Four languages on one phone line

The UAE's supply chain runs on English, Arabic, Hindi and Tamil, and it does not run on them separately. A single call to a trading company in Al Quoz might open with a hello in English, move into Hindi once the product is understood, and finish with a price read back in a mixture of both. Anyone who has bought materials here knows this. Anyone building a voice agent here discovers it in the first week.

The obvious approach, and why it fails

The obvious thing to do is to detect the language and match it. Speak what the supplier speaks, immediately, on every turn. It sounds like courtesy, and it produces some of the worst calls we have ever recorded.

The reason is that a phone call gives a language detector almost nothing to work with at the moment it most wants to decide. The first thing anyone says is "hello" — one word, often not in their first language, and frequently the only word spoken before a decision has to be made. "Yes" is not evidence. "Hmm" is not evidence. A greeting is the single least informative thing a person says on a telephone, and it is the first thing they say.

There is a second trap underneath the first one. Multilingual transcription writes down what it hears in whichever script it thinks it is hearing, so a supplier saying "ah, yes" in plain English can arrive as Tamil characters. An agent that reads the alphabet rather than the words then answers in Tamil — to someone who has been speaking English the whole time and now has to stop the call to ask what is happening.

Stay in English until told otherwise

So our agent does the unglamorous thing. It opens every call in English, whatever language it was greeted in, and it moves only on real evidence:

And it counts words rather than characters, precisely because the script a sentence arrives in is not the language it was spoken in. Fewer than about five words is never a reason to switch. If it is unsure whether that was a real sentence or just a greeting, it stays where it is.

Flipping between two languages mid-call is worse than picking one and keeping it.

Once it does switch, it stays switched — a supplier who asked for Hindi should not have to ask twice — and it writes each language in its own script rather than romanising it, because a price read back in transliterated Hindi is a price nobody can check.

The part that is harder than the languages

Arabic and Hindi both mark the gender of the person being spoken to, and an agent working from a transcript never hears a voice. Guessing from the sound of a name is exactly the kind of shortcut that offends somebody on the first call.

The agent therefore stays neutral until it has an actual signal — the supplier's own words about themselves, or a name they have given — and until then it rephrases around the problem: addressing the company rather than the person, and asking about stock and price, which carry no gender in either language. It is a small thing that costs nothing to get right and is very noticeable when it is got wrong.

Why an AI voice agent for supplier calls has to get this right

Because the alternative to an agent that speaks a supplier's language is not an agent that speaks English. It is a supplier who does not quote. The whole point of calling forty companies instead of eight is to reach the ones that are not already in your contacts — and in this market, a good number of those answer the phone in Hindi or Tamil and will happily give you a better price than the trader with the English website.

There is more on how a call runs in the common questions about the agent, or you can see the agent call a real supplier and hear the switching rule for yourself.