If anyone can rent the same intelligence, what is left that a competitor cannot simply buy?
A client told me, with some pride, that his firm had gone all in on AI. Everyone had a Gemini licence, and people were using them for everything — analysing data, drafting replies to customers, working up strategy. From the inside it looked like adoption done right: fast, enthusiastic, across the whole company. He wanted to know what to do next. The honest answer was that, on the part that might one day have been his, he had not yet begun.
Look at what was happening, from the surface down. Nearest the surface, and the thing that should have worried him first, was data walking out of the door: a hundred people pasting whatever they had to hand (numbers, contracts, customer details) into a tool owned by someone else, with no one having decided what may leave the building and what may not. Below that, quieter and more expensive, the model had no idea who he was.
It answers from the average of the Internet, not from anything his firm knew about itself; asked to analyse the company’s own numbers, it returns something plausible and generic, the work of a brilliant consultant who has never read your files. The speed is real, but the answers belong to everyone. And at the bottom, the part only a manager sees, no one was holding any of it together — a dozen people running a hundred prompts, each deciding privately how to ask and what to believe, not one of those decisions recorded, comparable, or owned. Everyone in the firm is now making decisions with a machine.
No one is accountable for any of them.
The three failures share a root. He had treated intelligence as the scarce resource and paid for access to it. But a frontier model is the least scarce input in the chain — on sale to every competitor he has at the same monthly price, and falling. What was scarce, and what he walked straight past, was everything the model could not supply on its own: his data, the context that would have made the answers his rather than anyone’s, the orchestration that would have turned a hundred private experiments into one system, the record that would have made a decision answerable. He bought the commodity and skipped the asset. The firm creates real value: people are faster, more gets done. And keeps none of it, because the same speed is for sale to anyone who signs the same contract. The gain runs straight through to his customers, as in a competitive market it must.
This is not a forecast about where things are heading; the commoditisation is already legible in what firms are spending. When Uber’s chief executive admitted the company had burned through its entire AI budget for the year in a single quarter and would shift most of its work to cheaper models, keeping the frontier for special cases, he was describing the same economics from the buyer’s side.
Research into legal work has found that cheap open-source agents, with an expensive model brought in now and then as an adviser, beat paying frontier prices for everything, at a fraction of the cost. The model layer is settling into a utility. No firm will hold a lasting edge there, any more than it holds one in the electricity it buys.
What is scarce is governed data
If the advantage cannot live in the model, it has to live somewhere the model cannot reach: in what stays scarce because it is yours. Here the usual answer — data — is half right and half a trap. Data is not scarce; every firm is drowning in it, his included, and he owned the data that mattered long before the licence arrived. What is scarce is governed data: data someone has taken the trouble to turn into a decision that is contextual, repeatable and answerable.
The raw material is everywhere; the refining is rare, because it is slow, unglamorous and always possible to defer to next quarter.
Retrieval-augmented generation, RAG, comes to the rescue
The technique has an ugly name — retrieval-augmented generation, RAG — and a simple idea behind it. Before the model answers, the system fetches the relevant passages from your own documents and makes it answer from those, citing them, rather than from whatever it absorbed in training. The model supplies the language; your material supplies the facts and the authority.
This is the difference between asking a clever stranger what your policy probably says and handing them the policy first.
A private hospital had a harder version of that problem: not one policy but hundreds, many of them in conflict. Its staff, under pressure in triage, in the front office, in a billing function tangled up with pre-authorisation, needed fast, exact answers grounded in the hospital’s own rules, in a place where a plausible answer is not an acceptable one. The information all existed, but it contradicted itself: a standing policy overridden by a time-limited minute, a FAQ that oversimplified, a manager’s email proposing a shortcut, a shift note that had hardened into informal practice.
The first attempt, a generic chatbot, did what generic models do — it treated every document as equal and produced a confident synthesis with nothing behind it. What made the system usable was not the model but the governing of the corpus: a hierarchy that settles which source outranks which, answers that cite the document and version they rest on, and a rule that when sources genuinely conflict the system stops and refers the question to a named person rather than guess. Any hospital can rent the same model.
That governed corpus is not for sale. I work through the case in full in a book I have coming out soon.
The European Union’s AI Act
A deadline was meant to make this concrete, and then it slipped. For the systems it treats as high-risk, the European Union’s AI Act demands the very things the licence-buyer skipped: documented lineage for the data a system learns from, automatic logging of what it did across its life, a named human kept in the loop.
Those rules were due to bite in August 2026; a reprieve agreed this spring is set to push them to the end of 2027. Most firms will read the delay as permission to stand the work down. Read the list again — data lineage, a record of every decision, someone answerable for it — and it doubles as a description of the one asset that does not commoditise. Deferring the compliance is deferring the advantage. The regulation is, almost by accident, a map of where the rent sits.
The same question: what a model can explain versus what a manager must understand is the subject of : AI in Management: Why Structural Thinking Still Matters for Better Decisions
