A voice agent on a Norwegian number is almost always configured for Norwegian. That is right for most of the calls, and wrong for a small and expensive minority: the customer calling from Sweden, the guest who lives in Denmark, the group buyer who never learned Norwegian.
Below is one such call, moment by moment. She dials a Norwegian service number at a quarter to eight in the evening, and she speaks Swedish. Nothing in the setup is obviously broken. But at six points in the call something has to have been decided in advance, or it ends somewhere nobody would have chosen.
19:42 – the call arrives
All the agent knows at this moment is which number is calling. Nothing else.
A Swedish country code is a hint, not a fact. Norwegians hold Swedish subscriptions, Swedes hold Norwegian ones, and a switchboard number says nothing about who is holding the phone. The country code is worth using as a starting point for which language the agent tries first. It is not worth building the rest of the call on.
19:42:02 – the opening line picks the language for the whole call
Somebody decided which language the agent greets in. That is a decision, including when it was made by not being made.
There is also a duty here. The agent has to disclose that it is artificial intelligence, and that is the cheapest requirement in the regulation to satisfy. But a disclosure given in a language the recipient does not command has been said without informing anyone. So the question is not whether the agent says it. The question is whether it says it again, in the new language, at the moment the call switches – and no vendor has answered that one on your behalf.
19:42:11 – she answers in Swedish
This is the moment that decides the rest.
Supporting two languages and being able to switch between them inside a single call are two different capabilities. They do not come bundled by default. Retell AI, which states this openly in its own documentation, notes that mid-conversation language switching depends on which speech recognition provider is handling the call: some support switching between every language they cover, while others transcribe one language per call. Coverage cannot be stitched together across providers mid-call.
That is worth taking into the next vendor meeting. "We support Norwegian, Swedish and Danish" does not answer the question. The question is whether the agent can move from one to the other eleven seconds in, on the same call, without hanging up.
One more detail in those same lists deserves a look: Norwegian appears as a single row. Not Bokmål and Nynorsk, not dialects, not tonal accent. One language, one entry – and the rest has to be tried on your own traffic before you believe it.
19:43 – the number
She reads out her order number, and she does it in Swedish.
"Fyrtioåtta" and "førtiåtte" are almost the same word for 48, and "almost" is the entire problem. A transcription locked to Norwegian forces what it hears into the nearest thing it knows, and the result is a number that looks perfectly correct in the log. Numbers are the hardest thing a voice agent hears – and a number heard in the wrong language fails silently.
The antidote is the same as ever: a check digit, a lookup that can reject the guess, or a confirmation read back to the customer. Read back in her language, so that she can actually hear that it is wrong.
19:44 – the handover
The agent cannot get further, and transfers.
Now a human is going to pick this up, and two things have to travel with it. One is almost always remembered: what the case is about. The other is almost always forgotten: which language she is speaking. Escalation is the part of the setup that is most expensive to do halfway, and a transfer that drops a Swedish-speaking customer into a queue where nobody has been told has done something worse than nothing. It has spent forty-five seconds gathering information that did not travel.
19:47 – what got written down
The call is over, and something was saved.
The question is what. The summary is written in one language, the customer's own words arrived in another, and somebody is going to follow the case up tomorrow morning. If the language sits as a field on the case, the callback goes out in Swedish. If it does not, the callback goes out in Norwegian and the whole round repeats – with the difference that this time she already knows it does not work.
Language is a property of the call, not a setting on the agent
The six moments have one thing in common. None of them is solved by a better model.
They are solved by somebody having decided what happens at each of them: which language the agent opens in, whether it can switch part-way, whether the disclosure is repeated when it does, what happens to a number read in another language, what the human is told at the transfer, and where the language is stored once the phone is down. All six are decisions. Most of them are made in practice by not being made.
Threll.ai builds voice agents in Norwegian, Swedish and Danish. In the Nordics, a customer changing language mid-call is not an edge case. It is a Tuesday.
