AI agents in banking: what changes for products, processes and APIs

AI agents carry out tasks on their own: they research, prepare decisions and call other systems to do so. For banks this is less a question of technology than of accountability, control and interfaces.

Products

Customers will hand tasks to agents, such as checking invoices, comparing offers or preparing payments. Banks have to decide whether to offer their own agents, give third-party agents access, or both. For payments, this reopens questions of strong customer authentication and liability, because today’s rules assume that a person approves every payment themselves.

Processes

Internally, agents already support analysis and documentation, for example when evaluating requirements, when creating API specifications and checking them against in-house standards, or in testing. The benefit only materialises once it is clear who reviews and approves the results. An agent does not replace accountability.

APIs

Agents use systems through interfaces. The Model Context Protocol (MCP) is an open standard for this: an MCP server exposes the functions of existing APIs as tools that an agent can call. Well-described APIs with clean specifications are the foundation. This makes API governance a prerequisite: which APIs may agents use, with which permissions, and how is their access monitored?

Anyone introducing AI agents should therefore start not with the model, but with product decisions, responsibilities and interfaces.