We sell an AI product, so read this with the appropriate suspicion. But we spend a lot of time talking to operators who have been pitched five of these in a year, and the confusion is real, so it is worth writing down what separates the useful from the decorative.
Things that genuinely work
Demand forecasting. Predicting how busy you will be is a well-understood problem with decades of research behind it, and the inputs — your own history, the weather, the local calendar — are all available. This is the one that changes what you do tomorrow morning, which is why it is the part we build.
Turning a photo of an invoice into data. Boring, unglamorous, and it saves real hours every week. If your bookkeeping still involves typing numbers off paper, this is the first thing to buy.
Drafting replies to reviews. Not a revolution, but it is fifteen minutes a day back and the output is honestly fine.
Finding the anomaly in your own numbers. Not a chat interface — an alert. "This Thursday is going to be nothing like your recent Thursdays" is a useful sentence. "Here is a dashboard" is not.
Things that are mostly marketing
"Ask your data anything." A chat box over your sales figures demos beautifully and gets used twice. The problem was never that you could not query your data. It was that you did not have time to look at it, and a chat box does not fix that — it just moves the effort.
Item-level prediction for a small menu. Forecasting how many of each dish you will sell sounds better than forecasting how busy you will be. In practice, in a forty-seat restaurant with a menu that changes, it needs a level of recipe maintenance nobody sustains past month two. Bakeries are different — there it genuinely works, which is why the German companies doing it well are all in bakery.
Anything that says "up to 97% accurate." More on this below.
Fully automatic ordering, for a restaurant. Works in retail and bakery, where the products are stable and the supplier list is short. In a restaurant with a changing menu and four suppliers who each have their own quirks, the automation is more work to supervise than the task it replaced.
How to test an accuracy claim in one question
Most accuracy numbers in this market are unfalsifiable, and the way to find out is to ask this:
"Can I see yesterday's forecast next to yesterday's actual takings, in the product, every morning?"
If yes, the number is checkable and you will know inside two weeks whether it is true for your venue. If it only ever appears on a website, it is a marketing figure and you have no way to test it.
The second question worth asking: what is it being compared to? "Up to 97% accurate" usually means the average error across a year, most of which is Tuesdays — and Tuesdays barely move, so almost anything scores well on them. The days worth measuring are the exceptional ones. Ask how it does on the first warm Saturday of the year, or the day before a public holiday.
We say up to 96%, and we show you yesterday's score every morning specifically so you can catch us out. That is not modesty, it is the only way an accuracy claim means anything.
What actually decides whether it helps you
Three things, and none of them is the model.
Does it reach you? The best forecast in the world is worthless in a dashboard you open twice a month. On a Friday night, if it is not a message on your phone, it does not exist.
Does it change a decision? If you read it and do nothing differently, it is entertainment. Prep quantities, the order, the rota — those are decisions. "Insights" are not.
Can you check it? Anything that cannot be wrong in a way you would notice is not making a claim.
The honest summary
For a single-site independent, most AI in hospitality right now is either solving a problem you do not have or solving one you already solved with a spreadsheet. The exceptions are the boring ones: getting paper into your accounts, and knowing how busy you are going to be.
That second one is worth real money, because everything downstream of it — prep, ordering, staffing, waste — is currently a guess made by a tired person at nine in the evening. But buy it on evidence you can check in a fortnight, not on a percentage on a homepage. Including ours.