Answering in their language is not the same as translating
Most AI chat products support a long list of languages the way a menu supports them: the words change, the product does not. The difference shows up fastest on the phone.
A translation layer is fine for a support article. It is much less fine in a conversation, where people use the words they actually use, whether a local product name, a colloquial size a place name nobody outside the city would recognise. An agent that translates your English content into Bangla and back will answer confidently and slightly wrongly.
Where it breaks first
Product questions. Someone asks for a size in the way it is described locally, not the way your catalogue spells it, and a translation layer maps it to the wrong item. This is not a language problem, it is a retrieval problem: the agent has to search your catalogue in the language the question arrived in.
And then on the phone
Voice is where the gap is starkest. Plenty of the fast, cheap speech models simply do not offer Bengali at all. That is a question of availability rather than quality. You either have the language or the phone rings out.
We built the voice side specifically so an agent can answer a call in Bengali with the same knowledge as the chat agent. Not a separate script, not a separate knowledge base, and not a translation of an English one.
What good looks like
- The customer never has to choose a language. The agent answers in the one they used.
- A product question in Bangla searches the catalogue, not a translated copy of it.
- When it escalates, the transcript reaches your team in the original language, with context.
- The same agent handles it on the website, in WhatsApp, and on the phone.