Banks face a two-sided challenge as artificial intelligence (AI) agents move from answering questions to taking action. On one side is the need to control the legal risks when customers delegate financial decisions; on the other is the need to contain the costs of deploying agents themselves.
A WilmerHale client alert and a Reuters report on McKinsey’s Global Banking Annual Review describe pressures that could reshape banks’ controls and economics. Customer agents may test established concepts of consent and accountability. Internally, agents’ interactions can generate computing bills that undermine expected savings.
The immediate compliance issue is determining whose instructions a bank is executing. WilmerHale’s analysis, focused on the U.K. Financial Conduct Authority framework, distinguishes authenticating an agent from establishing its authority. Software may legitimately access an account yet make a transfer or investment outside the limits its customer intended.
Under the U.K. Payment Services Regulations, payment authorization depends on payer consent through a procedure agreed with the provider. Strong customer authentication is a separate requirement. Establishing a connection between customer and agent does not necessarily settle whether a particular transaction was authorized.
An instruction to manage cash efficiently, for example, could leave uncertainty over permitted amounts, counterparties or products. Banks need to consider how authority is established, restricted and withdrawn, and when customers must approve further action.
Those concerns help explain the industry’s caution. Reuters reported that most banks currently block external agents from authorizing transactions within their systems, limiting their impact on deposits and profits.
Demand nevertheless appears substantial. McKinsey surveyed 46,000 retail banking customers worldwide. Among AI users, 72% said they would allow a banking agent to perform one financial task, although they would cap autonomous approvals at $500. The consultancy expects widespread customer use of agents within three years.
Delegation also complicates financial crime monitoring. WilmerHale describes an agent choosing payment methods, currencies or intermediaries to complete a supplier payment quickly and cheaply. Individually understandable actions could collectively produce unusual activity. An agent might also respond to a blocked transaction by seeking another route.
Existing anti-money-laundering and sanctions obligations still apply. The operational challenge is distinguishing legitimate automated behavior from fraud or evasion, and deciding when repeated activity requires human scrutiny.
Evidence becomes especially important when a customer disputes a transaction. Ordinary records may show what the bank received and executed without establishing whether the agent followed its instructions. Relevant material could include permissions, customer communications and action histories. WilmerHale cautions that banks should consider what evidence they need to accept agent-initiated transactions, even where responsibility for creating those records lies elsewhere.
Banks face a different control problem with their own agents in the form of unpredictable consumption of tokens, the units underlying AI processing charges. Reuters reported McKinsey’s warning that token spending could easily reach 20% of banks’ total labor costs.
Some banks have assigned spending oversight to teams called AI FinOps. Earlier approaches directed employees to use different models according to query complexity. Agent-to-agent communication makes that approach less effective because usage accumulates across interactions.
As a result, banks are becoming more selective about deployment, emphasizing applications with stronger expected returns. Alongside measurable productivity savings, they are examining whether additional revenue justifies higher AI spending, Reuters reported.
The two pressures can reinforce each other. Customer agents could move deposits from accounts paying no interest into higher-yielding alternatives, squeezing margins while banks absorb the costs of their own automation.
For management, the combined implication is that agent adoption requires both financial discipline and clear boundaries on delegated action. The U.K. Financial Conduct Authority currently relies on existing regulatory frameworks rather than proposing AI-specific rules, according to WilmerHale. Banks therefore face the practical work of adapting authorization, monitoring and recordkeeping now, while testing whether internal agents deliver enough value to justify their expense.
