- The National Restaurant Association put AI use at 26% of operators in February 2026, with only 6% using it for customer orders and 19% for marketing.
- A separate mid year survey of larger operators found 69% using or piloting AI for reporting and analytics, up from about 25% at the start of the year. Both numbers are real and they measure different restaurants.
- The order to adopt in is set by what happens when the machine is wrong, not by what sounds impressive. Reporting first, the phone second, the menu last.
- Allergen wording is not marketing copy. The FDA recognises nine major allergens, and food packed to order at the counter sits outside the packaged food labelling rules, which puts the burden on your staff and your menu rather than on a label.
- Among operators actively using AI, 61% reported lower food costs and 62% lower labour costs. Those are self reported and come from a vendor's own customer base.
- The cheapest first move for an independent is not an ordering bot. It is letting something read the sales data you already collect and never look at.
The phone rings at 7.40pm on a Friday. Two of the four people who could answer it are carrying plates and the third is on the pass. It rings out. Whoever was calling wanted a table for six on Sunday, and they have now called somewhere else.
That is the version of the AI question a restaurant actually faces, and it has almost nothing to do with the robot arm videos. It is about which of the small, repetitive, badly timed jobs in a restaurant can be handed to something that never gets busy, and which ones would be dangerous to hand over at any price.
Where are restaurants actually putting AI?
Mostly behind the counter, and mostly on paperwork. The National Restaurant Association's 2026 State of the Restaurant Industry survey found roughly a quarter of operators using AI tools at all, and the split across tasks is the interesting part. Restaurant Dive's summary of the association's AI adoption figures breaks it down.
| Task | Share of operators using AI | Who the output faces |
|---|---|---|
| Any AI tool at all | 26% | Mixed |
| Marketing, full service | 19% | Customer, reviewable before it ships |
| Marketing, limited service | 15% | Customer, reviewable before it ships |
| Administrative tasks | 10% | Internal only |
| Customer orders | 6% | Customer, live, unreviewable |
Six percent is the number worth pausing on. Voice ordering gets the coverage and the demos, and it is the least adopted thing on the list by a wide margin. Meanwhile 60% of operators told the association they plan to invest more in technology for customer experience, and just under half in back of house technology. The intent is broad. The deployment is narrow and cautious, which is the correct instinct.
Why do two surveys of the same industry disagree this much?
Because they are not surveying the same restaurants. Put the association's 26% next to Restaurant365's mid year report on operator AI use, which found 69% of operators actively using or piloting AI for reporting and analytics, up from around 25% at the start of 2026. Those two figures look irreconcilable until you look at the samples.
The association surveys the industry, which in the United States is overwhelmingly made up of single site independents. Restaurant365 surveyed more than 420 operators representing nearly 10,000 locations. That is an average of well over twenty sites per operator. Multi unit groups with a finance function adopt back office analytics early because they already have the data centralised and someone whose job it is to read it. A single site owner has the same data trapped in a point of sale they log into twice a month.
So the honest reading is this. Adoption is not one curve. It is a fast curve for operators who already had a reporting habit and a slow one for everybody else, and the gap between them is not about AI at all. It is about whether the numbers were being looked at before.
Restaurant365 reports 61% of its active AI users seeing lower food costs and 62% lower labour costs, with 88% reporting weekly time savings. Those come from a software vendor asking its own customers about the category it sells. That does not make them false, and it does mean they are the ceiling rather than the expectation.
The jobs, ranked by what happens when it is wrong
Adoption advice usually ranks tasks by return. Rank them instead by the cost of a mistake and the order changes completely, because in a restaurant the mistakes are not symmetrical. A wrong forecast means you throw away eight portions of fish. A wrong allergen answer means an ambulance.
| Job | What it does | Cost when it is wrong | Who has to check |
|---|---|---|---|
| Sales and labour reporting | Reads your point of sale data, answers questions in plain language | A bad decision you would have made anyway | Nobody, at first |
| Prep and inventory forecasting | Suggests par levels from history plus weather and local events | Waste, or an 8pm eighty six | The head chef, weekly |
| Rota drafting | Proposes shifts against forecast covers | An understaffed Saturday, or overtime | The manager, before it is published |
| Review replies and marketing | Drafts responses and posts | Embarrassment, recoverable | Anyone, before sending |
| Phone answering and bookings | Takes a reservation or a collection order | A lost table, a wrong time, a no show | The booking sheet, daily |
| Menu descriptions and allergens | Writes dish copy from a recipe | A hospital admission and a prosecution | Named person, every single change |
Work down that table rather than across it. The first two rows are where an independent should start, and they are also the two nobody photographs.
Should an AI answer the restaurant phone?
For bookings and collection orders, often yes. For anything involving an allergy, no. A phone agent is one of the few restaurant applications where the economics are unambiguous, because the alternative is not a human answering well, it is nobody answering at all during exactly the hours when the calls are worth the most.
The thing that decides whether it works is not the voice. It is the pause. A caller who hears a gap after they finish speaking assumes they have not been understood and starts again, which is how these systems tangle. We set out the timing budget that separates a usable agent from an irritating one in the piece on why an AI phone agent is judged on the pause, and the numbers there apply to a restaurant phone line exactly as they do to any other.
Two rules make a restaurant phone agent safe. First, it must hand off to a human the moment anyone says the word allergy, and it must do that without asking a follow up question. Second, it must never confirm a booking it has not written to the same sheet the floor staff read. A reservation that exists only in the agent's transcript is worse than a missed call, because the guest is now expecting a table.
Bookings taken by a machine also fail differently from bookings taken by a person. There is no small talk to reveal that the caller is unsure of the date, which nudges the no show rate up. The countermeasures are the ordinary ones we went through in the article on what actually gets a booked customer to turn up, and they matter more when the booking was frictionless to make.
Can a model write your menu descriptions?
It can write the adjectives. It must not write the allergens. Those are two different documents that happen to sit in the same paragraph on a menu, and treating them as one piece of copy is the most common serious mistake in restaurant AI use.
The FDA recognises nine major food allergens, sesame having been added by the FASTER Act with effect from 1 January 2023, and its guidance on food allergies draws a line that restaurant owners often miss. The packaged food labelling rules do not apply to food placed in a wrapper or box after a customer orders at the point of purchase. Food made to order is outside that regime. What fills the gap is your local food safety law, your staff's answers and the words on your menu, which is a heavier burden rather than a lighter one.
A model writing dish copy from a recipe will produce something fluent and confident about a dish it has never tasted, cooked in a kitchen it cannot see, on equipment it does not know is shared. Cross contact is invisible to it. So is the fact that your supplier changed the stock base last month. The workable division is that the model drafts the sensory description, a named human owns every allergen and dietary claim, and the menu carries a date showing when that human last checked it.
The failure mode here is not exotic. It is the same mechanism behind assistants confidently stating policies that were never written, which we took apart in the piece on why an assistant invents a policy. A model asked whether a dish contains dairy will answer. It will not say it does not know, unless it has been built to.
What about the reviews?
Drafting replies is a good use and publishing them unread is not. Restaurants collect more public feedback per pound of revenue than almost any other small business, and the reply is read by future customers rather than by the complainer. What matters is that it sounds like the person who owns the place, which is exactly the quality a general purpose model strips out. The method that survives contact with real reviews is in the article on replying to negative reviews without sounding automated.
How do you tell whether a tool actually saved you anything?
Measure the job before you automate it, for two weeks, on paper. This sounds like an excuse to delay and it is the only thing that separates a real saving from a feeling. Write down how long the rota took to build, how many calls went unanswered between six and nine, how many minutes went into replying to reviews. Then run the tool for four weeks and write the same three numbers down again.
The reason to do it this way is that AI tools shift work rather than remove it, and the shifted work is easy to miss because it lands on a different person at a different time. A rota drafted in two minutes and corrected in twenty five has cost you more than the forty minutes it replaced, and the manager who did the correcting will not mention it because correcting a draft does not feel like a task. The same trap catches AI bookkeeping, which we went through in the piece on where automated bookkeeping goes wrong for a one person business. The output arrives faster and the reconciliation takes longer.
There is one measurement that needs no baseline. Unanswered calls between your busiest hours are recorded by your phone provider whether or not you ever asked. Pull three months of that report before you buy anything. If the number is four calls a week, an agent will not pay for itself no matter how good it is. If it is forty, the calculation is already made and the only remaining question is whether the handoff to a human works when somebody says the word allergy.
What should stay out of the kitchen entirely?
Anything a diner eats on trust. That is a short list and it is worth writing on the wall.
Cook temperatures and holding times are the second entry for a reason that has nothing to do with capability. A model can recite the correct temperature perfectly. The problem is that a printed instruction carries authority in a kitchen, and a generated one carries the same authority with none of the accountability. When an environmental health officer asks who set that holding time, the answer needs to be a person.
Final prices are the third. If the number a customer is quoted comes from a generated string rather than from the same system that charges the card, you will eventually quote one price and take another. That is a consumer protection problem in most jurisdictions and a refund in all of them.
What one restaurant should do in the next month
Start with the data you already pay to collect. Almost every restaurant runs a point of sale that has been recording every item, every hour and every table turn for years, and almost no independent operator reads it beyond the daily total. Pointing a model at last year's sales by day, hour and dish, and asking it plain questions, costs nothing but the export and answers things you have been guessing at. Which starter dies on Tuesdays. What the real gap is between Friday and Saturday covers. Whether the 6pm to 7pm hour actually justifies the staffing you give it.
Then pick one repetitive written job and do it with review. Rota drafts against your own forecast, or the weekly supplier order, or the replies to last week's reviews. Before you automate the rota, be aware that most late shift changes come from bad delivery data rather than from demand. One job, done every week, checked every time, for a month. If it saves an hour a week it is worth keeping. If it creates an hour of checking, it is not, and you have learned that for the price of four weeks rather than an annual contract.
Only then look at the phone, and only if you are genuinely missing calls. Measure it first. Most operators believe they miss more calls than they do, and some miss far more. Your telephone provider can tell you the unanswered count by hour, and that number decides whether this is a real problem or a fashionable one.
What nobody should do is start with the customer facing order flow because it is the visible one. The industry's own adoption figures already say as much: six percent of operators have put AI on customer orders, against a quarter using it somewhere. The people running restaurants have worked out the ordering, and the trade press has not caught up.
The wider economics explain why the pressure to try something is real. The association's 2026 industry forecast puts sales at $1.55 trillion with real growth of 1.3%, more than nine in ten operators citing food, labour, insurance and energy costs as significant challenges, and 42% reporting their restaurant was not profitable. In that environment, an hour a week of a manager's time is not a rounding error. It is also not worth trading for a system that quietly gets an allergen wrong once a year.
If you are also selling online, whether that is collection orders, gift vouchers or a small retail range, the same discipline applies to the storefront. The pages need to state price and availability in a form a machine can read, because the machines are now the ones doing the reading. That is the reasoning behind generating a shop from a description of the business rather than dressing a template: the structured fields exist whether or not the owner knows what they are for.
Rank the jobs by what a mistake costs, not by what a demo looks like. In a restaurant that inverts almost every adoption list published this year.MaShop, 4 September 2026