Forget the Model, Look at the Harness

Every day, without fail, something turns up in my inbox offering to solve my inbound pipeline problem. Sometimes it’s the outbound instead, or the content, or the hiring. Same email, different sender, a hundred versions of the same idea: an AI that will take an entire function off my hands, starting Monday. Not one of the people sending them has the faintest idea what my pipeline actually looks like. That’s the tell — and it’s the same flaw running through almost everything being sold as AI right now.

The pitch skips the part that actually works

It treats the model as the whole system: switch it on, point it at your business, away it goes. That is never how the thing that actually works gets built.

What actually works is a raw model wired into everything else it needs to be useful. Your history. Your conventions. The decisions you already made and don’t want re-litigated every time. The specific facts about your specific business that live nowhere the model has ever read. None of that arrives with the model. All of it has to be built, by someone, for that one business, and kept current as the business changes. The model is the easy ten percent. The other ninety never appears in the sales email, because whoever sent it has never seen the inside of your business and isn’t offering to.

I found this out building my own

I spent six months building something like this for myself, and the ratio held. The model itself, the actual “intelligence,” turned out to be the smallest piece of the whole project (which surprised me more than it probably should have). Nearly everything else was harness: what it should remember and for how long, which of my own rules it should never override, what “finished” looks like for a given piece of work, how it tells me something needs my judgement instead of guessing at one. None of that is intelligence. It’s plumbing. And the plumbing is what turns a clever autocomplete into something you can actually trust with real decisions.

I’m still not sure how much of this generalises past my own setup. But the shape held everywhere I looked, and I looked hard, because the tell was in my own instinct before I’d even named it: before I let the thing touch a single piece of real work, I’d already written down what it was and wasn’t allowed to decide on its own. I hadn’t built a tool. I’d built a set of constraints with a tool attached to them.

So who actually wins?

Not whoever adopts fastest. That’s the bit the “adopt AI or get left behind” version of this gets backwards. Adoption isn’t the scarce thing — any business can switch a model on this afternoon. What’s scarce is the unglamorous, specific, slow work of building the context around it: memory that survives past one conversation, rules that stop it inventing an answer instead of saying it doesn’t know, judgement about what it should never be trusted to decide alone.

You can’t buy that off the shelf with a subscription, and nobody cold-emailing you can sell it to you, because they’d have to understand your business first. It has to be built, for your business specifically, by someone who knows both the tool and the business well enough to see exactly where the line sits. Build it well and you have a real advantage. Skip it and you’ve got a slightly better autocomplete wearing your business’s name to the meeting.

The harness is the job

Everyone’s watching the model, because the model is the part that’s fun to watch. The actual differentiator sits one level down, in the machinery nobody is selling: the memory, the conventions, the guardrails, the judgement calls about what a person still has to make. Get that right and the model underneath becomes almost incidental. Skip it, and you’ve built something confident, fluent, and completely unmoored from what’s actually true about your business.

Companion essay to What a sabbatical to learn AI actually looks like — the AI Sabbatical, post 1.