AI Calling Campaigns
Multilingual AI calling in your customer's language
Multilingual AI calling done properly: choosing each person's language, switching mid-call, translating disclosures and offer terms, and testing first.
Multilingual AI calling works when you treat language as data, not as a guess. Pass each person's preferred language with the contact, open the call in it, and let the agent follow if they reply in another one. Then translate everything that carries a legal or practical promise, meaning the AI disclosure, the offer terms, the opt-out and the help information, and have native speakers test it before the campaign goes live. With per-minute pricing the language itself costs nothing extra. The real cost is the half-day of work to get each language's wording right.
Most writing on multilingual voice AI is about the model: how many languages, how natural the accent, how fast it detects a switch. Those things matter, but they're rarely why a multilingual campaign fails. Campaigns fail because the list didn't say who speaks what, because the offer terms were translated by guesswork, or because the call was in Dutch and the follow-up text in English.
This guide covers the campaign side, the decisions you make before Milo, or any AI agent, dials anyone.
Why the customer's language matters on the phone
Text can be put through a translator. A phone call can't. Someone who doesn't follow the first ten seconds hangs up, and you don't get a second chance.
Two data points frame the size of the problem. In the EU's 2024 language survey, 59% of Europeans said they can hold a conversation in at least one foreign language, and 28% in two or more. Put the other way, about four in ten can't hold a conversation in any language but their own. And in CSA Research's survey of 8,709 consumers in 29 countries, 76% of online shoppers said they prefer to buy products with information in their native language, and 40% said they never buy from websites in other languages. That study was about websites, not calls, but there's no reason to think people are more tolerant of a foreign language spoken at conversational speed.
We saw a smaller version of this ourselves. In an October 2026 pilot with an online casino operator, calling players in the Netherlands, some players answered in Dutch. An agent that kept going in its opening language would have lost those calls. Language switching went from a nice extra to a requirement.
Step one: decide each person's language before the call
There are three sources for a language, in descending order of reliability:
| Source | Reliability | Use it for |
|---|---|---|
| The language the person chose with you (account setting, sign-up locale, support history) | High | The opening language of the call |
| What they say on the call | High, but only after the opening | Switching mid-call |
| The phone number's country | Low | A fallback when you know nothing else |
Country is a weak signal for two reasons. Several countries have more than one national language: a Belgian number could be a Dutch, French or German speaker. And every market has residents who prefer another language: a UK mobile might belong to someone who speaks Polish at home.
On Dial Milo you pass the language per person. In a CSV upload it's a column. Extra columns are saved on the contact as player details, as described in managing contacts and lists. In a webhook from a journey tool it's a field on the request:
{
"phone": "+447700900123",
"name": "Kasia",
"external_id": "evt-8812",
"language": "pl",
"offer_code": "FS20"
}
Milo opens the call in that language. If the person answers in another one, Milo switches and carries on in theirs. The full webhook set-up, including how outcomes flow back to Customer.io, is in Customer.io AI phone calls.
Step two: handle the switch gracefully
A mid-call switch is where people judge an AI agent. Three behaviours separate a good switch from a jarring one:
Follow, don't ask. If someone answers "Hallo, met wie spreek ik?" to an English opening, carry on in Dutch. Don't ask "Would you prefer Dutch?" They've already told you.
Don't restart the script. A switch should pick up where the conversation was, not repeat the greeting and disclosure from the top. Our pilot showed something similar with a plain "yes": a "yes" spoken in the middle of a sentence must not restart the pitch. Language changes need the same care.
Re-confirm what matters. If the identity check or the opt-out happened in one language and the call continued in another, confirm the key point once more in the new language. "Just to confirm, you'd like us not to call again?" is cheap insurance.
Code-switching, mixing two languages in one sentence, is normal for many bilingual people. Test it on purpose. A Spanish-English or Dutch-English speaker shouldn't be forced to pick one.
Step three: translate the parts that carry promises
The conversation itself is generated live. Some parts of every call shouldn't be. These are the sentences your legal, compliance or marketing team signs off, and each one needs an approved version in every language you call in.
| Element | Why it needs an approved translation |
|---|---|
| AI disclosure | The EU AI Act requires people to be informed that they are interacting with an AI system, unless this is obvious, and the Act has applied since 2 August 2026. A disclosure in a language the person doesn't follow isn't really a disclosure. |
| Who is calling and why | Telephone marketing rules in the UK require you to say who is calling. The same logic applies. |
| Offer and exact terms | A bonus or discount translated loosely is a different offer. "Up to" and "from" do not survive careless translation. |
| Opt-out wording | The person must be able to say no in their own language, and the agent must understand it. |
| Help and support information | For gambling in Great Britain, the Commission's code says operators marketing in a foreign language should provide responsible gambling information and the summary of terms in that language. |
On Dial Milo, the offer and its exact terms are written by you. Milo presents the offer once, never invents or changes a term, and gives conditions when asked. If you call in four languages, write and approve the terms in four languages. The script itself can stay in one language, because Milo speaks the person's language from it, but have a native speaker check the result on test calls.
Step four: keep the follow-up in the same language
The call is often the middle of a journey, not the end. A player accepts a bonus by phone, and an SMS and an email with the code follow. If the call was in Portuguese and the text arrives in English, much of the goodwill from the call is lost.
The simplest rule: the language field that started the call drives every message after it. If your journey tool sent language: "pt" to start the call, it already knows which template to use for the follow-up. With Customer.io connected, Dial Milo sends the outcome back as an event (milo_offer_accepted, for example), and your journey picks the Portuguese SMS and email from the same language attribute.
Where multilingual calling still breaks
Being honest about the weak spots saves you from finding them live.
Names, spelling and numbers. Names from one language spoken in another are where recognition fails most often. Any detail you'll act on, like a callback time or a phone number, should be read back for confirmation.
Regional variety. Flemish and Netherlands Dutch, Brazilian and European Portuguese, Swiss and German German all sound different. Test with speakers from the markets you're actually calling.
Voice fit. A voice that sounds warm in English can sound odd in Polish. Preview voices in each language you'll use. See choosing a voice.
Unknown callers on the campaign number. When someone who isn't on your list rings the campaign number back, Dial Milo answers with a short neutral line that's always in English, and points non-English speakers to email. Known players who call back get their campaign's script, as if the conversation had simply continued. Details are in when someone calls your campaign number back.
Worked example: one list, four languages
Illustrative figures. Say a UK online retailer is running a win-back campaign to 2,000 lapsed customers. Its account data shows the preferred language for 1,850 of them: 1,500 English, 200 Polish, 100 Portuguese and 50 Romanian. The remaining 150 have no preference recorded.
- Tag: 1,850 contacts get their language code from account data. The 150 unknowns default to English, since every number is a UK mobile and the country can't tell you more.
- Translate: the offer ("15% off your next order, for 14 days, minimum spend £40"), the opt-out line and the follow-up email are approved in Polish, Portuguese and Romanian. That's three short documents, not three scripts.
- Test: a Polish-speaking colleague takes ten test calls and finds that "minimum spend" was being rendered too loosely. The approved Polish terms are tightened before launch.
- Watch: after the first 200 calls, the transcripts show some of the 150 "unknown" contacts switched to Polish within the first sentence. The team records their preference in the CRM, so the next campaign opens in Polish.
The result is a campaign where nobody had to choose a language from a menu, and the CRM knows more about its customers afterwards than it did before.
A launch checklist for a new language
- 1
Tag the list
Add a language code to every contact you can. Where you can't, decide on a fallback and write it down.
- 2
Approve the fixed wording
Disclosure, reason for calling, offer terms, opt-out and help information, translated and signed off by whoever signs off the English version.
- 3
Native-speaker test calls
At least ten test calls per language: switch languages mid-call, spell a surname, ask for the terms twice, say no, say "stop calling". Read the transcripts.
- 4
Small live batch
Call a few dozen people in the new language first. Compare conversation rate, opt-out rate and outcomes with your main language before you scale.
- 5
Watch the first reports
Check the campaign's transcripts and summaries for misheard names, terms said wrongly or switches that restarted the pitch. Fix the wording, not just the setting.
What it costs
On per-minute pricing, language doesn't change the bill. A Dial Milo outbound call costs the phone line, priced for the exact number dialled and billed per started minute, plus AI at 9.57¢ per talk minute. A Netherlands mobile has a line rate of 10¢ and a German mobile starts at 9.4¢, so both come to roughly 19–20¢ a minute all-in, the same whether Milo speaks Dutch, German or English. You can price any number at /rates.
That's the real change multilingual AI makes. With a human team, every extra language is another hire and another rota to fill. With an AI agent, it's an hour of translation, an afternoon of test calls, and a column in your spreadsheet. For how this fits into a wider outbound programme, start with AI outbound calling. For inbound calls, see the bilingual answering service guide. For the rules on calling players in specific markets, see calling players in the Netherlands and UK Gambling Commission rules for marketing calls.
Frequently asked questions
- How does an AI phone agent know which language to speak?
- The most reliable way is to tell it. Pass a language code per person, such as en, nl or de, from a column in your list or a field in your CRM webhook, and the agent opens the call in that language. If the person answers in a different one, a good agent follows them instead of carrying on in the first language.
- Can I just pick the language from the phone number's country?
- As a fallback, yes, but not as the main rule. Country code tells you where a SIM was issued, not what its owner speaks. Belgium and Switzerland have several national languages, and every market has residents who prefer another language. Use the language the person chose with you, and use the country only when you have nothing better.
- Does a multilingual AI call cost more than an English one?
- Not on per-minute pricing. On Dial Milo the price is set by the number you dial, your caller ID country and the minutes of talk, and the language makes no difference. The extra cost of a new language is the work of translating and checking your script, offer terms and help information.
- What has to be translated besides the conversation itself?
- Everything a person relies on to make a decision or protect themselves: the AI disclosure, the reason for the call, the offer and its exact terms, the opt-out, and any help or support information. Follow-up texts and emails should arrive in the same language as the call.
- How do I test a new language before a live campaign?
- Have a native speaker take test calls and try to break them: answer in the other language, use regional slang, spell a surname, give a phone number and ask for the terms twice. Read the transcripts as well as listening, then run a small live batch and compare opt-out and conversation rates with your main language.
Keep reading
AI Calling Campaigns
AI outbound calling: how it works, costs and rules
AI outbound calling explained: what an AI phone agent does on each call, what a campaign costs per minute, the consent rules, and how to run your first list.
ReadAnswering Service
Bilingual & multilingual answering service, no surcharge
Why 'press 2 for Spanish' loses callers, what a bilingual answering service must actually do, and how to cover 20 languages with no per-language surcharge.
ReadIntegrations
Customer.io AI phone calls: add a call step to journeys
Add an AI phone call to a Customer.io journey: the webhook payload, idempotency, the events that come back, and a full welcome-to-deposit chain you can copy.
Read