Hospitality Is Not a Data Problem
Where AI belongs in a food business, and the one place it must never go.
I’ve watched AI pitched to restaurant operators as a recipe generator — design your next dish from sales data. I think that’s wrong, and not for the reason people usually give.
The usual objection is that the food won’t be good. That’s a capability claim, and capability claims age badly — models improve. The real argument is different: a great food brand runs on three things — art (the recipe, the taste judgment), science (the SOP, the checklist), and margin (procurement, forecasting, roster hours). AI has a completely different relationship to each one, and most mistakes in this sector come from treating them as a single system.
Margin is arithmetic under uncertainty, dull, and enormous in aggregate — give it to the machine entirely. An SOP is different: a machine can enforce one beautifully, checking that the temperature was logged and the sequence followed, across two hundred outlets at once. But it can’t write one, because an SOP is the fossilised record of a human judgment about what “good” looks like. The authoring is the value. The enforcement is the commodity.
The recipe sits apart from both. Every other input to a food business is purchasable — the SOP can be reverse-engineered, the supply chain is available to your competitor at a similar price, the software is on sale to the person across the street. The recipe isn’t. The accumulated judgment of someone who tasted a dish four hundred times and knew, on the four hundred and first, that it was finished — that’s the one input that isn’t for sale. Which is why a model that gets better at recipe generation makes this worse, not better: it generates a good dish for everyone, and you’ve traded your only non-copyable asset for an efficiency gain your competitor buys next Tuesday from the same vendor.
So the line isn’t front-of-house versus back-of-house. It’s simpler: does the task have taste in it? Wherever it does, a human decides. Wherever it doesn’t, the machine should run unsupervised.
I run a company that sells AI-executed growth, so this rule doesn’t exempt me either. Anyone can generate a thousand emails now. Almost nobody can decide which thousand people should receive them — the bottleneck didn’t disappear, it moved upstream to the one place nobody’s automated. The structure that survives is humans deciding and machines executing, not humans reviewing what the machine already decided. A technology that lowers the cost of a unit of output has little to offer a business whose product is that unit — and a great deal to offer the ninety per cent of the business that isn’t. Put the machine in the ninety per cent. Leave the ten alone. The ten is why anyone comes back.