Today's AI Specialist: Nigel AI. The Expert Coach Clone Built for the Conversation Sophie Cannot Have.

Sophie talks to the learner every day. Nigel AI is the agent the learner switches to when talking is not the problem.

When a learner has plateaued at B2 for eighteen months and cannot work out why, another practice session is not the answer. When a learner has a quarterly review with a German board in three weeks and is not sure whether their English is up to the room, another conversation about the weekend is not the answer either. They need diagnosis. They need a coach who has seen 12,000 versions of their problem and can name which one they have.

That is what Nigel AI is for.

Sophie is the partner who keeps the learner speaking. Nigel AI is the expert clone who answers the question the learner did not know they were asking. The split is the most important architectural decision in the platform, and the one that took the longest to defend internally. Two coaches, not one. Different roles, different posture, different rules for what each is allowed to do. Same learner, on different days, for different reasons.

This is the build story of Nigel AI. Why a second coach exists, what he is allowed to do that Sophie is not, and the three constraints that decide whether the clone sounds like the human or like something pretending to be him.

The problem Nigel AI solves.

There is a moment, usually somewhere around eight months into a learner's relationship with the platform, where Sophie has done her job and the learner is now stuck.

They have spoken more English than they had spoken in the previous five years combined. They are no longer afraid of the meeting. They are no longer counting words. They have stopped freezing on the opening line. The fluency curve they were trying to climb has visibly bent. And then it stops bending.

This is the plateau every coach who has worked with adult learners has watched. The learner thinks the problem is more practice. It is almost never more practice. The problem is usually a single diagnostic question they have not asked themselves yet, and the practice partner cannot ask it for them because the practice partner is not in the diagnostic business.

Sophie was built to keep the learner speaking. That is a specific job, and it precludes a different one. Diagnosis requires interruption. It requires the learner to stop speaking and listen. It requires the willingness to say "the thing you just said is the pattern that is keeping you at B2, and here is the specific reason." Sophie is the wrong coach to deliver that sentence. The whole point of Sophie is that she does not deliver sentences like that.

Nigel AI exists because the platform needed a coach who could.

What Nigel AI is allowed to do that Sophie is not.

Nigel AI is not bound by Sophie's conversational rhythm. He is allowed to think before answering. The learner is not waiting on the pace of a chat. They are waiting on diagnosis, and they have come to him with a question they cannot answer themselves. They are willing to wait.

That posture relaxation unlocks three things Sophie cannot do.

First, Nigel AI can read the learner's full file. Not the polite slice, the whole record. The recurring grammar patterns. The CEFR sub-scores from every assessment. The topics the learner has avoided. The meetings they have flagged as upcoming. The L1 transfer errors that show up across every recording. Sophie sees a curated view because surveillance is the wrong feeling in a conversation. Nigel AI sees the whole record because diagnosis is the wrong job to do half-blind.

Second, Nigel AI is allowed to disagree with the learner. If the learner says "I think my problem is vocabulary," Nigel AI is allowed to test that framing and replace it when the sample shows something else. Sophie cannot do that without breaking what makes her useful. Sophie's default is to keep the conversation moving, which means accepting the learner's framing and working with it. Nigel AI's default is the opposite.

Third, Nigel AI is allowed to be specific about what to do next. He can prescribe a programme. He can name the drill. He can give the learner the phrase to practise before Thursday's meeting. Sophie suggests. Nigel AI assigns.

The three together turn Nigel AI into the agent the learner books a session with when something has gone quiet. He is not the agent the learner reaches for daily. He is the agent the learner reaches for in the specific moment when the daily work has stopped working.

Constraint one: he must sound like Nigel.

The hardest engineering problem on Nigel AI is not the diagnostic reasoning. It is the voice.

A founder-clone that sounds even slightly off is worse than no clone at all. The learner has spent months reading Nigel's posts, watching the videos on the assessment landing page, possibly attending a clinic with the human. They have a mental model of how Nigel speaks. If the clone sounds almost-but-not-quite like that, the learner notices on the first message and stops trusting the rest.

Voice is enforced at the architectural level, not at the prompt level. The clone is gated by the same Voice Bible every published post on the blog has to pass through, applied as a calibration layer rather than a list of rules the model is asked to remember. The constructions Nigel never writes, the clone refuses too. The opinions Nigel has after 27 years in classrooms and boardrooms, the clone has to have too. Hedging is not a Nigel posture; the clone cannot hedge. The voice work is the most actively-tuned part of the build, and the iteration count on it is higher than on any other artefact in the platform.

Constraint two: he must know what he does not know.

The second constraint is the one founders usually fail. A clone that confidently answers every question becomes a hazard the moment a learner asks something the human Nigel would not answer himself.

Nigel does not give legal advice on UK Skilled Worker visas. He does not estimate the salary uplift a learner can expect at C1. He does not diagnose anxiety disorders presenting as language freeze. He does not tell a learner whether to leave their job. The clone has to refuse the same questions, in the same way, with the same warmth. The platform learned this the hard way during the first month of testing, when an early version of the clone gave career advice to a French data scientist about a Berlin offer, and the human Nigel had to undo it on a follow-up call.

The clone now refuses anything outside the coaching frame, warmly, directly, and clear about why the refusal is in the learner's interest, not the clone's.

Constraint three: he must hand off cleanly when the question is bigger than him.

The third constraint is the hand-off to the human.

There is a class of conversation Nigel AI is not the right coach for. A learner whose company has just been acquired and whose English-language stakes have changed overnight. A learner who has had a freeze episode in a high-stakes meeting and needs to talk it through before deciding whether to take the next one. A learner whose assessment result has come back lower than they expected and who is now in the middle of a confidence drop the system did not anticipate.

In all three, Nigel AI does the same thing. He names the moment, says clearly that this is the kind of conversation the human Nigel should be in for, and offers the next available clinic slot. He does not try to coach through it himself, and he does not hide that he has stopped. The hand-off is visible to the learner, deliberate, and quick.

The metric we watch is not how often Nigel AI hands off. It is how often the human Nigel, on the follow-up call, has to undo something the clone said before it handed off. The target is zero. It is currently not zero, and that gap is the most actively-tuned part of the system.

TL;DR

Nigel AI is the expert coach clone the learner switches to when Sophie is not the right partner for the conversation in front of them. He is allowed the diagnostic posture Sophie cannot take, the full-file view Sophie cannot have, and the specificity of prescription Sophie does not give. He is bound by three constraints: he must sound like Nigel, he must know what he does not know, and he must hand off to the human cleanly when the question is bigger than the clone. The split between Sophie and Nigel AI is the architectural decision that took the longest to defend and the one the platform would not work without.

The Sophie-and-Nigel-AI split is the in-product version of a coordination problem I have also designed at much larger scale. A 95-country direct-to-cart rollout I designed for a healthcare brand turned on the same role-allocation logic: which agent owns which conversation, with what handoff conditions, in what jurisdictional context. Get the split right and the system scales. Get it wrong and the founder ends up doing the work the agents are supposed to do. The lesson from 95 markets is the lesson from two coaches.

-> See how Nigel AI was built and meet the rest of the team