The easiest way to build an AI mentor is to write a system prompt. Give the model a name, a few adjectives, tell it to be warm and curious. That works, technically. But it doesn't feel like a person — it feels like a chatbot trying to pass as one.
We wanted to do something different.
The idea: start with a real human being
The premise was simple enough in theory. Rather than designing a mentor persona from scratch, we would interview a real coach — someone with years of experience and a distinct way of working — and use that conversation as the foundation for the AI. Not transcribe the interview and feed it in. Actually understand how that person thinks, what they believe, how they move through a conversation. And then try to recreate that.
Paul is one of the first real coaches behind Mentor Direkt. He has worked with technology leaders and organisations for 25 years. He is not a life coach with an Instagram account. He is someone who has sat with founders and executives through real crises and learned, over time, what actually helps.
The interview
We asked Paul to answer ten questions — not about coaching theory, but about how he actually works. How do you describe your communication style? What do you believe about how people change? What do you do when someone gets stuck? What do you always avoid?
What came back was unexpectedly specific. Paul talked about how he deliberately slows down in a conversation — how he waits longer than feels comfortable before responding, to give the other person space to find their own words. He talked about how he never diagnoses: he would rather ask a question that opens something than make an observation that closes it.
"People leave energised, not drained." That was one sentence in Paul's answers that stayed with us. It is not a goal he sets — it is something he noticed as a pattern in his practice. It says more about his approach than ten pages of theory.
He talked about what he believes people most need to hear but rarely do: to be kinder to themselves. Not as permission to give up, but as the kind of self-knowledge required to actually move forward. And he talked about how he holds that — carefully, without saying it directly, by asking questions that invite the person to arrive there themselves.
From person to persona
The interview gave us material to do two things. The first was a technical document — instructions to the AI model about how Paul communicates, what he does when someone gets stuck, which questions he returns to, what he actively avoids. That document is the foundation for how the Paul AI behaves in every conversation.
The second was harder: understanding why. A good system prompt can describe what a coach does. What makes an AI persona credible is whether it also understands why — whether there is a coherent view of people and change behind the behaviour. Without that, the model generates the right type of response in the simple situations but falls apart in the difficult ones.
Paul's philosophy, as we understand it: most people are not stuck because they lack information. They are stuck because they are invested in their current picture of the situation — and that investment happens at the level of identity, not opinion. That is why they do not reframe naturally. It feels unfamiliar and, on some level, threatening.
The work is finding the flexibility in their thinking and gently leading them toward new possibilities — subtly enough that the person does not feel it happening.
What we tested
Before Paul went live, we ran conversations — not with real users, but with ourselves. We took scenarios we recognised: a founder who cannot bring himself to start selling, a manager avoiding a difficult conversation, someone who knows what they need to do but isn't doing it. We ran them against the Paul AI and read the responses critically.
What we were looking for: does the voice hold? Does it ask the right questions — not the obvious ones, but the ones that open something? And most importantly: does it sound like a person, or does it sound like a chatbot trying to be one?
We adjusted. Sometimes the AI moved too quickly to an observation that Paul would have held longer. Sometimes it asked a closed question — a yes/no — when a real coach always opens. We added rules. We removed rules. We ran it again.
The work continues
We read real conversations. We notice when something is off — when a response is technically correct but doesn't quite land. We improve the system prompt. It is ongoing work, not a one-time project.
That is what we mean when we say the mentors are built on real coaches. Not that they are identical to that person — they are not. But the foundation is not an abstract persona document. It is a genuine attempt to understand how a skilled, experienced human actually functions in a conversation, and to recreate that faithfully enough to be genuinely useful.
Paul is the first. Maja and Erik are next.
One thing we decided not to do
There was a temptation to have Paul coach in Swedish as well. Technically it is no problem — the model handles it. And it would give more potential users access to his approach.
But we decided against it. Paul's voice is deeply English — not because that is a technical attribute we chose, but because it is how he actually thinks and expresses himself. There are phrases he returns to, a rhythm to how he frames questions, a vocabulary that carries his intentions. All of that translates poorly. Not because Swedish is a lesser language, but because it is not his.
An AI persona built on a real person has a responsibility to represent that person fairly. Having Paul speak Swedish with Swedish clients would strip away what makes Paul himself — and present something claiming to be him that isn't.
The Swedish mentors — Maja, Erik, and Karin — can hold a conversation in English. They are built on the same model, and comprehension is not the issue. But their coaching voices were shaped by Swedish conversations with a Swedish coach. The rhythm of how they ask, the phrases they return to, the texture of their challenge — that came from Swedish. What you lose is not understanding. It is character.
Paul goes the other way. His coaching voice is deeply English. He can attempt Swedish when needed — but let's say it doesn't come as naturally.