every use.

Define the permitted task narrowly

Write down what the tool may do and what decision remains with an authorised person. Summarising a case file, suggesting questions and deciding whether an account should be restricted are different tasks with different consequences. Define prohibited uses as clearly as permitted ones, especially where a tool can send messages, modify records or trigger downstream actions.

Record the model, version, retrieval sources and data-access permissions. Include AI functions embedded inside vendor products and staff productivity tools. An inventory based only on projects labelled “AI” will miss some of the most consequential uses.

Give the reviewer enough evidence

A reviewer needs access to the underlying material, a clear indication of missing information and an explanation of the task performed. For a case summary, require references to the relevant records and test whether those references genuinely support the statements made. A fluent answer without traceable support should not acquire authority merely because it appears in a familiar interface.

Separate fact extraction from assessment. A model may correctly extract a transaction date and still draw an unsupported conclusion about intent. The interface should make that distinction visible. Where the task requires judgement, the reviewer should record their own reason for accepting or rejecting the output rather than select a generic approval box.

Test oversight under realistic pressure

Use a controlled evaluation set containing ordinary cases, ambiguous records, contradictory evidence and intentionally incomplete files. Include attempts within retrieved text to redirect the system away from its approved task. Check whether the tool follows untrusted instructions or discloses information beyond the user's authority.

Measure substantive errors separately from style issues. A concise summary that omits the decisive evidence can be more harmful than a verbose one. Track factual support, omissions, inappropriate recommendations and unauthorised actions. Also measure reviewer behaviour: whether reviewers spot the errors, how long it takes and which cases they escalate.

Examine a concrete failure

In a hypothetical pilot, an AI assistant summarises a customer complaint and says the customer accepted a proposed resolution. The source record only says that the proposal was sent. The reviewer is processing a large queue and approves the summary without opening the correspondence.

The remedy is not just a warning to be careful. The bank can require explicit source evidence for acceptance, prevent the tool from closing the case and route ambiguous responses to a human queue. The evaluation should then test whether those changes reduce the specific error without making the workflow unusable.

Control changes and fallback

Vendor updates, retrieval changes and prompt edits can alter behaviour even when the interface looks the same. Define which changes require renewed evaluation, who approves them and how the previous version can be restored. Preserve enough evidence to reconstruct important outputs, subject to the applicable privacy and retention rules.

Provide a workable fallback for outages or a suspended model. Staff should know which manual process to use and which activities must pause. Avoid a fallback that assumes the same unavailable provider will supply the missing information.

Judge the resulting decisions

A useful governance report connects the tool to outcomes: error types, affected cases, correction time, human overrides and residual limitations. Compare performance across relevant case types, including less common or more complex populations. Efficiency gains should be considered alongside whether the bank can still explain and correct its decisions. Effective oversight is demonstrated by the ability to detect, challenge and stop a harmful output in the actual workflow.