Pillars 3 and 4 become real code. Wire each language to its own system prompt and its own KB shard. The agent doesn't just respond in Swahili — it responds with Swahili-native business facts. Same stack, 5 distinct customer experiences. Under $300 marginal to go from 1 to 5 languages.
Each step is concrete. By the end: 5 native brains running on the same agent. Same code. 5 distinct customer experiences.
One system prompt per language. Same agent personality, expressed natively. FR: polite local idioms. SW: greeting opens a small conversation. RN: honorifics for elders. Store as PROMPT_FR, PROMPT_EN, PROMPT_SW, PROMPT_AR, PROMPT_RN. NEVER translate the EN prompt with Claude — write each from scratch.
For each language, build a separate set of documents (FAQs, product specs, policies) — written or reviewed by a native, NOT auto-translated. Chunk into 500-token pieces, embed with text-embedding-3-small or Voyage AI, store under per-language namespace (Pinecone) / metadata column (Supabase) / collection (Chroma).
Detected language code → matched system prompt + matched vector namespace. Configuration is data, not branching code everywhere. Adding a new language = adding one config entry, not a code rewrite. The detector decides; the router looks up.
Same retrieval-augmented generation pattern from Lesson 3.3 — but constrained to a single language shard. Embed customer question → search vector DB scoped to matched namespace ONLY → top 3 chunks → inject into matched system prompt → Claude generates. Native answer, native tone, native facts.
Ship one language at a time. Pick your strongest market first. Send 10 real customer questions to your native reviewer. Catalog: off tone, wrong fact, broken idiom. Patch prompt + KB + re-test. 5 languages over 5 weeks > 5 languages in 1 week (all mediocre).
Top AI Africa deploys the full stack — language detection, native prompts, native KB per language. WhatsApp agents sounding native in French, English, Swahili, Arabic, Kirundi. Free 15-min strategy call.