The Position

Every few months, the models get better. Not slightly better: materially better, at things that were impossible the quarter before. The price falls at the same time. This is not a temporary condition of an immature market; it is the market. Capability that costs something today will be ambient tomorrow, the way computation itself went from a budget line to a utility nobody itemises.

We call this layer substrate: models, agents, automations, the raw capability anyone can buy, rent or download. Substrate is necessary. Nothing we build works without it. But necessity is not advantage, and the confusion between the two is currently mispricing an entire industry.

An automation is a feature. A feature is a line item, and platforms absorb line items. The email drafter, the meeting summariser, the document extractor: each was a company once, and each became a checkbox in software the customer already pays for. This is not a failure of execution by the people who built them. It is what platforms do, and the absorption is usually finished within a year of the capability becoming reliable.

So the question that matters about any proposed AI work is not whether it can be built. Almost anything can be built. The question is what happens to it when the substrate underneath improves, because it will improve, on someone else’s schedule, for free, for everyone, including whoever competes with you. Work built against a model’s current weakness is scaffolding with an expiry date. When the weakness goes, the scaffolding goes, along with everything that was charged for it.

The buyer’s version of this mistake is subtler than the vendor’s. A tool is demonstrated; it does something impressive; a budget is found. Nobody asks what the demonstration would look like eighteen months later, performed by the platform the organisation already licenses, for nothing. A wrapper on current ability is not an asset. It is a countdown, and the countdown is running on somebody else’s clock.

What endures is a composition: several capabilities arranged around a real process, producing an outcome that was not available before at that cost, speed or scale.

A composition is not a bigger feature. It differs in kind, not in degree. A feature does one thing to one artefact. A composition rearranges how a process runs (who touches what, in which order, with what checks) and its value lives in the arrangement, not in any single part. That is why platforms do not absorb compositions: a platform can see its own artefacts, but it cannot see your process. The arrangement is invisible from outside the organisation it serves.

The clearest example predates this industry by seventy years.

The shipping container was never a better box. Corrugated steel, standard dimensions, corner castings: a competent workshop could make one in a day, and the idea was decades old. Yet the container rebuilt world trade, and the box itself was the least of it.

What actually happened was a re-composition. Ships were redesigned around the cell guide. Cranes replaced gangs of longshoremen. Ports moved from city piers to deepwater terminals with rail behind them. Truck chassis, customs paperwork, insurance and labour agreements were renegotiated around one standardised unit. Any one of those changes, taken alone, was marginal. Arranged together, they cut the cost of moving goods so far that the location of world manufacturing changed.

And the man who executed it was not a shipbuilder. Malcom McLean ran a trucking company. What he understood was freight: where it waited, where it was stolen, where the cost actually accrued, which was overwhelmingly in the handling, not the sailing. Shipbuilders had every technical advantage and no reason to see it. The composition was visible only from inside the problem.

The resistance is instructive as well. The container was fought longest by the people closest to shipping: ports that had just invested in the old piers, lines that had perfected break-bulk loading, unions whose agreements priced the old handling. They were not wrong about their interests; the composition made much of what they had optimised irrelevant. Which is the last property of a real composition, and the one that makes it uncomfortable: it does not improve the existing process. It replaces the frame the process was optimised within. Judged from inside the old frame, it usually looks like a worse box.

That is the uncomfortable part of the doctrine, and the reason we hold it: finding compositions is not an AI skill.

It is knowledge of the problem. How the system works. How the process actually runs, as opposed to how the organisation chart says it runs. What the people inside it do all day, and what they quietly built to survive it: the spreadsheet beside the system, the group chat where coordination really happens, the one person who simply knows. What genuinely breaks, and who pays when it does. None of this is in the model, and none of it responds to prompting. It is learned in the room, slowly, by asking what things cost.

This is not an argument for ignorance of the tooling. You cannot compose capabilities you do not understand; a McLean who had never priced a crane would have got nowhere. It is an argument about sequencing. Understand the outcome first and the mechanism second, and point the mechanism at the outcome, never the reverse. Much of the AI work on the market runs the sequence backwards: a capability in search of a process, demonstrated to an audience instead of measured against a cost.

It means the interesting question about an organisation is not “where could AI go?” The honest answer is everywhere, and it means nothing. The interesting question is: where is a cost already being paid? In lost deals. In errors somebody absorbs quietly. In a person whose holidays are a business risk because the process lives in their head. Where the cost is real it is being paid now, in labour and workarounds, and the composition that removes it can be judged by the oldest test there is: whether it stops being paid.

It also means saying, more often than is commercially comfortable, that AI is the wrong instrument. Much of what is called an AI problem is process debt: a missing definition, two systems never reconciled, an approval step nobody can defend. A rules engine costs an order of magnitude less than we do, and where it is the answer, it is the answer. The practice of finding compositions includes the discipline of reporting when there is none.

The substrate will keep improving whether or not anyone composes with it well. That is what makes it substrate, and that is why it is not the point. The point is what it was pointed at.