One agent. 25 languages. Zero duplicate workflows.

We detect from a stack of signals — the shopper's Accept-Language header, browser locale, prior conversation language, IP-derived country fallback, and explicit user picks via a language switcher. The detected language drives the agent's response language, recovery email locale, and product card labels in lockstep.

No. The retrieval layer is multilingual at the embedding level — an English product description will surface for a Spanish-speaking shopper based on semantic similarity. You can still upload localised variants per language for tighter brand voice; we promote them via the reranker when the shopper's language matches.

Yes. The recovery-email worker passes the detected shopper language into the prompt; the model writes the email in that language, including product card labels (price, stock, CTA). Per-locale overrides let merchants polish phrasing for brand-critical languages.

The widget renders RTL automatically when the detected language requires it. Layout direction, message bubble alignment, and CTAs all flip. Email templates also flip RTL where supported.

Yes. If a shopper switches mid-thread — e.g., starts in English then writes a Spanish question — the agent detects the switch and either continues in their new language or falls back to a 'I can help in Spanish or English — which would you prefer?' prompt. The behaviour is configurable.

Conversation analytics gain a 'language' slicer so you can compare conversion, AOV, and AI-attributed revenue by shopper language. Per-language conversation tags surface so you spot, say, that German shoppers ask sizing questions twice as often.

The list is curated based on actual LLM quality, not just 'the model claims to speak it.' If your store needs a language outside the 25, we'll evaluate quality with you and either add it or recommend a partial-feature alternative — usually retrieval-only without the full agent.