The System of Record in Your Head
The build did not stall on the technology. It stalled on a sentence nobody could write.
Ask an owner why the AI project stalled and you will hear about the technology. The model wasn’t accurate enough. The tool never understood the business. The vendor promised more than the software could do. I have quoted the failure statistics myself, and they are doing honest work as a warning. But in an owner-managed business the diagnosis is usually wrong — the model was rarely the thing that failed.
The record nobody made
Every business keeps a system of record. The money has one; that is what the ledger is. The customers have one, even if it is a spreadsheet everyone swears at. But the thing that actually runs the business — how it decides — has a system of record too. What a good quote looks like. When a job is refused. Which customer’s complaint moves the schedule and which waits until Friday. The record of all that is the owner’s head. Nobody wrote it down. Nobody needed to — the owner was always in the building.
There is a reason it never got written, and it is not laziness. Expertise buries its own workings. The better you get at a decision, the less you can say about how you make it; twenty years in, the steps have become invisible to the person performing them — the expert’s curse. The philosopher Michael Polanyi called it knowing more than we can say. I call it the reason the room goes quiet when someone finally asks how the price actually gets set.
Where the build actually stalls
Now put an AI system into that business. The systems we build earn their accuracy through constraint: every step passes a gate — a written statement of what a right answer looks like, set before the system runs. Every likely failure has its correction decided in advance, and the moments that need a person are named up front as stop conditions rather than discovered later as surprises. None of that is exotic. It is the business’s own rules, made explicit enough to be checked.
Which is where the build meets the problem. A gate needs a written expected result — and you cannot write one for a process nobody can state. A system cannot be checked against a standard that exists only in one head. So the build stalls at the first honest attempt to say what the technology is supposed to do. And the model takes the blame, because the model cannot argue back.
Vin Vashishta made a neighbouring point about enterprise AI recently: the bottleneck sits in the gap between the business and the people building for it, not in the model. At the scale of an owner-managed business the translation is blunter. Both sides of that gap are usually one person — and the gap runs between what the owner knows and what the owner can say.
Why the wrong diagnosis survives
Everyone in the room prefers it. For the vendor, a model that failed can be replaced by next quarter’s model on next quarter’s licence — the diagnosis renews the contract. For the owner, “the technology isn’t ready” postpones a decision, while “nobody here can say how we decide” sounds like an accusation. And the curse does the rest: from inside, the process does not feel unwritten. It feels obvious. Every expert experiences their own judgement as common sense — which is why nobody notices the record was never made until something that is not a person needs to read it.
The inversion
Here is what took me longest to accept in my own build. I run my practice on written routes — every system I rely on was forced through the same gates I have just described. Writing them was the humbling part. The first draft of every route exposed a rule I thought I had and didn’t: two cases I treated differently for a reason I could not state, a threshold that moved depending on the day. The agent, once the route existed, was close to the easy bit.
That order holds. The forced writing-down is the asset. The agent is almost a by-product. What the writing produces is worth more than what it displaces: decisions that have survived being looked at; the discovery that the same question gets three different answers on the same corridor; a business that no longer needs one particular head in the building to know what good looks like.
Instinct — the judgement that catches a confidently wrong number — has to be protected and deliberately grown. That is true, and so is this, and the line between them is precisely what the writing-down finds. Most of what a business calls judgement turns out to be procedure nobody ever wrote down. What is left when the writing stops — the part that will not go onto paper — is the instinct that deserves the name. Until you write, you cannot tell which is which. You are protecting a mixture, and paying instinct’s wages for procedure’s work.
None of this needs a vendor. It starts with one process, a blank page, and a question: what should happen here, and how would we know it was right? Write until the answers stop changing. If AI arrives later, it arrives with something to hold on to. If it never arrives, the most important process in the business is finally out of one head.
What the head was
The failure statistics will keep circulating and the models will keep improving, and neither has much to do with whether your build works. A system can only be as explicit as the business it serves. The head was never a bad system of record — it just doesn’t scale, and it takes weekends. If the writing-down is work you would rather not start alone, you know where to find me.