The test
AI is worth paying for in a restaurant when it removes work you would otherwise pay a person to do, using your own operating data. Menu data entry, cross-referenced reporting, demand forecasting and anomaly detection meet that bar today. Most of what is marketed as AI in restaurant software does not.
Works today
Turning a photo, PDF or existing web menu into a structured menu with categories, items, prices and modifiers. Removes days of data entry, and it is checkable — you can see whether it got the prices right. This is the clearest current win, and not coincidentally it is the single largest barrier to switching POS at all.
Asking "which items sold most last week" instead of building a report. Useful in proportion to how well it is grounded in your data. Ask what happens when it does not know the answer.
Flagging unusual voids, discounts, refunds, comps or waste. Often the highest-value application available, and rarely the one being marketed, because "we watch for things going wrong" sells less well than a chat box.
Predicting covers or item demand from history plus weather and local events. Genuinely useful with enough history. Treat predictions from a system with two weeks of your data as decoration.
Promising, not yet dependable
- Automated ordering from forecasts. The forecasting is usable; letting it place supplier orders unsupervised is not, and the failure mode is expensive.
- Dynamic menu pricing. Technically straightforward, commercially delicate. Guests notice, and remember.
- Labour scheduling. Good at the mathematics, weak on the human constraints that make a rota work.
- Voice ordering. Real deployments exist, mostly at chains with the volume to absorb the error rate.
What to ask before believing anything
- Is it grounded in my data? A model reasoning about restaurants in general is a search engine with better manners.
- What does it do when it is unsure? Silence is a feature.
- Can it act, and can I stop it? Draw the boundary explicitly around anything touching prices, orders or books.
- How much history does it need? If the answer is "none", the forecast is a guess.
- Where does my data go, and is it used for training? In writing.
The unglamorous prerequisite
All of this depends on your operating data being decent. Miscategorised items, inconsistent modifiers, unrecorded waste and cash sales that never reach the books produce confident, fast, wrong answers.
The most valuable AI preparation is not choosing a vendor. It is getting your stock and sales data into a state worth analysing — which pays for itself even if you never buy an AI feature at all.
Where ThaliPOS fits: menu import from a photo or existing online menu is available to pilot restaurants; the question-answering assistant and morning brief are in development and labelled as such. The ThaliPOS AI page →