The chatbot costume.

Last week I planted the flag: the seat is the product. This week, the first thing six months of building taught me. It contradicts most of what this industry is currently being sold.

Every hospitality conference this year has the same demo. An AI assistant answers a guest’s message instantly, charmingly, in any language. The room nods. Somebody says "game changer". Somebody always says "game changer". Everyone goes home.

Here is what the demo never shows. A real guest doesn’t ask a demo question. She asks this: "Can we get a table for four on the terrace this Friday around eight — and does the tasting menu have a gluten-free option?"

One WhatsApp message. Now count what’s required to answer it correctly. Live availability, not yesterday’s. The floor plan, because "the terrace" is a specific set of tables, not a vibe. The current menu, not the PDF from March. Allergen data structured well enough that "gluten-free" returns an answer the kitchen will stand behind. And ideally her history, because if she sat on the terrace twice before, the right answer starts with "your usual spot?"

The language part, the part the conference shows, is now the easy bit. The hard bit is that in most venues this information lives in five places, and three of them are a manager’s memory.

The manager is off on Tuesdays.

I’ll admit this wasn’t the plan. In February I assumed the clever conversational layer would be most of the work. It has been perhaps a tenth of it. The rest has gone into the unglamorous business of making a venue’s own data trustworthy enough that an AI is allowed to open its mouth: menus with allergens structured properly, availability that is genuinely live, guest context that follows the guest from WhatsApp to the website to the front door. (There is a version of this letter where I pretend I knew that from the start. This is not that letter.)

So when an operator tells me a vendor’s AI "went live in a day", I now know exactly what that means. It went live without any of that. It will answer fast, confidently and wrongly — and here the jokes stop, because in this industry a wrong answer about an allergen isn’t a bad user experience. It’s a safety incident with a brand attached.

Which is why the data slog is worth it. Every operator I’ve ever worked with, from quick service to fine dining, London to Dubai, wants the same three things. More bookings tonight. More spend per head from the guests already coming. Fewer of the quiet leaks: the no-shows, the abandoned bookings, the guest who came twice and never came back and nobody noticed.

That’s it. That’s the whole industry. Everything else is administration.

AI done honestly is the first technology I’ve seen that can move all three. Not by chatting, but by answering the terrace question perfectly at 11pm when no host is on shift. By noticing the vanished regular. By turning a review into something the kitchen does differently on Thursday. It earns that role only if the plumbing beneath it is real. Systems of record store what happened. What this industry needs are systems of revenue, built to change what happens next.

That gap is what I’m building into, and Table for One will document it weekly: where restaurants quietly leak money, what the incumbents have stopped noticing, what AI genuinely changes in a dining room and what remains slideware. Including what breaks. Plenty does, and the breakages are usually more instructive than the launches.

Next week: the anatomy of a lost booking, and the five places a guest disappears between deciding to eat out and sitting at your table.

— B