The FCA finds useful AI in clearer letters and better handovers

Customers at one financial firm said its revised letters were clearer and arrived sooner, even when their complaints had been rejected. The firm had used a language model to suggest better wording, retained staff review, and tested the communication with a small group of customers before deployment.[1]

The Financial Conduct Authority published the example on 19 October in its account of good and poor AI practice. Drawn from regulatory work and industry engagement, the examples introduce no new obligations.[1]

The letter is a revealing place to begin. A customer can understand an unwelcome answer, see the reason for it and decide what to do next. That is a different achievement from persuading the customer to like the decision. It gives the firm a useful way to judge an AI writing tool: does the person now understand the substance?

Understanding is already part of the rulebook

The FCA’s Consumer Duty guidance requires communications to meet customers’ information needs and equip them to make informed, timely decisions. Firms should consider the intended audience, product complexity and communication channel, and test, monitor and adapt their communications where appropriate. The guidance says information about switching or complaining should be at least as clear as information used to sell a product.[3]

That creates a substantial job for language tools. Dense correspondence can be reorganised around the decision, its reasons and the next available action. But shortening a sentence is only part of that job. The rewrite must preserve a condition, exception or deadline that changes what the customer can do. A polished paragraph that quietly drops one of those details has simplified the wrong thing.

Another firm’s controls reduced unexpected model behaviour but left responses so generic that some customers were steered towards unsuitable options. Consistency and usefulness need to be assessed together.[1]

A good handover carries the context

In her published speech accompanying the findings, FCA executive Jessica Rusu described a firm maintaining a live list of signals that its chatbot could flag to human advisers when a customer might need extra support. She also highlighted the need for a clear route to speak to a person and for accountable ownership of AI systems.[2]

The regulator’s standing guidance on vulnerable customers gives that route a practical purpose. It calls for staff who can recognise and respond to customer needs, systems that record and retrieve those needs, and flexible service and communication channels. Monitoring should establish whether the support is actually meeting those needs.[4]

Consider a hypothetical customer who explains that they cannot read a document in its current format. A useful AI response might establish the format they need and pass that request to the right team. A subsequent adviser who can see the request starts with the problem already understood. If the customer has to repeat the entire conversation, the transfer has preserved the queue while losing the help.

The new examples recognise that continuity: some firms let consumers change channels without repeating information. Smaller firms also used selected outputs and contrasting test cases to check their systems.[1]

Measure the task the customer came to complete

For a team buying an AI assistant, these examples suggest concrete comparisons. Can recipients explain a letter’s meaning? Can someone with an additional support need reach a suitable channel? Does the next staff member receive enough context to continue? A shorter response time becomes more valuable when it accompanies progress on those tasks.

A clearer letter and a better handover are modest outcomes individually. They also occur at the points where customers experience an institution most directly. The FCA’s examples make those everyday encounters a useful starting point for deciding where financial AI earns its place.

Report an error or suggest a correction