A venue owner asks the obvious question: everyone says AI can help my restaurant, so why does nothing work? The honest answer is that most projects never had a chance, because they started at the wrong end.
Here is the usual sequence. Someone buys a tool. The tool needs data. The venue's data is a PDF menu, an Excel rota, a recipe folder that stopped being accurate in 2023, and a head chef who knows why the beef takes eleven minutes and not nine. The tool ingests the PDF, produces something confidently wrong, and within three weeks nobody opens it.
The project did not fail at the model. It failed at the assumption that the knowledge was already written down.
Hospitality knowledge lives in people, not documents
Walk into any venue and ask where the standard for the signature dish is written. You will get one of three answers, and only one of them is a document. In most kitchens the real specification is a person: the chef who trained everyone else, who corrects a plate on the way to the pass, who knows the supplier substitution that works and the one that does not.
That knowledge is real, it is valuable, and it is completely invisible to any software you buy. So when a tool promises to answer staff questions, it answers from the documents, which are the thin, out-of-date shadow of what the venue actually knows.
The question is not whether the AI is good. It is whether anything you gave it was true.
The three failures, in order of frequency
Capturing nothing and expecting everything. The venue points a tool at existing files. But the files were never a knowledge base, they were compliance artefacts. Nobody wrote them to be read by staff mid-shift, so nothing useful comes back out.
Capturing text, not structure. A slightly better project scans everything. Now there is a pile of unstructured text. The tool can find a page, but it cannot answer a question that spans two of them, because nothing tells it that this allergen note relates to that recipe which is served on that section.
Structuring for the software, not the shift. The most sophisticated failure. Everything is captured and structured beautifully, and it still dies, because reaching it requires a laptop in an office and the question was asked with two hands full at 20:15 on a Saturday.
Every one of these fails before the model is involved. The model is the last five percent of the problem and it is the only part anyone talks about.
What working backwards looks like
Start from the moment the question gets asked. A commis needs to know something, on the floor, right now. Work backwards from there and the requirements fall out in reverse order: it has to be reachable on a phone, so it has to be answerable, so it has to be structured, so it has to be captured from the person who actually knows.
That last step is the one everyone skips, because it is the only one that cannot be bought. It requires sitting down with the chef and the floor lead and asking the questions that get the knowledge out of their heads, in a form that survives them leaving.
It takes about ninety minutes. It is not glamorous, it involves no models, and it is the entire difference between a project that works and a project that quietly stops being opened.
The uncomfortable implication
If your venue is disorganised, that is not a reason to wait until you are ready. It is the opposite. The venues with the least written down have the most to gain, because the gap between what the operation knows and what is reachable is at its widest.
The venues that struggle are the ones that assume the folder is the knowledge. It never was. The folder is what was left over after the knowledge walked out at the end of the shift.