If your firm uses AI to translate customer letters, the FCA's new good practice report makes clear that you remain responsible for the result, and it counts staff review of AI-drafted letters as good practice. A qualified linguist who reads both languages, with a review record you can show a supervisor, is the most direct way to meet that expectation.

Would anyone on your team notice the error?

Picture an arrears letter going out to a customer whose first language is Polish. To save time, it passed through the AI translation feature in your customer communications platform. The English version told the customer to contact you within 14 days to avoid further charges, and the translation dropped a single word, so it told them they did not need to contact you. Nobody in your collections team reads Polish, so the first sign of the problem is a complaint.

On 19 October 2026 the FCA published AI in financial services: good practice and areas for improvement. It states that "firms can outsource AI services, but remain responsible for them", and it lists "assuming AI outputs are clear without testing them" as an area for improvement. The report does not discuss translation, yet every principle in it applies to a customer letter that an AI tool has translated.

AI is now part of everyday operations in your sector. In the Bank of England and FCA's 2024 survey, 75% of firms were using AI, 33% of use cases came from third parties, and 46% of firms said they had only a partial understanding of the AI they use. In a speech launching the new report, Jessica Rusu, the FCA's chief data, information and intelligence officer, said the regulators' next survey shows 86% of firms now use AI, and that firms need "clear accountability, risk-based controls, human oversight, escalation triggers".

Why translated communications carry a hidden Consumer Duty risk

Under the Consumer Duty, PRIN 2A.5.3 requires your communications to be "likely to be understood by retail customers" and to equip them to make decisions that are "effective, timely and properly informed". The rule covers every communication in any form, so a translated letter is held to the same standard as the English original.

‘The customers most likely to receive a translated letter are often the ones the Consumer Duty is looking to protect. If a firm cannot check what its translated letter says, it cannot show the customer understood it and good outcomes are being achieved.’ said Charles Roe, former Director of Mortgages, UK Finance, and former Head of Mortgages and Mutuals, FCA.

The 2021 Census found that 5.1 million people in England and Wales, 8.9% of residents aged three and over, had a main language other than English, and 880,000 of them could not speak English well. The FCA's guidance on the fair treatment of vulnerable customers names low capability, such as poor literacy, as a driver of vulnerability and asks firms to make sure "all communications and information about products and services are understandable".

The errors that cause harm are usually small. In a peer-reviewed study in JAMA Internal Medicine, researchers checked machine translations of 647 sentences from hospital discharge instructions: 92% of the Spanish and 81% of the Chinese sentences were accurate, yet 2% of the Spanish and 8% of the Chinese sentences had the potential for significant harm. A fluent sentence with one reversed meaning reads as correct to anyone who cannot check it against the source, which is why a monolingual reviewer cannot catch it.

What good practice looks like

The FCA's report describes firms using AI to draft customer letters where "staff reviewed all the changes before approving the final wording", and another where "the firm's case officers then reviewed and sent the letters". For a translated letter, that reviewer has to read both languages, which in practice means a qualified linguist. ISO 18587, the international standard for post-editing machine translation, sets requirements for "full, human post-editing" and for the competences of the people who do it.

The report also addresses dependence on technology suppliers. It notes that firms "may not be able to control changes made by large technology providers", and describes good practice firms that used contract terms to obtain notice of significant model changes. When the engine behind your translation tool is updated, output that was acceptable last month can change with it, so your review process needs to catch the change before your customers do.

Finally, the report asks firms to test whether AI output is clear. For a translated communication, that means testing it with people who speak the language, in the same way you would test the English version.

How we verify AI translation at Guildhawk

Guildhawk is a language intelligence company for regulated industries, and our approach rests on the principle the FCA has now set out: the firm that uses AI remains accountable for what it produces. Every AI output in Guildhawk’s in-house secure translation tool, GAI Translate, can be verified by an accredited human linguist, and each job exports a review record with a full audit trail that can be shared with your compliance team or the regulator, if required.

Before a new AI model reaches client work, our technology team runs it through a structured validation programme. It combines several automatic quality measures with error-based evaluation drawing on ISO 5060, the international guidance on evaluating translation output. When early results led us to upgrade the underlying model, we ran the validation again on the upgraded version. Live work produced with AI is still reviewed by a human linguist.

Our linguists carry out full post-editing to the standard of a publishable human translation. Their written guidelines tell them to read the source before the machine output, to check every negation such as "do" and "do not", to look for anything added or omitted, and to treat suggested terminology as unconfirmed until it has been checked against the client's termbases and reference material. Reviewers report recurring errors so that the next output improves. Behind this sit 25 years serving regulated industries, ISO 27001 certification, (which we were the first language service provider to achieve), and the Queen's Award for Enterprise.

A checklist for AI-translated customer communications

Use these six questions to test your own process, whoever translates for you.

• Do you know every point where AI translates a customer communication, including features built into third-party platforms?
• Who is accountable for the accuracy of each type of translated communication, and does your senior management responsibilities map show it?
• Does a qualified linguist who reads both languages review the translated versions of collections, complaints, claims and vulnerability communications before they are sent?
• Have you tested translated communications with customers who speak the language, as you would test the English versions?
• Does your provider's contract give you notice of changes to the AI model behind your translations?
• Can you produce a record of who reviewed each translation, what they changed and when?

If your firm is sending AI-translated letters to customers today, the FCA has set out where responsibility lies for their accuracy and who is accountable.

Talk to us now about putting a qualified linguist and an auditable review record behind everyone.