1The Wolf arrives in person
There is a body in the car, the car is in a friend’s garage in the suburbs, and the friend’s wife is home from the night shift in an hour. Two men in blood-soaked suits are arguing about whose fault it is. Their boss makes one call, and the man he calls says he is thirty minutes away and will be there in ten. He arrives in a dinner jacket, introduces himself, “I solve problems,” walks out to the garage, looks at the car for a few seconds, asks who owns the house, asks for coffee, and starts giving orders. Forty minutes later the car is clean, the men are in borrowed clothes, and the Wolf is gone.
The scene is played for laughs and it is the most accurate portrait of a good consultant on film. He was not the best in the world at cleaning cars. He was the person who came to the house, looked at the actual mess with his own eyes, took responsibility for the outcome in front of the people who needed it, and left when it was done. Everything he knew about cleaning cars he could have written down, and it would have been useless to the men in the suits, because the difficulty was never the cleaning. The difficulty was the hour, the wife, the two frightened men and the boss on the phone, and none of that was in the car.
Humans create better outcomes the closer they are to the problem. That sentence sounds almost naive in 2026, when every week brings a new model, a new benchmark and a new claim that the advisor is about to be replaced by an agent, and I hold it more firmly than I did before the models arrived. Closer to the problem is not the same thing as closer to the technology, and the whole of this essay is the distance between the two.
2The moat was never the model
In May 2023 a memo written inside Google leaked and was published under the title it deserved: we have no moat, and neither does OpenAI. Its author had watched open models catch up with the labs’ best work in weeks, on laptops, for the cost of a dinner, and concluded that the models themselves would not be what anyone paid for.1 It was read as a note about foundation models. It was a note about everything downstream of them, and most people who quoted it stopped a paragraph too early.
If the model is not the moat, then the software built on the model is not one either, because anyone can build it. That is the part that has taken three years to sink in, and it is sinking in now through share prices. The tools that let a person describe an application and have it written for them crossed a hundred million dollars of annual revenue within a year of launching, and in the six months to April 2026 the fund that tracks software companies lost more than a quarter of its value while the companies in it reported good earnings.2 We spent two decades turning people into operators of software, one seat at a time, and the thing that made that business a business was that software was hard to make. Now anyone can build what matters to them, and there will be a Cambrian explosion of it, which is not a market but a flood.
A flood changes what is scarce. When nobody could build, building was the scarce thing and the vendor sold it. When anyone can build, the scarce thing is knowing what to build, for whom, in what order, against which of the forty constraints that were not in the request, and being the person who answers for it when the wrong one was chosen. Software stops mattering, and that is the good news, because software was never the point; the outcome was, and the outcome has always belonged to whoever was closest to it.
3Trust is how people survive too much choice
Trust is not a warm feeling. It is a mechanism for coping with a world too complicated to check, and a person extends it to whoever is present and answerable, because that is the cheapest way to reduce a thousand choices to one. Niklas Luhmann wrote that down in 1968, in a book about how any society functions at all given how little any member of it can verify, and it is the best account I know of why the Wolf gets the call.3 You do not trust him because you have audited his method. You trust him because he came to the house, because his name is on the result, and because you can find him tomorrow if the car turns up.
Everything that makes the enterprise hard is of that kind. The chief financial officer who cannot sleep is not kept awake by a technical problem. She is kept awake by a supply chain that breaks every quarter, a reporting line that will not admit it, a regulator who moves the rules in a memo, a process that is theatre and another that everyone thinks is theatre and is the only thing holding the company up. The knowledge of which is which is what an economist once called the knowledge of the particular circumstances of time and place, the kind that never makes it into a report because nobody would put it in one, and the people who hold it are the people who have worked beside it for years.4 A model can be handed the report. It cannot be handed the years.
The industry has been telling those people’s clients that they no longer need them: that agents will replace advisors, that software will eat consulting, and the claim is backwards, because the harder the problem, the more of it lives in context, politics, compliance and the one detail that makes this client unlike every other, and the more it needs a human who can be trusted with that and be found afterwards. A large language model is a commodity, rented by the token from several suppliers at prices that keep falling; bringing one into production in a bank or a hospital is a question of reliability, ownership, and who holds the data and the intellectual property when it goes wrong, and nobody has ever trusted a software vendor very far with any of those, the labs least of all. AI is going to eat software, not people, and the people it will feed are the ones whose clients already trust them.
4Software has to disappear for the person to appear
The strange thing about the most human interface ever built is that it only works when the software behind it goes away. Mark Weiser said in 1991 that the most profound technologies are the ones that disappear, that weave themselves into everyday life until they are indistinguishable from it, and he was describing the opposite of the enterprise application, which announces itself two thousand times a day.5 A model that speaks the language of the person doing the work can make the screen unnecessary, and a screen that is unnecessary is one fewer thing between the consultant and the client.
That is the tension in the market. The vendors want the model inside their screens, because the screen is what they own; the consultant wants the model to remove the screen, because the screen is what stands between her judgement and the client’s problem. We chose the second side, and we chose the hardest environments to do it in, healthcare, compliance, payments and the largest companies on earth, because that is where trust at scale is the whole problem and where the tools that build anything for anyone cannot tell each other apart. Our answer is a single concept, the workblock: human expertise, a model and the business logic of one specific job, packaged so that it is reusable, governed and auditable, and so that it runs against the live systems, writes into the ledger and learns from the transactions. The consultant builds it. The client runs it. What the consultant leaves behind is no longer a deck; it is working software, in the client’s name, that the consultant is answerable for.6
Most software companies charge a licence and then compete with their customers’ advisors for the relationship. We do the opposite, and we charge for the outcome, because an outcome is the one thing that aligns the vendor, the partner and the client on day one, and because a licence is a bet that the client will not notice.7
5Closer to the problem
The Wolf leaves nothing behind but a clean car, two men in T-shirts and a problem that no longer exists, and he gets paid for that and not for the hours. For thirty years the consulting trade has been mocked for leaving PowerPoints, and the mockery was half fair, because a deck is what you leave when the only thing you can build is advice. That constraint is gone. The person who knows the client can now leave the thing the advice described, and the vendors who spent those thirty years selling seats to committees are discovering that they were never near the problem at all.
Boutique or global, specialist or generalist, the kind of firm matters less than the belief it holds. The one worth holding is the one the models have made true rather than obsolete: the humans closest to the problem will outperform the humans furthest from it, no matter how good the model gets, and the model’s job is to make the distance between them disappear.8