AI in financial communications: lessons from the FCA

To use AI for customer communication, firms need to put in place the right skills and methods. These include good content reviews, customer testing, and clear instructions for AI tools and agents. A solid content policy should guide this work. The FCA’s new publication on AI in financial services gives food for thought on how to use AI to improve customer understanding.

On 19 October 2026, the Financial Conduct Authority published AI in financial services: good practice and areas for improvement. The publication gives examples of how firms use AI and apply existing expectations, including Consumer Duty. It introduces no new requirements.

For content teams, these findings suggest several practical steps. The recommendations below focus on the everyday work of writing and reviewing customer communications.

AI can help make financial content clearer

FCA case study E describes a firm using AI to improve replies to customer queries and complaints. Staff reviewed the letters before sending them. The firm also tested the communications with a small group of consumers before introducing the tool. Consumers reported clearer letters and faster responses, including when their complaints were rejected.

The useful point here is the connection between writing and testing. Before changing a communication, decide what customers need to understand and do. Then check whether the change helps them.

For example, a reply about a declined insurance claim should explain the reason and the customer’s options. That purpose can guide the writing instructions and the review.

Give AI a clear brief and reliable sources

The FCA identifies the use of approved source material as good practice. For writing teams, this means agreeing which information an AI tool can use and checking whether that information is suitable.

An instruction such as “make this simpler” leaves several decisions open. The team still needs to define the reader and the purpose. It also needs to decide which details must stay and which terms need an explanation.

For example, an AI tool might rewrite a letter about loan repayments. The instructions should identify the source and explain which figures the tool must preserve. They should also state how to present the amount due and the payment date.

Rules about voice and tone belong in that brief too. A request to sound friendly gives limited direction. Examples of suitable wording can help make the instruction more precise.

If an AI agent can draft and send messages, it also needs clear limits on when it may act. The team should decide when it must stop and refer a case to a person.

Give reviewers a clear task

A ‘review’ can become a meaningless catch-all if the role isn’t clearly defined and if there are not clear review criteria.

Putting a reviewer into the workflow requires decisions about what is within their scope to review, and how much time and experience is needed to do the review. Reviewers need access to the source material. They also need authority to correct a response or stop it from being sent.

A review should cover meaning as well as wording. For example, a shorter explanation may leave out a condition that changes the customer’s options. A polite reply may acknowledge a complaint without answering it.

Test whether customers understand

The FCA criticises firms that assume AI outputs are clear without testing whether different customer groups understand them.

Testing should start with the customer’s task. For a repayment letter, this could mean checking whether someone can identify the amount due and explain what happens next.

Ask people to explain the message in their own words. Ask them to show what they would do next. These tasks can reveal misunderstandings that a general question about clarity might miss.

Include customers with different communication and accessibility needs. Keep a record of misunderstandings and use the findings to revise the content.

Check whether everyone can use the whole service

The checks need to cover the whole service. This includes how customers find information, complete a task, and get help.

It also needs to include all customers. For example, testing should include customers who use screen readers or need more time to respond. It should also check whether people can move to another support channel.

If an AI agent acts for a customer, the communication should explain its actions.

Use a content policy to connect the work

A content policy can bring these decisions together. It can explain which uses of AI are suitable for your communications and who is responsible for each check.

The policy should also give teams a practical way to handle uncertainty. For example, it could require staff to check the source before answering questions about an exclusion. It could also set rules for recording errors and reviewing changes.

A policy workshop can help teams work through real examples and agree what to do. The result should be guidance people can understand and follow during their normal work.

How Narratology can help

Your team may already use AI to draft letters, simplify documents or review content. Narratology can help you agree how to do that work and how to judge the results.

• Agree your AI content policy. AI policy and risk workshops help your team decide where AI can help and which checks are needed.
• Give AI better writing instructions. Turn your content standards and tone guidelines into practical prompts, with approved examples for AI tools and agents.
• Find problems in customer communications. Plain-language reviews and customer testing help you see what people understand and where they need more support.
• Build your team’s skills. AI-assisted writing training and coaching help writers and reviewers judge the content and apply consistent standards.

Contact Narratology to discuss your AI content policy, writing process or customer testing needs.

Source

Financial Conduct Authority. AI in financial services: good practice and areas for improvement. Published 19 October 2026; updated 20 October 2026. Sections 1.2, 2.2 and 3.4.