I Built an AI Recovery Companion. Here Are the Ten Reasons It Shouldn’t Exist.

My engineer asked me why anyone would use Sol over ChatGPT. I didn't have the perfect answer that night. So I went away and built one properly.

It was the end of a long day of coding, the kind where you’ve been staring at the same broken thing for six hours and your patience has worn down to nothing. Adam, who builds Sol with me, looked up and asked the question flatly, without any diplomacy left in his voice.

“Why would anyone use Sol anyway, when they can just use ChatGPT?”

He wasn’t being difficult. He was tired, and he was right to ask. If the person who builds the thing every day doesn’t have a settled answer to that question, nobody else is going to find one either.

I had the answer, but I knew it could be more robust. So I went away and built one properly, and sent it to him.

That felt, at the time, like solving a single awkward question from a colleague. It’s taken me longer to admit it’s actually the first of many questions I’ve had to sit with just as squarely - questions that, if I answered them badly, would mean Sol shouldn’t exist at all. This is me showing the working on all of them, starting with Adam’s.

Why not just use ChatGPT?

The honest answer is that generality is the weakness here, not the strength. ChatGPT will do anything you ask, which is exactly why it’s the wrong tool for someone quietly wondering if they drink too much. Ask it “do I drink too much” and it will either reassure you or lecture you - and either way, it loses the person in front of it.

The clearest way I’ve found to explain the difference is a medical library versus a GP who’s known you for years. ChatGPT is the library: every paper, instant, free, more information than any one person could hold. Sol is the GP. For many questions, the library wins. But when something’s actually wrong, and you’re frightened, and you’re not even sure you want to know the answer, you don’t want a library. You want the person who greets you by name, remembers your history, and knows that the gentle question works better on you than the hard fact. The library can’t do the one thing that matters most in that moment. It can’t meet you where you are.

Sol carries a clinical architecture ChatGPT structurally cannot: Motivational Interviewing, FRAMES, Recovery Capital, a design built specifically for someone still ambivalent about their own drinking, who knows the worst move you can make with an ambivalent person is to push. And it carries things a general-purpose model was never built to hold - a UK governance wrapper, a safeguarding layer, memory of this particular person’s journey, and no commercial incentive to maximise how long they stay talking. Sol isn’t trying to keep anyone engaged. He’s trying to help someone get somewhere.

That’s the real answer to Adam’s question. But it opened a door I couldn’t shut again, because if I was willing to interrogate that assumption properly, I owed the same rigour to everything else people might reasonably doubt about what I’m building. Is that down to my autism, or just a need to be incredibly clear, or both? Time will tell.

The trust problem

The polling everyone quotes at me says “only 6%” of UK adults would turn to an AI platform with a drinking concern. It’s real data, and I take it seriously - but it asks whether people would confide in a generic tool like ChatGPT, not a purpose-built companion with no login wall and no clinical gatekeeping. The number that actually matters isn’t how many people trust “AI” in the abstract. It’s how many people currently tell no one at all, which is a much larger figure, and it’s the entire reason Sol exists. Asking “would you trust AI” before someone has met a specific product is like asking “would you trust a website” before Google existed.

There’s a harder version of this doubt, one I don’t get to wave away with statistics: won’t something this frictionless just become a hiding place - the new “no one,” with better manners? It would, if I built Sol to keep people. I’ve deliberately not. Sol is grounded in motivational interviewing, which means its instinct isn’t to hold on, it’s to help someone move - toward a meeting, a GP, a friend, a morning that looks different from the last one. Success isn’t measured by how long someone stays in conversation with Sol. It’s whether they walk away more able to face the people in their life, not less.

The safety problem

This is the one I’ll least apologise for being strict about. Sol has a built in risk framework and escalation logic, running from low-level concern through to hard-stop conditions where the conversation’s normal shape ends entirely and the person is pointed toward real, immediate human help. When professional support is needed, Sol says so plainly and then stays - the referral is a moment inside the conversation, not the end of it, because the standard pattern of detect-risk-then-refer-then-vanish is a liability response dressed up as care, and it’s exactly what causes people to say “yes, I’ll do that” and never follow through.

Scope, by the way, is a discipline, not an afterthought. Sol is currently built for alcohol, and only alcohol. Every time I've been tempted to stretch it toward adjacent territory - other substances, wider mental health, anything outside what the clinical architecture was actually built and tested for - the answer has stayed no, for now. That kind of expansion sits on the long-term roadmap, not the near one, and it stays there until there's a purpose-built framework and a formal clinical partnership behind it, built for that specific thing. I'd rather say that plainly than pretend it's further away than it is.

The workforce problem

I’ve had this one asked directly, and it deserves a direct answer: “is this just a cheaper substitute for keyworkers?” It isn’t, and the reasoning matters more than the denial. Workforce strain - caseload pressure, supervision gaps - is a named risk in the sector’s own standards. Sol isn’t pitched as replacing the people who carry that caseload. It’s pitched as covering what an overstretched, capped workforce structurally cannot provide between sessions: the evenings, the weekends, the gap between referral and first appointment, the long quiet stretch after discharge when relapse risk is highest and formal contact has stopped.

There’s also an honesty asymmetry worth naming, because it isn’t a threat to the keyworker relationship, it’s a different thing entirely. People often perform recovery for workers - they manage impressions, say what keeps them out of trouble. Sol has no power over anyone’s prescription, referral, or housing, so there’s nothing to perform for. That’s not a replacement for the relationship a good keyworker builds over months. It’s a different kind of honesty, available at a different hour.

The proof problem, and the system problem

Is there evidence this actually works? The honest position, not the defensive one: digital brief intervention as a category is well-evidenced and already commissioned across the UK. Large language model delivery specifically is newer - early, not absent, moving fast, with peer-reviewed pilots landing within the last year. Sol’s own testing framework exists to contribute real data to that evidence base rather than wait for someone else to build it. If pushed further, the right answer is to say so plainly, not to over-claim a settled evidence base that doesn’t yet exist.

And why would an already-stretched, already-tooled-up service want another digital product at all? Because the category is already somewhat proven - DrinkCoach did that work, PHE-backed, with a real health economy outcome behind it - and Sol isn’t asking anyone to fund something unproven as a category. It’s extending a category that already has commissioner trust into the part of the journey - before referral, after discharge - that most commissioned tools were never built to reach.

The caveat I keep in my back pocket

Here’s the part I don’t say often enough in public, and probably should.

This entire argument - the clinical architecture, the governance wrapper, the safeguarding layer, everything that separates Sol from the ChatGPT window sitting open in another tab - holds only for as long as it’s genuinely true.

The moat isn't the underlying model. Sol is built on the same Claude models that power general-purpose AI tools, and that was a deliberate choice, not an accident - Anthropic's own approach to safety and its documented behaviour around sensitive, personal conversations made it the right foundation for something handling exactly this kind of disclosure. But the model was always the foundation, never the moat. The moat is everything built on top of it: the motivational interviewing framework, the safeguarding architecture, the pre-contemplation design, the years of testing against real scenarios and real people.

The day any of that stops being real, Adam’s question becomes fair again, and I won’t have an answer. Which is precisely the argument for protecting all of it, properly, for as long as I’m building this at all.