1 November 2026
AI in a company does not start with software. It starts with a list.
The sentence I hear most often, when artificial intelligence comes up in a company, is a version of “which software do we buy?”. It is the wrong question, and it arrives too early. A tool without a list is a piece of furniture in a room you have not measured. It fits. Then you walk around it.
Four steps, and the first is a list
I do not sell a package. I listen to how the company works, I mark what repeats, I build an agent or a tool only there, and a person checks the result. Four steps. I call it the custom method so I do not give it a name cleverer than the work. The first step is not technical. It is a list.
The list is not an inventory of wishes. It is the things that happened more than once last week and took someone’s time. A question to the owner. A file that had to be found. A passage copied from a PDF into a sheet. A quote rebuilt with the same lines. A report asked for “like last time”. If you cannot write ten lines of that kind, you are not ready to buy anything. You are ready to watch.
Listening and the map
Listening is for that. Not to confirm the org chart. The org chart says who should do what. The day says who actually does it, and where people stop. I stay long enough to see the difference. Then I map. Mapping is not drawing a pretty process. It is marking the points where the work repeats and eats the time of the people who decide. The rest, for now, stays put. AI placed where there is no repetition is theatre.
Imagine a construction firm of eight people. It is an example, not a client, and I do not have an eight-person site whose accounts I am publishing. The owner wants “something with AI” because everyone is talking about it. If I start from the software, I offer a generic assistant that does not know their estimates, their suppliers, the way they name a variation. If I start from the list, dull and precise items appear: where the latest estimate is, who approved the variation, which PDF is the one that counts, what can be told to the supplier without calling the owner back. Those items can become a tool. The rest of the company, that day, does not need to be reinvented.
Build only on those items
Building comes after, and only on those items. An agent that answers questions already answered. A handoff that does not have to be copied again. A reminder that does not depend on one person’s memory. It is not a catalogue. It is something made for that process. That is why I do not have a single price list: the thirty-minute review is free, and the map, the solution and the ongoing work are on request, because they depend on what the list contains.
Then someone checks
The fourth step is the human check. I have written about it separately, and I do not treat it as optional. A tool that produces an output with nobody looking is not efficiency. It is a new way to be wrong quickly. The person who checks is not a drag on the project. They are the project.
I have a concrete precedent, and it is not an AI case sold to a client. At To Be, between October 2025 and March 2026, I worked as a project manager on photovoltaic and agrivoltaic projects totalling over ten megawatts. Part of the work was writing procedures, ISO-compliant SOPs, and putting digital planning and reporting tools in place. It was not a toy. It was an attempt not to redo the same steps out loud every week. The lesson I keep is this: first make the repeated thing repeatable, then put a machine on it. If you skip the first piece, the machine photographs the mess.
At Atlas too, where I am CEO and co-founder, the point is not a slogan about technology. It is that data, if it arrives scattered, does not become a reliable document on its own. I write about that in the note on paper and PDFs. Here the principle is enough: the list comes before the tool, in an imaginary firm of eight people as in a benefit company that does diagnoses and reports.
There is a simple way to tell whether the list is true. Read it to the people who do the work, not to the people who describe it. If they say “yes, I redo this every Tuesday”, you are there. If they say “it would be nice if…”, you have written a wish. Wishes can be heard. They do not become an agent on Monday morning.
I do not have a percentage of firms that buy software in the wrong order, and I will not invent one. I have the opposite habit, taken from the jobs I have done: the counter, the procedures, the reports. First you write down what comes back. Then you decide whether it is worth taking it out of someone’s hands. The software, when it arrives, has to recognise that list. Not the other way round.