Most writing about AI on a kiosk is a list of capabilities. Capabilities are not the constraint. Data is. Three uses survive contact with a real restaurant, each needs a specific kind of history before it does anything, and the other dozen popular ideas fail for reasons that are visible before any code is written.
Three uses, and the data each one needs
The test for whether a use is real is simple: name the data it needs, then check whether the shop has it. A use whose data does not exist is a demo, however good the demo is.
| Use | Data it needs | Shop has it? | Fails by |
|---|---|---|---|
| Add-on recommendation | Item co-occurrence: which items appear in the same order | Yes, if orders are recorded at item level, and after roughly 50 orders per item | Recommending the same three items to everybody |
| Menu translation | A reviewed glossary of dish names and modifiers | No, and this is the blocker | Translating a dish name literally, which reads as nonsense to the customer |
| Anomalous order flagging | Historical labels: which orders were refunded, cancelled or complained about | Partly. Refunds are recorded; complaints usually are not | Flagging large orders, which are the ones the shop wants |
One: recommendation, and why fifty orders per item is the threshold
A recommendation engine works from co-occurrence: this item appears with that one more often than chance would predict. With a menu of forty items and a thousand orders, the average item appears twenty-five times. At that volume the co-occurrence numbers are noise, and the system produces recommendations that are effectively random or, worse, reflect the shop’s own promotions rather than customer behaviour.
The practical threshold is around fifty orders per item before the pairs become stable enough to act on. A shop doing 240 orders a day reaches that in about ten days for a forty-item menu, which is why the recommendation use is the one that works in practice. A shop doing 60 orders a day needs six weeks, and during those six weeks the system should be showing best-sellers rather than pretending to personalise.
What to show while the data accumulates
A plain best-seller list, ordered by margin rather than by volume, outperforms a recommendation model with insufficient history. It is also honest: the customer sees what other people actually order, which is a true statement, rather than a guess presented as a suggestion.
The failure mode worth designing against is sameness. A model trained on a small menu will converge on the three items with the highest co-occurrence and recommend them forever. That is not personalisation, it is a static list with extra steps, and it stops working the moment the customer visits twice.
Two: translation, and why the glossary is the whole project
Translation looks like the easiest use and is the one most likely to embarrass a restaurant. Machine translation handles sentences well and dish names badly, because a dish name is a proper noun with a story attached. A literal translation of a regional dish produces something between confusing and unappetising, and the customer is standing in front of a screen with nobody to ask.
What makes translation work is not a better model, it is a reviewed glossary: every dish name, every modifier and every allergen note, translated once, checked by somebody who has eaten the food, and then used as the only source. With that in place, a translation layer is reliable and cheap. Without it, the output is unpredictable in a way that damages the brand on the first foreign customer.
The order of work is therefore unglamorous. Write the glossary first. It is a day of work for a forty-item menu and it is the difference between a feature and a liability.
Three: flagging the odd order, and the label problem
Flagging is the use with the clearest operational value and the weakest data. The idea is to notice the order that is about to go wrong: a table ordering four mains and no drinks at 8 pm on a Friday, an order with an unusual combination, a repeat customer whose pattern changed.
To learn what wrong looks like, a system needs examples of wrong. Refunds and voids are recorded and can serve as labels. Complaints and remakes usually are not, which means the system learns from the smaller and less important half of the problem. A shop that wants this use should start by recording remakes and complaints at the point they happen, which is a workflow change and takes a month to become a habit.
The second problem is that the obvious signal is the wrong one. Large orders are statistically unusual and commercially desirable, so a naive model flags exactly the orders the shop wants to encourage. Any implementation needs that false positive designed out before it reaches a kitchen screen.
Where the computation runs
A kiosk can do its inference on the device or send it to a server. The choice is not about capability, it is about what happens when the network is down and who holds the data.
| On the device | On a server | |
|---|---|---|
| Latency | Under 100 ms, no round trip | 300 ms to 2 s depending on the link |
| When the network drops | Keeps working | Degrades to no recommendation, or a cached list |
| Data leaving the shop | None | Order contents, which is commercially sensitive |
| Updating the model | Requires a firmware or app update | Immediate |
| Cost | Hardware has to be sized for it | Ongoing inference cost per order |
For recommendation, on-device is usually right: the model is small, the latency matters at the point of decision, and the data does not need to leave. For anything that learns across shops, a server is the only option. The useful question is not which is better but which failure the shop can tolerate during a lunch service.
The board puts a floor under that choice. The Q-series is quoted with a J1900, a Core i3, an i5 or an i7, and at the Celeron end of that list an on-device model is not worth attempting. The K-series goes the other way, with an RK3288 or RK3399 board and Android, which runs a small model comfortably and nothing much larger.
What the shop changes before any of this works
Every use above depends on the order record, and most order records are messier than the owner believes. Four habits are worth building before buying anything, because they cost nothing and they are the actual prerequisite.
- Record sales at item level, not at category level. A till that reports “food, 84 orders” cannot support a recommendation, and the fix is a catalogue rather than a model.
- Keep one name per item. A menu where the same dish appears as three variants of spelling splits its own history into three thin sets, and each one falls below the threshold.
- Record the sold-out moment, not just the sale. Knowing that an item ran out at seven in the evening is the most useful signal a kitchen can have, and it is almost never captured.
- Label the order that went wrong. A remake or a complaint noted at the point it happens is a training example; the same event remembered a week later is a story.
None of these is artificial intelligence, and all of them are what makes artificial intelligence possible. A shop that does the four and never buys a model still ends up with better decisions, which is a reasonable place to stop.
The three that do not work yet
Each of these appears in almost every article about AI kiosks and each fails for a reason that has nothing to do with model quality.
- Voice ordering. A restaurant at peak is a noisy room, and speech recognition degrades in noise in a way that typing does not. The failure is not a wrong order, it is a customer repeating themselves twice and then walking to the counter, which is worse than never offering the feature.
- Dynamic pricing. It works commercially in some categories and fails in food because the customer can see the person next to them paying a different price. The cost is trust, and it is not recoverable with a discount.
- Face-based personalisation. It requires consent that a kiosk cannot practically obtain, it stores biometric data, and in most jurisdictions the compliance burden is larger than the benefit. The recommendation can be built from the loyalty account instead, with no camera at all.
How to tell whether any of it is working
Three measurements, and the third one is the one that matters.
| Measure | What it tells you | Where it misleads |
|---|---|---|
| Acceptance rate of the recommendation | Whether the suggestion is relevant | Rises when the suggestion is cheap and obvious, so it rewards timidity |
| Basket size on kiosk orders versus till orders | The overall effect | Confounded by who chooses the kiosk in the first place |
| Repeat visits by the same customer | Whether the experience is worth coming back for | Slow to move, and affected by everything else in the shop |
A recommendation system that raises basket size by six percent and reduces repeat visits by two percent has lost money, and only the third measurement shows it. That is why the sensible deployment order is recommendation first, on a small scale, with the repeat measurement running from day one rather than added later.
FAQ
Do I need AI on a kiosk at all?
No. A well-organised menu with photographs and a sensible default increases basket size without any model. AI adds value when there are enough orders to learn from and a menu where the right add-on depends on what the customer already chose.
How much order history is enough?
Around fifty orders per item before co-occurrence stops being noise. Below that, show a best-seller list ordered by margin, which is honest and works.
Is on-device inference really necessary?
For a small recommendation model it is the better default, because it keeps working when the network does not and keeps order data in the shop. For anything that aggregates across sites, the server is the only option.
Where to go next
Name the data a use needs before believing the use. Two of the three that work need history the shop already has; the third needs a day of glossary work first.
The throughput and payback arithmetic for a kiosk Installation-day checks and the five recurring kiosk faults Screen size, brightness, payment module and service access · self-ordering, self-checkout or check-in · where AI actually pays in a POS workflow
Specifying the hardware? The hardware side is covered in Cafe POS, station by station — how 48 characters on a ticket change the printer you need.