One customer call, link by link: seven places it breaks

On 10 September, OpenAI released GPT-Live-1 in the API at $0.05 a minute. The price explicitly covers the voice layer only – the part that listens and speaks. Deeper reasoning and tool calls are delegated to a text model the developer picks, and pays for separately.

That is an accurate description of something that has been true for a while: a voice agent is not one thing. It is a chain of links that each have to hold in turn, within a few seconds, in a fixed order. Here is one perfectly ordinary call – a shop owner ringing his supplier about a delivery that arrived wrong – and what has to hold at each link.

The phone rings

The first link has nothing to do with AI. The call has to travel from the caller's operator, onto your number, through your telephony provider and to the agent, ideally inside three or four rings.

It is also the only link with nothing to fall back on. If it fails, the caller hears nothing at all. Who owns the number, and where the call goes when the agent does not answer, is decided here – not in an argument about which model to use.

The first sentence

The agent says something. It is the one sentence a hundred per cent of callers hear, and it has to do four things at once: say who is answering, say that this is an AI, offer a way out, and ask one question open enough that the caller starts talking.

The failure is nearly always the same. The sentence was written last, in five minutes, by someone who wanted to be finished.

"Yeah, hi – it's about a delivery that came yesterday"

The caller tells you everything at once, in the wrong order, with a detour about how he rang last week as well. Nobody speaks in form fields.

What has to hold here is that the agent does not try to steer him back to step one. It should pick up what it got for free – that this is about a delivery, that it arrived yesterday, that something is wrong – and ask only for what is missing. An agent that asks about something it has already been told is the fastest way there is to lose a caller.

The order number

Now a string of digits has to be read out, and this is the weakest point in the whole chain – more so in the Nordic languages, where the numbers can be said two ways and both are correct, while the phone line throws away much of what separates one digit from another.

What has to be true here is not that the agent hears correctly every time. It is that it reads the number back when it is unsure, instead of writing down something that merely looks plausible.

The four seconds when nobody says anything

The agent has the number. Now something has to be looked up, and the lookup takes time. In a chained setup this is usually where the call falls apart: the agent goes quiet, the caller assumes the line is dead, and then both talk at once.

This link is precisely what the new generation of models is built for. OpenAI calls it "silent context management" – handling silence and background noise without interrupting the conversation or narrating every step out loud. The language-learning service Speak reports that interruptions during learners' thinking pauses fell by almost 80 per cent against its previous, turn-based setup.

But whether those four seconds feel right is not settled by the model alone. It is settled by whether anyone has decided what the agent does while it waits – and what it says if the lookup never comes back.

The answer the agent is not allowed to give

The lookup comes back: wrong item shipped, and it is the supplier's fault. The caller asks whether he gets a credit note.

This is where the call stops being about technology. The agent is perfectly capable of composing an answer. The question is whether anyone has decided it may. A yes to a claim binds you whether it comes from an employee or from an agent, and that boundary should be written down before the first call, not after the first complaint.

The handover

The agent transfers. If the person who picks up opens with "hi, what's this regarding?", the preceding four minutes were wasted – and worse, the caller has now explained the same thing twice to the same company.

It is the context that dies at a handover, and it dies however good the voice model was. This link does not come with the model. It has to be built.

What a model swap actually fixes

Seven links. One of them is the voice model.

The useful thing about the price now being quoted separately is that it makes clear what an upgrade can and cannot solve. A better voice model helps with links three, four and five. It does nothing about who owns the number, what the agent is allowed to promise, or what the human is told when they take over.

Threll.ai builds voice agents in Norwegian, Swedish and Danish. Walk the seven links of your own setup and mark the ones a new model would improve. They are rarely the same links that are actually holding the project up.