Opinion · Frontline Intelligence
Speak to think.
In January, we ran a pilot with the Global TB Caucus, a parliamentary network with presence in over seventy countries. Fifteen people took part. They were scattered across four continents, from Indonesia to Zimbabwe to Brazil to the Americas. Some had been doing the job for a few months. Others had been doing it for a decade. The pilot was a tight, well-structured sprint, concise by design, organised so we could move fast and learn fast. Each participant completed twelve bespoke voice simulations of the conversations that make up their actual work: introducing the Caucus to a sceptical Member of Parliament, briefing on regional data, refusing a cash request cleanly, coaching a junior colleague through a difficult moment.
We had built the platform around what we assumed was the obvious thing. The agent. The simulated other side of the conversation. We spent a few intense weeks on its voice, its accents, its responsiveness, its ability to push back. We move fast, that is how this team ships. We expected feedback to be about the agent. Was it convincing. Did it feel real. Did it ask the right questions.
The feedback was about something else entirely.
The most experienced participant in the pilot was Matt Oliver, who has been pitching this cause to parliamentarians for over a decade. He had every reason to be the most sceptical user we had. By the end of the sprint, he sat down with us and said this:
“The valuable bit is being able to explicitly practice and articulate out loud. The thing I see as the art of those kinds of exchanges is how fluently and cogently you can put together the stuff that comes out of your own mouth.”
He said the valuable bit was when he was talking. Not when the agent was responding. The agent could have been almost anything as long as it gave him something to react to. The point was the practice itself, the act of forming sentences under pressure, finding the right line, hearing it out loud, knowing it could be better, going again.
He was not the only one who said this. We just had to listen for it.
Alesia Matusevych has been a Regional Manager for six years, covering Eastern Europe. Her version of the same observation was that the platform helped her structure her thinking. She did the exercises in the mornings, before her real calls with parliamentarians, and what she took away was a clearer way of organising what she was going to say. The simulation gave her a low-stakes space to find the structure. The structure was what she carried into the high-stakes room afterwards.
Joel Musampa joined the Caucus four months before the pilot. He covers Central West Africa. He found the experience difficult, in a good way. His description was that the simulation forced him to be diplomatic and cautious with his word choice, in the same way he would have to be with a real MP. For someone still developing the patterns, the value was in the act of producing careful speech, again and again, until the careful speech started to feel natural.
Three users. Three continents. Three different levels of experience. The same observation underneath all of it. The value was in the speaking, not in the listening.
This was not what we had been optimising for.
We had quietly assumed, like most of the voice AI industry assumes, that the product was the agent. Build a better agent, and the value follows. Better voices, sharper personalities, more emotional range, faster latency. Make the simulated other side of the conversation as close to a real human as possible, and the user will get a better workout.
It is a reasonable hypothesis. It is also wrong, or at least it is wrong as a centre of gravity. The pilot quietly showed us that beyond a certain threshold of realism, more agent quality does not produce more user value. What produces user value is something else: an environment in which the user is willing to keep talking, willing to take the risk of saying something imperfect, willing to listen to themselves and try again. The agent matters only insofar as it sustains that environment. Past that, it disappears into the background.
The agent is not the product. The agent is a forcing function. The product is the environment that makes the user comfortable enough to keep speaking honestly.
So this is the strategic decision we have made, and we are stating it openly because we think it explains a lot of what we will and will not build over the next year.
We are not in the race to make our agent more human. Beyond a workable level of realism, that race has diminishing returns. We are in a different race, which is to design user-led calls. Calls in which the user is the centre of gravity. The agent says enough to give the user something to push against, and then it gets out of the way. The simulation structure is deliberately simple, because simplicity protects the user’s flow. The scoring evaluates the user’s speech, not the agent’s behaviour. Every piece of friction we have removed in the last three months has been on the user side: how they enter the call, how they recover from a stumble, how they read feedback, how they decide to go again.
This produces a clear opinion about what voice AI training is for. It is not a Turing test. It is not an attempt to fool the user into thinking they are speaking to a real person. It is an attempt to create the conditions under which the user does the work of speaking, hears themselves doing it, and improves.
The phrase we have started using inside the company is speak to think. The act of articulating something out loud is not a delivery mechanism for a thought already formed in the head. It is the place the thought gets formed at all. Anyone who has tried to explain a complex idea to a colleague and only really understood it themselves halfway through the sentence knows this. The simulation is the colleague. Or rather, the simulation is the space in which a person can be that colleague to themselves.
Designing for that is a different discipline to designing a more lifelike agent. It is closer to the work of a good interviewer, or a good coach: knowing when to speak and when to stop speaking, knowing how to hold the space, knowing that the most valuable thing in the room is the person you are listening to.
We did not arrive at this by sitting in a room and theorising about voice AI. We arrived at it by deploying to fifteen people across four continents and listening to what they told us. The conviction that came out of that experience is now driving where we put our engineering hours, our research time, and our intellectual property. We are building for the user’s voice, not the agent’s.