AI Translation and Multilingual Customer Service for UK Businesses: What Works, What to Check and What It Costs (2027)

AI translation is good enough for routine customer service in most major languages: an enquiry in Polish answered in Polish from your systems, a tenancy letter drafted in Romanian from the English original, a phone call handled in Spanish by a voice agent, a candidate chased on WhatsApp in Urdu. It is not good enough on its own for anything where a mistranslation has legal, medical or safety consequences, where a qualified human translator or a bilingual member of staff reviews. The design is the same as any AI service: the assistant is grounded in your own answers and records, it says it is an AI, it escalates to a person when unsure or when the matter is consequential, and a sample of conversations is checked weekly by someone who speaks the language. At UK fixed-price rates, adding languages to an existing assistant is small; a multilingual enquiry assistant on web, WhatsApp and phone is a custom system at £8,000 to £25,000, with translation adding little to the per-conversation running cost.

Reliable for: Routine service in major languages
Human review for: Legal, medical, safety-critical
Cost: Small on top of an existing assistant

Where it works

Written enquiries by email, web chat and WhatsApp in most major languages, answered in the customer’s language from your own systems with the same grounding and escalation as an English conversation. Standard documents, letters, guides, FAQs and forms, drafted in another language from the English original for a bilingual person to check. Phone, through a voice agent that detects the language and continues in it, for booking, status and routine questions. Internal use: staff reading a supplier’s email or a customer’s complaint in a language they do not speak. Sectors where this pays most are lettings, care recruitment, hospitality, clinics and any business whose customers or workforce are multilingual.

Where it does not

Contracts, tenancy agreements, medical instructions, safety notices, consent forms and anything a regulator or a court might read: here a mistranslation is a liability, and a qualified translator or a bilingual professional reviews before use. Dialects and less common languages, where quality drops and the assistant should say so and escalate. Tone in complaints and sensitive matters, where a fluent person judges better than a model. The rule is the same as for accuracy generally: an agreed threshold, measured on real samples by someone who speaks the language, decides what the assistant handles unattended.

The checks

Five checks that decide what the assistant handles unattended in each language.

01
A native or fluent speaker scores fifty real conversations per language before go-live against the threshold, and a sample weekly after
02
The assistant states it is an AI in the customer’s language and offers a person
03
Consequential matters (money, complaints, health, legal, safety) escalate to a person regardless of language
04
Standard documents in other languages are reviewed by a bilingual person before they are issued, and the reviewed versions are reused rather than regenerated
05
Language detection is confirmed with the customer rather than assumed

The data rules

Translation is processing, and conversations in any language are personal data with the same rules: business plan or API under a data processing agreement, retention set, a DPIA where the volume or the data warrants it. Where the content is health or legal, the region of processing may matter and the residency guide applies. Nothing about another language changes the position; it just makes it easier to forget that a Polish complaint about a landlord is as sensitive as an English one.

What it costs

Adding languages to an assistant you already run is mostly the checking: a bilingual reviewer’s time for the fifty samples and the weekly sample, and a small amount of prompt work; the model handles the language at little extra token cost. A new multilingual enquiry assistant on web, WhatsApp and phone, integrated with your systems, is a custom system at £8,000 to £25,000 fixed, with the running cost per conversation stated before go-live and dominated by messaging and telephony rather than translation. Human translation for the consequential documents is bought separately, once per document, and reused.

Common questions

Can AI handle customer service in other languages?

For routine enquiries in most major languages, yes, grounded in your systems with escalation to a person. Not unattended for legal, medical or safety-critical content, or for less common languages where quality drops.

Is AI translation accurate enough for business use?

For routine service and internal reading, generally. Measure it: a fluent speaker scores fifty real samples per language against a threshold before go-live and a sample weekly after.

Can an AI phone agent speak other languages?

Yes. A voice agent can detect the language and continue in it for booking, status and routine questions, confirming the language with the caller and escalating to a person when the matter is consequential.

Should we translate contracts and medical documents with AI?

Only as a first draft for a qualified translator or bilingual professional to review. A mistranslation there is a liability; the reviewed version is then reused.

What does multilingual AI customer service cost?

Little on top of an existing assistant beyond bilingual checking time; a new multilingual assistant across web, WhatsApp and phone is £8,000 to £25,000 fixed, with translation adding little to the per-conversation running cost.

Do GDPR rules apply to translated conversations?

Exactly as to English ones: processor agreement, retention, DPIA where warranted, and region of processing where the content is sensitive.

AI translation and multilingual customer service for a UK business: what works, what to check, and what it costs.