Journal / Kirill Naumov
AI in restaurants: a useful copilot starts with the manager’s questions
A practical guide to using AI in restaurant operations—from sales and menu questions to carefully controlled actions on inventory and staffing.
By Kirill Naumov ·
Restaurant AI does not have to start with a robot waiter or a dramatic transformation plan. A more useful opening is a manager’s everyday question: which menu items changed this week, what should I check before ordering, or what are guests mentioning in reviews? A good assistant can gather the relevant information and show what to inspect next. In a busy restaurant, fewer mystery dashboards and more time to make the next decision is already a respectable ambition.
Toast’s June 2026 report offers one current example of how operators use an AI assistant embedded in a restaurant platform. The company says it analysed anonymised activity from more than 125,000 US restaurant locations using Toast IQ in Q1 2026; sales and revenue, menu and inventory, and guest and marketing were common topics. This is vendor-reported activity from users of one product—not an independent estimate of the whole restaurant industry, and not proof that the product caused better financial results. It is a useful clue about the questions operators are bringing to AI.
Four workflows worth examining
**Sales and shift review.** Ask the assistant to compare sales by daypart, channel, location or menu category, then ask it to show the figures behind the summary. ‘Lunch weakened’ should link to dates, stores and a defined sales measure. Without that, it is a prompt for investigation—not a reason to change the rota or menu.
**Menu and inventory.** AI may flag items whose sales, availability or ingredient use deserve attention. Low sales alone do not mean a dish should disappear; contribution, preparation time, seasonality and its role in the menu matter too. Use a short review list with the underlying data visible, and let an authorised person approve stock or menu changes.
**Guest feedback.** AI can group reviews by topic and draft a reply. It should not invent a resolution, promise compensation or write as if it personally served the guest. Route safety, discrimination, payment disputes and complaints about a named employee directly to a responsible manager. Faster attention is useful; removing a human from the difficult conversation is not the point.
**Staffing and preparation.** Forecasts may help consider demand, labour or prep quantities, but weather, a nearby event, a large booking or a supplier delay can change the picture quickly. Ask the system to show its assumptions. Let the manager adjust the plan, record why and compare forecast with what happened. A visible override is good operational feedback, not a defect to hide.
Draw a line between advice and action
The important design question is not just what a model can answer, but what it may change. For an early pilot, keep it read-only or let it prepare changes for approval: a stock suggestion, a draft reply or a proposed schedule. Only consider direct actions once the team can see who authorised them, what information they used and how a mistake can be undone.
Restaurant data has gaps and local context. A sold-out dish can look like weak demand; a missing clock-in can look like a labour anomaly; a guest review can be sarcastic. AI may speed up the first pass, while the people running the shift interpret what actually happened. If the assistant cannot explain its source or uncertainty, it has not earned permission to act.
A four-week pilot that earns its place
Week one: choose one recurring decision with a named owner. Week two: check that menu names, modifiers, inventory, staff roles and timestamps mean the same thing across systems. Week three: run in review mode and compare the assistant’s suggestions with the manager’s decisions; ask where it saves time and where it misses the shift’s reality. Week four: look at accuracy, time to a decision, staff corrections and the operational or guest outcome you can genuinely measure.
If the pilot cannot show a better decision or useful time saved, do not scale it because the demo was charming. Restaurant AI earns its place when it helps prepare, check or communicate a real decision. The manager still owns the call, the team needs clear escalation rules and the business needs to know what changed. Start with one workflow, reliable-enough information and an honest measure.
Explore the AI for Retail workshop → · Discuss an AI diagnostic
Further reading: Toast’s 90 Days with Toast IQ report. The report describes usage of Toast’s own product; its figures should not be read as industry-wide results.