What a leader actually buys
Nobody who runs a company wakes up wanting an agent. He wakes up with a quote that was typed twice, a month-end that closes on the twelfth, a customer who waited four days for an answer that took four minutes to write. He wants that to stop, and he wants to be able to explain to his partner why it will.
That is what he buys: a decision he can defend. The system comes second. It is the proof that the decision was right, not the reason for taking it.
The AI market sells the other way round. It sells licences, then use cases, then a workshop to find out what to do with them. A year later the licences are renewed, the use cases are a slide, and the quote is still typed twice.
Why the tool first fails
A tool bought before a decision has three ways to fail, and we have watched all three.
It fails by dispersion. Everyone uses it a little, for something different, and nothing changes in the accounts because nothing was supposed to. Use without a decision is a cost line with a good conscience.
It fails by substitution. The tool does the task that was easy to automate and leaves the one that mattered. The person who used to do both now does the hard half and watches a machine do the easy half, slightly worse.
It fails by abandonment. Nobody owns it. The vendor sells the next version, the integrator has moved to the next client, the person who championed it has changed jobs. Six months later it is a login nobody remembers.
In all three cases the tool was not the problem. The absence of a decision was. A decision names the task, the number it must move, the person who owns it, and the date on which someone looks at the number again. A tool cannot supply any of that.
What a four-page memo changes
So we start with a memo. Four pages, written in ten working days, with the leader present on the first, the fifth and the last.
Page one is the decision: what the company will do, and what it will stop doing. Page two is the numbers: what it costs today, what it should cost after, how we will know. Page three is the risks: what could make this wrong, and what we would see first. Page four is what we do not do: the three tempting things we looked at and set aside, with the reason.
Then, and only then, a prototype. It runs on a real sample of the leader's data during the mandate, and it is shown on day ten. It performs no irreversible action. It exists to answer one question: does the decision hold when it meets the data?
Sometimes it does not. The memo says so, and the mandate ends with a plan for what to do instead. We will write about that separately, because it deserves its own note.
Most of the time it does. And then the leader holds something no licence gives him: a decision his board can read without us in the room, a system he has seen run on his own figures, and a build plan with a scope, a timeframe and a price. He can sign it or put it in a drawer. Either way, he decided.
A generic example, because we never cite a client. A distributor had bought a chat assistant for its sales team the year before. Usage was high, results were invisible. The memo did not recommend a better assistant. It recommended that quotes stop being typed twice, named the two systems that had to talk to each other, and set a single number to watch: hours per week spent re-entering. The prototype moved that number on a sample in four days. The assistant stayed. The decision was elsewhere.
That is the whole method, as much as we will ever say of it. The decision first. The system as proof. The leader in the room when both are made.
ElahPublished on October 21, 2026