AI customer support in Bemba and Nyanja: what actually works for a small Zambian business
A shop running customer questions and orders through WhatsApp Business quickly runs into a gap that has nothing to do with the product being sold: a lot of customers don't write to it in English. They write in Bemba, in Nyanja, or in a mix of both with English dropped in mid-sentence, because that's how they'd naturally text a friend. An AI assistant or chatbot built and tested only in English either misreads the message entirely, replies with something that doesn't match what was actually asked, or the business ends up routing that customer to a phone call anyway — which defeats the point of automating the conversation in the first place.
The way this actually gets solved isn't a single model that's fluently bilingual in Bemba and Nyanja the way general assistants are in English. It's usually a chain: the incoming WhatsApp message first goes through a translation model that explicitly covers the language in question — Meta's NLLB-200 translation project, for example, includes both Bemba and Nyanja among the roughly 200 languages it supports, which most general-purpose chat models don't handle nearly as well — and the translated English text is what actually gets sent to the assistant generating the answer. The reply then gets translated back into the customer's language before it goes out over WhatsApp. To the customer, it looks like one continuous conversation in their own language; underneath, it's two translation steps wrapped around the actual assistant.
Where this still falls short is worth being honest about. Translation quality for Bemba and Nyanja is meaningfully behind what the same models manage for English or French, simply because there's far less written text in those languages to have trained on, and short, informal WhatsApp messages — abbreviations, code-switching mid-sentence, dropped words — are the hardest case for any translation layer to get right. A pricing question or a complaint is exactly the kind of message where a garbled automated reply going out under a business's name does real damage, so a system worth using needs a confidence check that hands the conversation to a person rather than guessing when the translation looks shaky.
We build WhatsApp Business and chatbot integrations for Zambian businesses, and wiring in a translation layer alongside the assistant — plus the rule for when to hand a conversation to a person instead of letting the bot guess — is part of what we set up before it goes live. If your business runs customer support over WhatsApp and quietly assumes every customer is comfortable typing in English, that's usually a real slice of your customers you're currently serving worse than you think.