Building a Multilingual AI Agent for East Africa

A customer messages your WhatsApp support line: 'Boss nilibuy hii product wiki iliyopita na bado haijafika. Si uniambie status?' That's English, Swahili, and Sheng in three clauses. A monolingual AI agent will misread this, ask for clarification in formal English, and frustrate a customer who was perfectly clear. Multilingual support in East Africa is a practical, not theoretical, requirement.

The Specific Language Challenge

East African urban markets, particularly Nairobi and Dar es Salaam, have communication patterns that differ from what most multilingual AI systems are built to handle. The challenge is not just that customers write in Swahili — most modern LLMs handle formal Swahili reasonably well. The challenge is code-switching: the natural alternation between English, Swahili, and Sheng within a single message or even a single sentence.

Sheng is particularly difficult: it's a continuously evolving urban slang that blends Swahili with English, Kikuyu, Luo, and other languages. Terms that were Sheng two years ago may be mainstream Swahili now, or may have shifted meaning. No static training dataset fully captures it.

Language Detection vs. Language Handling

Most multilingual chatbot implementations work by detecting the language of the input and then routing to a version of the agent configured for that language. This breaks down for code-switched inputs: the language detector classifies the input as 'English' (because more words are English) and routes to the English agent, which then misses the Swahili and Sheng content.

The better approach is to not detect and route at all — instead, use a model capable of handling mixed-language input natively, with a system prompt that explicitly acknowledges the multilingual context and instructs the agent to respond in whichever language the customer has used. This keeps responses natural and maintains the customer's own linguistic register.

Configuring the Knowledge Base for Multilingual Markets

Your knowledge base needs to contain content in the languages your customers use. An English-only knowledge base serves a Swahili-speaking customer poorly — even if the AI model can translate, the retrieved chunks will be in English and the response will feel foreign. Maintain parallel versions of key documents (product terms, FAQs, policies) in Swahili and English, and let the retrieval system pull from both.

• Maintain key FAQs in both English and Swahili as separate chunks
• For Sheng-heavy markets, add a glossary of common terms to the knowledge base
• Configure the agent to match the formality level of the customer's message
• Don't force formal Swahili if the customer is using casual Sheng — it reads as condescending
• Test with native speakers before launch — automated testing won't catch register mismatches

Response Language Strategy

The clearest heuristic: respond in the language the customer used most in their last message, at the same formality level. If the customer writes in Swahili, respond in Swahili. If they write in English, respond in English. If they code-switch, respond in whichever language their question was phrased in. Attempting to match their exact code-switching pattern is technically impressive but often feels unnatural — a slight formality gap is more comfortable than an AI trying to sound like it grew up in Nairobi.