Most AI strategies are elaborate ways to avoid deciding

I once built a machine whose entire job was to decide. AIDA read a hospital’s blood bank — the units on the shelf, the demand across the ward, the supply two facilities away — and it made a call about red cells. Transfuse or hold. Move the units or leave them. On the hospital data I ran it against, there was no neutral position: a decision was made, or a decision was made by default, and the default cost wasted product and cold-chain vans idling between facilities. Another doctor and I built it from nothing, and building it taught me the shape of a real decision. It is irreversible. It is owned. It closes a door.

Most of what passes for AI strategy has the opposite shape. It is built to keep every option open.

The tells are familiar because they are respectable. A roadmap with four horizons. A steering committee that meets fortnightly. A vendor bake-off with a scoring rubric, a shortlist, and a second shortlist. A proof of concept that proves the concept, and then a second one to be sure. None of this is stupid. Much of it is skilled and expensive, and that is exactly what makes it dangerous, because work this fluent is easy to mistake for progress. The working group is not lying when it says it is busy. It is simply busy in the one direction that asks no one to be wrong.

Here is the difference. Activity accumulates. A decision subtracts. After a good workshop you have more options than you walked in with; after a real decision you have fewer, and one of them is now load-bearing. The comfort of a roadmap is that everything on it is still possible. Nothing has been given up.

There are only three decisions that actually bind, and every organisation knows which they are even when it will not name them. Architecture: what becomes true of this system that you cannot cheaply undo. Economics: where the money is made, and what you will refuse to spend it on. Sequence: what ships first, which is also a ruling on what does not. Everything else is elaboration. You can run a two-year programme that touches all three and settles none of them. Many do. The people who run them are promoted for it.

This is why the middle of my Modus is a room and not a document. Decide means the people who can actually say yes are in one place at one time. The three choices go on the wall. They are argued to a conclusion, written in plain sentences a stranger could act on, and signed. The signature is not ceremony. It is the moment the door closes and the options collapse into a commitment with someone’s name on it. Most strategy decks are unsigned on purpose. An unsigned plan can always be revised into never happening.

A model that decides about blood cannot hedge. It cannot open a second proof of concept while the shelf empties. I built it to commit, because the alternative — no call — was itself a call, and usually a worse one. The organisation on its fifth roadmap has made the opposite choice and hidden it from itself: that not choosing is the safe one.

Deciding is the only act that can be wrong, which is why it is the expensive one. The roadmap protects you from being wrong. It cannot make you right. Only the signature can.