Designing AI for multilingual customers: language, channels and data

Key takeaways

• Meet customers where they already are: WhatsApp, LINE, Telegram, email or web chat, depending on the market.
• Language is not only a translation problem. Tone, politeness and mixed-language messages are part of the design.
• Voice is still a primary channel in many markets and remains under-automated.
• Data protection law should shape the architecture from the start, whichever regime applies.

Much of the AI tooling sold to enterprises is designed around a default customer: English-speaking, email-first and based in North America or Europe. Deploy it for customers elsewhere and the gaps appear quickly. Customers do not write emails; they send messages. They do not stay in a single language; they move between two in the same sentence. The experience feels foreign because it was built for somewhere else.

These are the lessons we keep returning to when designing for customers in many languages and markets.

1. Go where customers already are

In every market, customers have a messaging app they live in. It might be WhatsApp, LINE, Telegram or plain SMS and email. It is where they talk to friends, follow brands, book services and ask businesses questions. Making customers leave it to use a web form or a separate app adds friction that a competitor on their channel will not have.

The design implication is that the conversational layer should sit natively on the channels customers already use, with the same knowledge and the same hand-off to people whichever channel the customer arrives on. That is why our concierge, Primues, is built as one layer across all of them.

2. Language is not only translation

Treating a language as a translation target produces replies that are technically correct and socially wrong. Languages differ in how they signal politeness and register, how formal a business message should sound, and how often people mix in terms from another language. A reply that is too blunt can feel rude. One that is too formal can feel cold.

• Define the tone your brand uses in each language and test replies against it with native speakers.
• Handle mixed-language messages naturally, including product names and technical terms.
• Respect the honorifics and politeness conventions appropriate to the customer and the context.
• Test with real messages, including slang, typos and voice-note transcripts, not only clean examples.

3. Do not forget the phone

In many markets the phone remains a primary channel for appointments, payments and support, and it is among the least automated. Callers wait in queues, clinics are tied up answering routine calls, and reminders are skipped because nobody has time. Natural-sounding voice, with live hand-off to a person, is a significant opportunity. It is also hard to do well, because latency, tone and interruptions matter far more in speech than in text.

4. Design for data protection from the start

Customer conversations contain personal data, and the law that protects it differs by country: the EU and UK GDPR, Thailand's and Singapore's Personal Data Protection Acts, India's Digital Personal Data Protection Act and many others. The practical implications for AI systems are similar everywhere: collect only what is needed for a stated purpose, keep records of how data is processed, control who and what can access it, set retention periods, and be ready to notify the regulator promptly if there is a breach.

It is much easier to design these properties in than to add them later. Systems that operate with scoped permissions, keep a complete record of their actions and handle sensitive steps through approval are in a far better position to meet these obligations, and organizations operating in several countries should plan for all the regimes that apply. Our data protection and AI checklist lists the questions to ask.

Putting it together

None of these lessons is exotic. They come from taking the customer's actual environment seriously: the channel they use, the language they speak, the way they prefer to be addressed and the law that protects them. AI that respects those details feels natural. AI that ignores them feels like an import.

If you are planning an AI initiative across languages or regions, we would be glad to compare notes. You can speak with Primues to get started, or see what this looks like in customer support automation and on our multilingual AI automation page.