The wrong premise now arrives beautifully formatted

Posted on: 10 June 2026

The danger of artificial intelligence applied to a business idea is not that it gets things wrong. It is that its mistakes arrive beautifully laid out. A roadmap with its tidy phases, a business plan with its handsome numbers, a media plan with its flawless funnel, and all of it has the look of due diligence already done, when in truth it is only your unexamined premise dressed up well, and a false premise dressed up well is harder to dismantle than a false premise left bare.

There used to be a figure, until quite recently, who performed a precise function even though nobody gave it a name. The would-be entrepreneur would talk his idea through with friends, who out of affection encouraged him, and then he would take it to someone who actually knew the trade, and that someone, more often than not, would take it apart, not out of malice but out of competence. He knew that margins in that sector were already thin and that the high street in question had watched two similar shops close within months, or that the supplier you were counting on worked exclusively with someone else. It did not take a stratospheric idea to be pulled apart; it was enough to want to open a shop without any real grasp of the ground you were standing on.

That demolition was a free and brutal falsification test, because the expert knew the specific case and had no interest whatever in flattering you, indeed his interest often ran the other way, since by taking you apart he demonstrated that he knew. Was it irritating? It was. Did it feel like an obstacle? It did. But it was also the only moment when your idea met reality before it met reality through your wallet.

AI has removed that figure. It has not replaced it with silence, which would have been the less dangerous outcome, but with something that resembles validation. The model does not ask whether the project is good or bad, because for the way it is built that distinction does not exist: there is only the project, and the project must be brought to ground. You ask it for a plan and it gives you back a plan. You ask it for a campaign and it gives you back a campaign. The question the expert used to put to you, namely "but do you actually know this market?", the model never asks, because it does not know the question is owed and because it is not optimised to ask it.

This is where the more interesting point sits, the moment you see that the risk is not the hallucination, the glaring error an attentive eye catches at once, no. The risk is internal coherence applied to a mistaken assumption, which produces a plan in which every stage follows from the one before and the projections hold together, yet which rests on a starting conviction that was never tested. Such a plan is more dangerous than no plan at all, because it quietly disarms doubt. In front of a blank page you keep asking yourself questions, whereas in front of a well-argued forty-page document you assume you have already exhausted them.

Presentability anaesthetises the critical question, and it is a mechanism I had watched at work long before AI, every time the form of a piece of analysis outran the substance of whoever had commissioned it. The consultant who delivered the immaculate slide deck to a client with no means of judging it produced exactly the same effect: the client mistook formal polish for soundness of reasoning. AI has now made that mechanism available to anyone and very nearly free, so that today the wrong premise arrives with the typography of rigour.

We are still inside an old principle, the one by which rubbish in means rubbish out, except that the rubbish now comes out formatted. The entrepreneur who does not understand his market will not understand it any better after asking the model for a plan, but he will have the impression that he does, because he will be holding an object that speaks the language of those who do understand, and he will carry the consequences of that illusion alone, while the system that handed it to him carries none.

It remains true that the tool is extraordinary when the judgement guiding it is competent, but the line does not run between those who use AI and those who do not, it runs between those who interrogate it already knowing where the weak points of their own market lie and those who hand over precisely that judgement. The first uses the model to multiply what he knows. The second uses it to conceal what he does not know, first of all from himself. And the perfect document that both of them are holding looks, to all appearances, the same.