AI for restaurants
Covers, no-shows, food cost. Three numbers decide whether a restaurant has a good month, and all three are lost to the same cause: everybody who could fix them is on the floor during service. That is the gap AI for restaurants is genuinely for, whatever the photographs of robot arms suggest.
Service time is the constraint
A restaurant has two hours at lunch and three at dinner when every member of staff is unavailable for anything else. Almost every operational failure traces back to that window.
- The phone rings at 20:30. Nobody can take it. That was a table for six on Saturday. It is the single most expensive unanswered call in any small business, and it is invisible because it leaves no trace.
- No-shows are treated as weather. They are not. A confirmation that actually reaches someone the morning of, and a waitlist that can be offered the table within minutes of a cancellation, moves that number measurably.
- Food cost is a month behind. Supplier prices move, the menu does not. By the time anyone recalculates, a dish has been sold at a loss for six weeks. Nobody is lazy; the arithmetic just needs doing on a day nobody has.
- The review nobody answered. A one-star review, four days old, unanswered, sits at the top of the page a new customer reads. Answering it well is fifteen minutes that never exist.
The four worth building, in that order
- Booking by phone and message, answered during service, writing into the same reservation system the floor already uses. Not a second book.
- Confirmations that reach people, and a waitlist that fires the moment a table frees, ranked by party size against the gap.
- A weekly food cost recalculation from supplier invoices as they arrive, with only the dishes whose margin actually moved surfaced. Not a spreadsheet nobody opens.
- Review drafts in the restaurant's voice — for a human to send or bin, never posted automatically.
What it must never do alone
- Change a price on the menu. It says which dish moved and by how much. The decision belongs to whoever owns the concept.
- Post publicly. A reply to a bad review that goes out unread is how a bad night becomes a bad week.
- Overbook. The rule about how tight you turn tables is the restaurant's, and it has to be written down before anything automates against it.
- Argue with a customer. Anything with a complaint in it stops and goes to a person by name.
Do not build this if
Three cases, and all three are common. If you still take bookings in a paper diary, automate the diary first — there is nothing for any of this to write into, and no amount of clever answering fixes that. If you run one site with a quiet phone and the owner on the floor every service, the owner is the system, and a build will not beat them; come back when there are two sites or a second sitting. And if supplier invoices reach you as photographs of paper in a chat group, that is fixable, but it is the first job rather than the second, and it is cheaper than anything else on this page.
Where the same shapes already run
- Promotion guard — margin defended by a scheduled job rather than by remembering: it keeps products out of discounts nobody agreed to. A menu whose cost moved is the same shape of problem.
- Team briefing bot — the dull discipline that makes the rest work: the day's numbers and the day's list in the team chat every morning, with nobody assembling them.
Questions from the floor
What does AI for restaurants actually do first?
Answers the phone during service and writes the booking into the system you already use. It is the cheapest thing to build and the easiest to measure: count the calls that currently end in voicemail between 20:00 and 22:00 and multiply by an average cover.
Will it stop no-shows?
It reduces them. A confirmation that reaches someone on the day, plus a waitlist that can fill a cancelled table within minutes, is most of the available gain. Anything promising to eliminate no-shows is selling you something.
Does it need to connect to our booking system?
Yes, and if it cannot, it should not be built. A parallel diary is worse than no automation at all, because now the floor has two sources of truth on a Saturday night.
Can it handle the food cost work?
It can do the arithmetic every week instead of every quarter, from the invoices as they land, and show you only the dishes whose margin moved. Deciding what to do about it stays with you — that is a concept decision, not a calculation.
We are a group of four sites. Does that change anything?
It changes the order. With several sites the comparison between them is usually worth more than any single automation: the same dish costing differently across sites is a supplier conversation you cannot have without the numbers side by side.
Two things to check before you spend anything
First, whether your reservation system will let anything else read from it and write to it. If it will not, none of this works, and you are better off knowing that before a conversation than after a proposal. Second, how many calls came in last Saturday between 20:00 and 22:00, and how many were answered. Bring both and the free hour will be enough to settle it.
Book that hour →