How does AI really create value? And how much of it does your firm actually keep?
I can watch the value being created in my own work, and I have never once measured where it goes. AI lets me deliver consulting that is several times better than it would have been, and faster. Real value, on every engagement. Some of it stays with me: I do more in less time. A good deal goes to the client, who gets a sharper result at a quicker turnaround. And a slice goes to the firms whose models I run, metered by the token. Three recipients, one of them me, and I could not tell you the split. I apply the counterfactual to a client’s marketing spend without mercy. I have never applied it to my own.
That blind spot is the whole problem in miniature. Almost every business case for AI answers one question: will this create value? And treats the answer as the end of the matter. The question a chief executive is actually paid to answer comes next: will the firm keep any of the value it creates? Those are different questions, and the gap between them is where money quietly disappears.
Value, once created, leaks in one of three directions, and a firm with no structural reason to hold it loses it down all three.
It leaks to customers. When a productivity gain is available to you, it is usually available to your competitors within a quarter or two. In any market with more than one serious player, the gain competes away into lower prices or better service, and the buyer walks off with it — the market working as designed, not breaking. This is most of what is happening to my own consulting margin: as every adviser learns to do more in less time, the advantage runs straight through to the client.
It leaks to the supplier. Much of the value of an AI capability accrues to whoever sells the input that is genuinely scarce: the model provider, the cloud, the few firms that own the compute. The firm building on top sits in the thinnest part of the stack — the part easiest to replicate and most exposed.
I learned this at first hand. I built a product, FinModeler, around a wizard that turned a handful of inputs into a professional-grade financial model. For a while it felt close to magic, and it sold — one-day licences, steadily. Then the model provider did something worse than compete with me on price: it built the capability I was selling straight into the spreadsheet, in the shape of Claude for Excel, and the licences stopped selling almost overnight. There was nothing left underneath to charge for.
It leaks to copycats. If the advantage rests on the model alone, and the model is for sale to anyone with a corporate card and an API key, there is no advantage. There is a feature, and features are matched. The gain is real and temporary, which is another way of saying it is the customer’s gain on a delay.
Capture does sometimes hold, and the cases where it does show what it takes. Ant Group used transaction data to issue small-value credit to merchants in close to real time — loans that traditional underwriting treated as too small and too slow to be worth making at all. The technology created a product rather than discounting one. But the reason Ant kept the value is easy to miss behind the cleverness of the model. The model mattered; the advantage came from the structure around it: payment rails the borrower already stood on, transaction data no rival could see, a channel the borrower had no reason to leave. The model was for sale; the position was not.
None of this is peculiar to AI, which is both the reassuring part and the warning. General-purpose technologies create enormous value and let the average firm that adopts them keep very little. Electricity remade the twentieth-century economy, and the durable rents went to a few infrastructure owners and to customers who got cheaper goods — not to the thousands of factories that electrified and competed the savings straight back into their prices. Aviation, as Warren Buffett liked to point out, transformed the world and was a graveyard for the capital poured into it; he once said a far-sighted capitalist at Kitty Hawk would have done his successors a favour by shooting Orville Wright down. The aircraft flew; the investors were ruined. Creation and capture are not the same event, and mistaking one for the other has destroyed more money than most technologies have saved.
The discipline for this already lives in the firm, in a function rarely asked about AI strategy. Finance learned, painfully, to separate what happened from what the decision caused. eBay found that its paid search advertising was destroying value: each dollar spent returned a loss of about sixty cents, because most of the traffic it paid for would have arrived anyway. The spend “worked” on every dashboard and captured nothing, because it created nothing the firm did not already have. The capture question is one turn harder. Assume the AI does create value the firm would not otherwise have had. Now ask what finance asks of everything else: of that value, how much would have stayed with us regardless, how much will competition take, how much will the supplier take, how much can anyone with the same tools take next quarter? What survives those cuts is the part worth funding as a source of profit.
And the cuts are not fixed. My licences did not stop selling because I had mispriced them; they stopped because the boundary between what I sold and what the platform gave away moved, fast and without warning. A capability that captures value this year can be absorbed the next, when the layer beneath it rises into the thing you were charging for. Anyone who tells you the boundaries of capture have settled is guessing. They are still moving, and they are moving towards whoever owns the scarce layer.
Failing the capture test is not a reason to abandon a project. A capability you cannot keep may be one you cannot afford to skip, because the alternative is letting a competitor build it and turn the gain against you, handing the saving to your shared customers as a price you then have to match. That is a real reason to spend. It is a different reason, and it changes everything downstream: you fund the work as table stakes rather than as a profit centre, you size it accordingly, you stop demanding a return it was never going to produce, and you measure it against the cost of falling behind rather than a payback that does not exist.
The projects that build durable advantage are the ones with a structural answer to the capture question — a proprietary problem, a channel customers will not leave, a decision only you know how to make, a signature only you can put your name to. Most of those are not technical assets; they are things the firm already had before the model arrived, refined into something a rival cannot buy off the shelf. That is its own subject, and the one I will take up next. For now the point is narrower: the model is for sale to everyone, so the model is never the answer to where the value stays.
The business cases worth funding twice are the ones that survive the second question. Plenty will not — mine did not — and learning which before the money is committed is most of what strategy is for. As for FinModeler, the tool still works, better than ever; I will simply have to earn from it somewhere the floor cannot rise to meet me.
