Multilingual Customer Service on Autopilot: Serving Customers in Any Language

Quick answer

Multilingual customer service on autopilot is realistic for a small UK business today, with tools like Tidio's Lyro and Fin costing from roughly £25 to £100 a month plus a one off native speaker check of your policies. Automate language detection, translated policies and simple replies about hours, delivery, orders and returns, and keep complaints, refunds and anything regulated with a person who has a clear handover process.

A Polish customer emailed a client of mine at 11pm on a Tuesday asking whether a sofa would fit through a standard door. Nobody in the business spoke Polish, so the email sat in the inbox for two days before someone pasted it into Google Translate, answered in English, and lost the sale. That is the whole problem in one story, and it is far more common in Britain than most owners think.

Here is the short version before I go deep. Multilingual customer service on autopilot is now realistic for a small UK business, and the good tools cost less per month than a single agency phone call in Polish. But "autopilot" is doing a lot of work in that sentence. What you can automate well is language detection, first replies, order and booking questions, and translation of your own written material. What you cannot automate safely is judgement, complaints, and anything that touches refunds or regulated advice. Get that split right and the tools pay for themselves within a month. Get it wrong and you'll be apologising in a language you can't read.

Why This Matters More in Britain Than the Sales Pages Admit

The vendors love to talk about "going global". For most of the businesses I work with, the customers who need another language are already on the high street. The Office for National Statistics reported from Census 2021 that 8.9 percent of people in England and Wales have a main language other than English, which is over five million residents. Polish is the largest at 612,000 people, then Romanian at 472,000, Panjabi at 291,000 and Urdu at 270,000. Romanian speakers increased sevenfold between the two censuses.

That is not a "global expansion" opportunity. That is the plumber in Slough, the letting agent in Leicester and the online garden centre in Bradford. Roughly 4.1 million of those five million told the census they speak English well, so they'll cope with your website. But coping is not the same as buying with confidence, and when something goes wrong with an order, people fall back to the language they think in. If you trade anywhere near a large Polish, Romanian or South Asian community, you already have multilingual customers. You're just making them do the translating.

There is a widely quoted figure from CSA Research, a US firm, that 75 percent of consumers are more likely to buy from a website in their own language. I've seen it cited in the Invent guide to multilingual AI support and a dozen other vendor pieces. It is an American survey and I wouldn't stake a budget on the exact number holding here, but the direction is obviously right.

What Autopilot Actually Means Here

Let me be precise, because the phrase gets abused. There are four layers, and they are not equally automatable.

The first layer is detection. The tool works out what language the customer is writing in from the message itself, or from their browser settings. This is the easiest part and every serious tool does it well now.

The second is translation of your own content. Your FAQ, your delivery policy, your booking confirmation, your returns process. This is one-off work that you check once and reuse thousands of times. It is safe to automate because you control the source text.

The third is live conversational replies. An AI agent reads a customer's Romanian question, looks up the answer in your English knowledge base, and writes back in Romanian. This is where the newer tools have made an enormous leap, and it is also where the risk lives.

The fourth is handover. When the AI can't help, it passes the conversation to a human with a translated summary attached. This layer decides whether the whole thing works or humiliates you.

Autopilot, properly defined, means layers one and two fully automated, layer three automated for a defined list of question types, and layer four designed so carefully that the customer never notices the seam.

Translation Layer Versus Native Multilingual AI

This is the single most important technical distinction, and the sales pages bury it. There are two ways a tool can "support 100 languages".

The old way bolts a translation service onto an English chatbot. The customer's Polish goes into Google Translate, comes out as English, the bot answers in English, and the answer goes back through the translator. Every pass loses meaning. Idioms break and product names get translated when they shouldn't be. As the guide published by Fin, which has an obvious commercial interest but is right on this point, puts it, the strongest multilingual agents generate answers directly in the target language rather than running a translation layer between the reasoning and the customer.

The new way uses a large language model, which is the same family of technology behind ChatGPT, that reads and writes fluently in dozens of languages natively. You write your help content once in English. The AI reads a Romanian question, understands it directly, and composes a Romanian answer directly. No pipe in the middle.

Ask any vendor one question: "If I write my knowledge base in English only, does your AI answer in Polish natively, or does it translate?" If they hedge, it's a translation layer. Not useless, but a different product with a lower ceiling.

The Tools I'd Actually Consider

A note on money first. Most of the tools below bill in US dollars, so the pound figures in this article are approximate conversions and will move with the exchange rate. Where I know a UK price in sterling I've said so.

Fin by Intercom

Fin is the AI agent that made per resolution pricing a category. Intercom renamed itself Fin this year and Salesforce completed its acquisition in September 2026. Pricing hasn't changed yet, but I wouldn't sign a multi year deal without a break clause. The published Fin pricing page charges $0.99 per outcome, which is roughly 75p to 80p, and you're only charged once per conversation even if Fin takes several actions.

The multilingual side is the reason it's on this list. Fin handles around 45 languages in general release, detects the customer's language automatically, and answers from a single English knowledge base. For a business with Polish, Romanian and Urdu speaking customers, that alone can justify the cost.

The catch is the maths. Getmacha's detailed breakdown of Intercom Fin pricing notes a 50 outcome monthly minimum, so the floor is about £38 a month before you've resolved anything, and Intercom's own seat licences sit on top at roughly £23, £66 or £103 per person per month depending on tier. Featurebase's analysis of Fin found that real world resolution rates in Intercom's own case studies run between 42 and 50 percent. Plan on 40 percent. If you get 300 conversations a month, that's 120 resolutions, so about £95 in Fin fees plus your seats.

Best for: an established online shop or service business doing 300 or more support conversations a month across several languages, that already uses or is willing to move to Intercom, Zendesk, HubSpot or one of the other help desks Fin plugs into.

Tidio with Lyro

Tidio is the one I see most often in small UK businesses, mostly because it's cheap to start and the Shopify integration is good. Its AI agent is called Lyro. Tidio's own pricing page gives you your first 50 Lyro conversations free, then you buy a monthly quota. A review at YourGPT puts the Lyro add on at $32.50 a month on annual billing for 50 conversations, which is around £25, rising as your quota grows. The base Tidio plans are roughly £23 a month for Starter and £46 for Growth.

The multilingual story is more modest than Fin's. Tidio detects the visitor's browser language and can auto translate conversations in the shared inbox. Lyro will answer in multiple languages from your content, but in my testing the quality in Eastern European languages sits a notch below the native LLM tools. Fine for delivery times and opening hours. I'd not trust it with a nuanced complaint.

The real warning about Tidio is the billing structure, not the language support. Chatarmin's analysis of Tidio's pricing found the plan stack jumps from about £46 to about £590 a month with nothing in between, and Lyro conversations, human conversations and automated flows are three separate quotas. Small businesses regularly find their £46 plan costing £120 once the add ons are switched on. Budget for that from the start.

Best for: a small online shop or booking led business on Shopify or WordPress, doing under 500 conversations a month, that wants live chat, a basic AI agent and multilingual handling in one cheap tool.

DeepL for your written content

DeepL is not a chatbot. It is a translation engine, and for European languages it is the best one a small business can buy. If your problem is "I need my returns policy, product descriptions and email templates in Polish and Romanian", this is the tool, not an AI agent.

DeepL bills in dollars or euros depending on where you sign up, so treat these as approximate. According to the eesel guide to DeepL's plans, the free tier gives you 50,000 characters a month, which is about 8,000 words, with no card required. The Individual plan is about £7 a month on annual billing for 300,000 characters, and Team is around £22 per user per month with shared glossaries. That glossary feature is why I push clients to Team rather than Individual: you can lock your product names and brand terms so they're never translated into something odd.

One caution from the same guide: DeepL's quality lead is strongest in German, French, Spanish, Polish, Dutch and similar European languages. For Japanese, Korean and Chinese, several structured comparisons put it behind the general purpose AI models. If your customers are mostly South Asian language speakers, DeepL supports far fewer of those languages than you'd hope, and I'd test it on real customer messages before paying.

Best for: any business that needs its own written material translated well and consistently, and wants a human to review it once before it goes live.

Google Translate

For understanding an incoming message so you know roughly what the customer wants, it's fine and free. For replying, it's the option I'd warn you off. The output reads like a machine wrote it, because one did, and customers can tell. In the UK, where the ServiceNow Consumer Voice Report found 69 percent of consumers already say AI chatbots miss tone and emotional cues, a clunky machine translated reply confirms every suspicion they had.

What About the Big Enterprise Names

You'll come across Zendesk AI, Salesforce Agentforce, Freshdesk's Freddy and a host of voice platforms. For a business with under ten staff, skip them. Zendesk's AI sits behind plans that cost more per seat than most small businesses spend on support altogether, Freddy's translation is an add on tuned for major languages, and the voice platforms quoted in the Rework comparison start at contract values north of £100,000 a year. None of it is aimed at you.

The Compliance Part Nobody Puts on the Sales Page

I've watched owners set up a multilingual bot on a Friday afternoon and only learn the rules when a customer complains.

Two UK regimes apply at once. The first is data protection. The Information Commissioner's Office is the regulator, and its guidance on AI and data protection makes clear that a chatbot processing customer names, emails and support messages is fully inside UK GDPR. In practice for a small business that means four things. You need a documented lawful basis, which for support enquiries is usually legitimate interests or performance of a contract. Your privacy notice has to say that AI is used and how. The customer must be able to tell they're talking to a machine from the interface itself, not from a line in your privacy policy. And you should complete a Data Protection Impact Assessment before launch, because the ICO has said most new AI uses involving personal data count as high risk. A DPIA is a written check of what could go wrong and how you'll prevent it, and for a simple support bot it's a two hour job.

The bit that is specific to multilingual work is data transfer. Nearly every AI tool sends your customer's message to servers in the US or EU for processing. The EU is fine under the UK's adequacy regulations. The US is only fine where the provider is signed up to the UK extension of the Data Privacy Framework or you have appropriate contract terms in place. Check the provider's data processing agreement for the words "UK" and "international transfers" before you sign. If they only mention EU GDPR, ask.

The second regime is consumer law, and it's the one that has changed dramatically. The Digital Markets, Competition and Consumers Act 2024 gave the Competition and Markets Authority direct enforcement powers from 6 April 2025. As Pinsent Masons explains, the CMA can now fine a business up to £300,000 or 10 percent of global turnover, whichever is higher, for unfair commercial practices including misleading actions and omissions, without going to court first.

Why does that matter for translation? Because a misleading statement in Romanian is still a misleading statement. If your AI tells a customer in Polish that returns are free when your English policy says the customer pays, you have made a misleading claim to a consumer. "The bot translated it wrong" is not a defence the CMA has shown any interest in. The CMA has said it will focus first on the most egregious cases and objectively false information, so a one off error in a small shop is not going to trigger an investigation. But a systematic wrong answer repeated to hundreds of customers is exactly the kind of pattern the new regime is built to catch.

The practical upshot is simple. Everything your AI is allowed to say about price, delivery, returns, cancellation and guarantees should come from a written source you control, and you should have had a native speaker check the translated versions of those specific items once. That's a few hundred pounds of translator time to protect against a regulator with a 10 percent of turnover stick.

Welsh, Scotland and the Rest of the UK

The census figures above cover England and Wales, but the pattern of Polish, Romanian and South Asian languages leading the list holds across the UK. Welsh deserves a specific mention. The Welsh Language Standards apply to public bodies and some utilities, not to your café or garage, so you're not legally required to serve customers in Welsh. But the native LLM tools handle Welsh reasonably well now, and offering it costs nothing once your bot is set up. In parts of Gwynedd and Anglesey it's the difference between being local and being an outsider.

A Workflow That Works for a Business With Fewer Than Ten Staff

Here's the setup I recommend, in order, with rough costs. The order matters.

Week one, translate the fixed content. Open a free DeepL account. Paste in your FAQ, delivery policy, returns policy, booking terms and your three most used email templates. Translate into your top two customer languages, which for most UK businesses will be two of Polish, Romanian, Panjabi, Urdu or Arabic depending on where you are. Then pay a freelance native speaker to review just those documents. Through ProZ or a local agency, expect £60 to £120 per language for that volume. It's the best money in the whole project.

Week two, put those translations somewhere the AI can read. If you're using Tidio, that's the Lyro knowledge source. If you're using Fin, it's the help centre. Don't upload your entire website. Upload the checked policies and the FAQ, and nothing else yet.

Week three, turn on automated replies for a narrow list only. I use five categories to start: opening hours, delivery times, order status, how to book or buy, and how to return or cancel. That's it. For everything else, the AI is instructed to say, in the customer's language, that it will pass them to a person, and to collect the order number or booking reference so the human doesn't have to ask.

Week four, build the handover. This is the layer everyone skips. When a conversation escalates, the human sees three things: the original message in the customer's language, a machine translation into English, and a one line summary of what the AI already told them. The human replies in English, the tool translates the reply, and here's the rule I insist on: for anything involving money, the human reads the translated reply back through DeepL into English before it sends. Round trip checking catches about 90 percent of the howlers.

Month two, review every escalated conversation. You'll find two or three question types the AI could safely have handled. Add them one at a time, with the same native speaker check on any new policy wording.

Total cost for this in month one, for a business on Tidio: roughly £25 for Lyro, £23 for the Starter plan, nothing for DeepL free, and £120 to £240 for human review. Call it £250 to £300. Ongoing, £50 to £70 a month. For a business on Fin at 300 conversations, roughly £95 in resolutions plus seats, so £150 to £250 a month, with the same one off review cost.

Where Common Advice Is Wrong

The standard advice on this topic has holes in it.

The first is "just support every language". The tools make it trivially easy to tick 45 boxes, so people do. Don't. Every language you switch on is a language you can't check. Pick the two you can afford to have reviewed by a human, launch those, and expand only when you've seen a real customer in a third language more than a handful of times.

The second is "the AI will learn from your past conversations". If your history is three years of a stressed owner replying in a hurry, the AI learns to reply like a stressed owner. A checked FAQ beats ten thousand old emails.

The third is treating a chatbot as a substitute for a person rather than a filter in front of one. The UK evidence on this is brutal. A Capterra UK poll reported by Consultancy.uk found that 51 percent of UK consumers say chatbots never understand their request, and research from Cavell found 79 percent of UK adults prefer customer service from someone in their own country. That last figure matters here. A Polish speaking customer in Wolverhampton would rather deal with a UK business that handles Polish competently than be routed to a call centre abroad. AI plus a local human is the combination people want, and it's the one small businesses are best placed to offer.

An Honest Reality Check

There are businesses for which the right answer is not to do this at all. If you get fewer than a handful of non English enquiries a month, pay a freelancer on retainer to handle them by email within 24 hours and put a line on your website saying you welcome enquiries in Polish and Romanian. That's £30 to £50 a month, it's a human, and it's better than a bot for that volume.

If your service involves regulated advice, so anything financial, legal, medical or involving immigration, I would not let an AI reply to customers in a language you can't read. The Financial Conduct Authority's Consumer Duty applies to communications regardless of language, and the risk of a plausible but wrong answer in a language you can't audit is too high for the money you'd save.

One technical weakness is worth naming. Language models are far better at languages with lots of text on the internet than at those without. Polish and Romanian are well served. Panjabi, Sylheti, Somali and Tigrinya are patchy, and some tools claim to support them while producing text a native speaker would find odd or occasionally offensive. If your community speaks one of those, test with a real speaker first.

Questions to Settle Before You Buy

Do you actually know which languages your customers use? Pull three months of enquiries and count. Most owners guess wrong.

Can your team read the escalations? Someone needs to own the handover inbox, and they need a workflow for round trip checking replies about money.

Has a human checked the policy translations? If not, you're not ready to switch on automated replies, whatever the tool.

Does the provider's contract mention UK GDPR and international transfers by name? If it only references EU law, get it in writing.

Have you written the DPIA? Two hours, one document, and the ICO's own templates make it painless.

What happens when the tool is wrong? Decide now who apologises, in what language, and how you make it right. Write it down.

Six yeses and you're ready. Four, and you should start with DeepL and the freelancer retainer and come back to the chatbot in three months.

Something to Do This Week

Pick one thing. Open a free DeepL account, paste in your returns policy, translate it into the language you see most in your inbox, and send it to a native speaker to check. That single document, done properly, is the foundation of everything else in this article. Multilingual customer service on autopilot isn't really about the software. It's about having a small set of things you're certain are right, in every language you offer, and then letting the machine repeat them at 11pm on a Tuesday so you don't have to.

Sources

Office for National Statistics, Language, England and Wales: Census 2021: https://ons.gov.uk/peoplepopulationandcommunity/culturalidentity/language/bulletins/languageenglandandwales/census2021
Invent, Best AI Tools for Multilingual Customer Service (2026): https://www.useinvent.com/blog/multilingual-ai-best-practices-chatbots-virtual-assistants-and-ai-customer-service-2026-guide
Fin, Best Multilingual AI Agents for Customer Service in 2026: https://fin.ai/learn/best-multilingual-ai-agents-customer-service
Fin, Fin AI Agent Pricing: https://fin.ai/pricing
Getmacha, Intercom Fin Pricing: The Per-Resolution Cost Breakdown: https://www.getmacha.com/blog/intercom-fin-pricing
Getmacha, Intercom Pricing Explained (2026): Plans, Seats & Fin AI Costs: https://www.getmacha.com/blog/intercom-pricing-explained
Featurebase, How Much Does Intercom's Fin AI Cost? 2026 Pricing: https://www.featurebase.app/blog/fin-ai-pricing
Builts.ai, Intercom Fin AI Review: We Tested It on 500 Tickets (2026): https://builts.ai/blog/intercom-fin-ai-review/
Tidio, Tidio Pricing: https://www.tidio.com/pricing/
YourGPT, Tidio Review 2026: Pricing, Lyro Limits & Who It's Actually For: https://yourgpt.ai/blog/comparison/tidio-review
Chatarmin, Tidio Pricing 2026: $59 to $749, Nothing In Between: https://chatarmin.com/en/blog/tidio-pricing
eesel AI, DeepL Pricing in 2026: Every Plan, Every Limit, and One Gotcha Most People Miss: https://www.eesel.ai/blog/deepl-pricing
Langbly, DeepL API Pricing 2026: Developer vs Growth: https://langbly.com/blog/deepl-api-pricing-guide/
Rework, Best AI Tools for Multilingual Customer Service in 2026: https://resources.rework.com/tools/ai-tools/best-ai-tools-for-multilingual-customer-service-2026
Information Commissioner's Office, Guidance on AI and Data Protection: https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/artificial-intelligence/guidance-on-ai-and-data-protection/
Bratby Law, Is There a UK AI Act? UK AI Regulation in 2026: https://bratby.law/uk-ai-regulation-what-the-law-says/
Pinsent Masons, DMCC Act Overhauls UK Consumer Law Enforcement and Expands Consumer Rights: https://www.pinsentmasons.com/out-law/analysis/dmcc-act-overhauls-uk-consumer-law-enforcement
CMS, The DMCC Act: Consumer Elements Come Into Force From 6 April 2025: https://cms.law/en/gbr/legal-updates/the-dmcc-act-consumer-elements-come-into-force-from-6-april-2025
Slaughter and May, A New Age Dawns for Consumer Protection in the UK: https://www.slaughterandmay.com/insights/new-insights/a-new-age-dawns-for-consumer-protection-in-the-uk/
ServiceNow, AI + EQ = the Future of Customer Experience (Consumer Voice Report 2025): https://s205.q4cdn.com/537566246/files/doc_news/AI--EQ--the-Future-of-Customer-Experience-2025.pdf
Consultancy.uk, Chatbot Frustration Sees UK Customer Satisfaction Slump: https://www.consultancy.uk/news/33122/chatbot-frustration-sees-uk-customer-satisfaction-slump
CFOtech, UK Consumers Are Turning Their Back on AI Chatbots (Cavell research): https://cfotech.asia/story/uk-consumers-are-turning-their-back-on-ai-chatbots

Frequently asked questions

How many customers in the UK speak a language other than English?

The ONS reported from Census 2021 that 8.9 percent of people in England and Wales, over five million residents, have a main language other than English. Polish is the largest group at 612,000, followed by Romanian, Panjabi and Urdu. Around 4.1 million of them speak English well, but many still prefer their own language when something goes wrong with an order.

What is the difference between an AI chatbot that translates and one that is truly multilingual?

A translation layer bolts a service like Google Translate onto an English bot, so every message is translated in and out and meaning is lost each way. A native multilingual AI reads the customer's question directly and writes the reply in that language from your English knowledge base. Ask the vendor whether the AI answers in Polish natively or translates, and if they hedge, it is a translation layer.

How much does multilingual AI customer service cost for a small business?

On Tidio, the Lyro AI add on is around £25 a month for 50 conversations on top of a base plan of roughly £23 to £46. Fin charges about 75p to 80p per resolved conversation with a 50 outcome monthly minimum, plus seat licences. Most tools bill in dollars, so the pound figures are approximate, and you should add £60 to £120 per language for a native speaker to check your policies.

Is Google Translate good enough for replying to customers?

It is fine for understanding what an incoming message says so you know what the customer wants. It is a poor choice for sending replies, because the output reads like a machine wrote it and UK consumers already say chatbots miss tone. Use DeepL for your written policies and a native LLM tool for live replies instead.

Do I need to comply with UK GDPR if I use an AI chatbot?

Yes. The ICO's guidance makes clear that a chatbot processing names, emails and support messages is fully inside UK GDPR. You need a documented lawful basis, a privacy notice that mentions AI, an interface that makes it obvious the customer is talking to a machine, and a Data Protection Impact Assessment before launch. Also check the provider's contract covers UK international data transfers, not just EU rules.

Can I get in trouble if my AI gives a wrong answer in another language?

Potentially, yes. Since April 2025 the CMA has direct powers under the DMCC Act to fine businesses up to £300,000 or 10 percent of global turnover for misleading consumers, and a wrong claim about returns or prices is misleading in any language. The CMA is focusing on serious and systematic cases first, but a repeated wrong answer to hundreds of customers is exactly the pattern it targets. Have a native speaker check your translated policies once.

Which languages should a UK business support first?

Count your actual enquiries over three months rather than guessing, because most owners get this wrong. Nationally the largest groups are Polish, Romanian, Panjabi and Urdu, but it varies a lot by town. Launch with the two languages you can afford to have checked by a human and only add a third once you have seen it in real enquiries more than a handful of times.

What should the AI handle and what should stay with a person?

Start with five categories only: opening hours, delivery times, order status, how to book or buy, and how to return or cancel. Everything else should escalate to a human with the original message, an English translation and a short summary attached. Complaints, refunds and any regulated advice should never be answered by the AI in a language you cannot read.

Is it worth automating multilingual support if I only get a few foreign language enquiries?

Usually not. If you get fewer than a handful a month, pay a freelance translator on a small retainer of £30 to £50 a month to answer them by email within a day and say on your website that you welcome enquiries in those languages. That is a human, it is cheaper, and it works better than a bot at that volume.

Does DeepL work for all languages my customers might speak?

DeepL is excellent for European languages such as Polish, German, French and Spanish and gives you 50,000 characters a month free. It supports far fewer South Asian and African languages, and independent comparisons put it behind general AI models for Japanese, Korean and Chinese. Test it on real customer messages before paying, and get a native speaker to review anything about money.