AI Upselling for Restaurants: What Actually Lifts the Ticket

AI upselling is software that suggests an add-on with every order: the raita with the biryani, the second round before the table goes quiet, the fries nobody remembered to offer. Done properly it lifts the average ticket by roughly 5 to 10 percent. You will see 30 percent advertised. That number is borrowed from somewhere else, and this piece will show you exactly where. The useful part isn't the headline figure anyway. It's understanding why the lift happens at all, because that tells you which tools will work in your dining room and which will just sit there.

Key takeaways

• AI upselling suggests a relevant add-on on every order, instead of only when a server remembers.
• A realistic lift is 5 to 10 percent on the average ticket. Treat anything above 15 percent as a claim to check, not a benchmark.
• The widely quoted "up to 30 percent" comes from self-service kiosk research, not from AI.
• The one specific AI figure in circulation, Toast's 6 percent, describes one restaurant in a pilot and measures order volume, not order value.
• It works because it is consistent, not because it is clever. Coverage beats personalisation.
• Frequency is the setting that matters most. An assistant that suggests something every single time trains guests to ignore it.

What AI upselling actually is

Strip the branding and there are three things happening.

A pairing rule. Something has to know that a biryani goes with a raita and a masala dosa goes with a filter coffee. This can be set by hand, learned from your own order history, or both.

A trigger. Something decides when to make the suggestion: as the item is added, at checkout, or partway through the meal when a table is ready for another round.

A channel. The suggestion has to reach the guest. On a server's handheld it becomes a prompt for staff. On the guest's phone it goes straight to the person deciding.

That's it. Whether a large language model picks the pairing or a lookup table does, the commercial mechanism is identical: a relevant suggestion, in front of a guest, at a moment they can act on it.

Keep that in mind, because it explains most of the disappointment operators report with these tools. If a system knows your menu perfectly but only shows the suggestion after payment, it will sell nothing.

Where the "30 percent" number really comes from

This is worth being blunt about, because you are going to be quoted it.

Search for AI upselling and you will find "up to 30 percent" repeated across vendor blogs as though it were an AI result. Follow it back and it lands in two much older places, neither of which is about AI.

The first is self-service kiosk research. Industry trackers put the average ticket increase from kiosks at 15 to 30 percent. That figure compares a kiosk against a counter with no prompt at all, usually in fast food. It measures what happens when a machine asks a question that a queue-pressured cashier never asked.

The second is staff upselling training, and here the numbers do not even agree with each other. Trade press puts well-trained servers at up to 20 percent on check size, while vendor guidance elsewhere quotes ranges as wide as 5 to 30 percent. None of these trace back to a published study. They are assertions that have been repeated until they read like findings, and they describe humans being trained, not software being installed.

So the 30 percent is either kiosk data wearing the wrong label or a trade-press figure with no origin. If your restaurant currently has servers who never suggest anything, some genuine upside exists for you. If your floor already sells well, you are not going to add 30 percent by installing software, and any vendor implying otherwise is quoting a number they have not earned.

The one real AI figure, read properly

There is a specific, checkable AI upselling number in circulation, and it is worth reading carefully because almost everyone repeats it wrong.

On Toast's Q1 2025 earnings call, CEO Aman Narang said that a restaurant in its AI pilot "found that a menu upsell tool increased its average order volume by 6%," as reported by PYMNTS.

Read that again, because two words do a lot of work.

"A restaurant." Singular. One site in a pilot programme, self-reported by the vendor on an earnings call. That is a promising anecdote, not a study.

"Average order volume." Volume, not value. Order volume and average ticket are different measurements, and the distinction matters when you are forecasting revenue.

Now look at how it gets repeated. Multiple pages render this as restaurants, plural, seeing a 6 percent lift in average order value. Both changes push the claim in the same convenient direction. Nobody is necessarily lying; a number gets copied, a word slips, and the copy gets copied.

The honest position is that AI upselling is early, the public evidence is thin, and the most-cited figure is one restaurant's pilot result for a metric people keep misquoting. Anyone presenting this as settled is selling.

Why it works: coverage, not cleverness

Here is the opinion this piece exists to make, and it runs against the marketing.

Vendors sell AI upselling on intelligence. The pitch is personalisation: it learns each guest, it knows their history, it picks the perfect pairing. That is the interesting part technically and it is mostly not where the money comes from.

The money comes from coverage.

Think about your own floor on a Saturday. Your best server suggests the dessert maybe eight times in ten on a quiet Tuesday. During the 9pm crush, with six tables waiting and a kitchen backing up, that drops to two in ten, and the drop happens precisely on the night with the most covers. The suggestion isn't badly targeted. It simply never gets made.

Software does not get busy. It makes the suggestion on the fortieth order of the night exactly as it made it on the first. Going from suggesting on a fraction of orders to suggesting on all of them is a large change in the number of suggestions, and a modest, believable percentage on the ticket. Which is exactly the shape of a 5 to 10 percent lift.

This reframing has a practical payoff. If coverage is the mechanism, then the questions worth asking a vendor change completely. Not "how smart is your model?" but "on what share of orders does a suggestion actually appear, and can I see that number?" The first question is unanswerable. The second has a report behind it or it doesn't.

It is the same logic behind restaurant upselling techniques for human staff, with the failure mode removed. The tactics were never the hard part. Doing them on the four hundredth order of the week was.

What a realistic lift looks like in rupees

Put numbers on it, because percentages hide how modest and how worthwhile this is at the same time.

Take a restaurant doing 1,000 orders a month at a ₹400 average ticket. That is ₹4,00,000 in monthly sales.

The realistic row is a good outcome. It is roughly ₹3.4 lakh a year from the same covers, the same menu, and no new staff. It is worth doing.

The advertised row is the tell. If a piece of software reliably added ₹14 lakh a year to an ordinary restaurant, it would not need a blog post to sell it. Every restaurant in the country would already run one, and the price would not be a monthly subscription.

Our own experience, running this at real restaurants, is a lift of about 5 to 10 percent, and up to 15 percent at our first restaurant. We build one of these, so weigh that accordingly. But we would rather publish the number we actually see than the number that converts better.

For the wider set of levers around this figure, the average order value guide covers pricing, bundles and menu design, and menu engineering covers making the high-margin items easier to choose in the first place.

How to tell real AI upsell from a recommended-items widget

Most tools sold as AI upselling are a "customers also ordered" strip. Some are more. Five questions separate them.

1. Does it know your menu, or a generic category? A pairing engine that suggests garlic bread with a thali has learned from someone else's restaurant.
2. Where does the suggestion appear? In the ordering flow, while the guest is still deciding, or after the order is placed. Only the first one sells.
3. Can it handle a question, not just a click? "Is the paneer tikka spicy?" is the moment before a decision. A system that answers it can suggest with it. A widget cannot.
4. Does it report coverage? Ask for the share of orders where a suggestion was shown, and the share accepted. If the vendor cannot show both, they cannot show you the lift either.
5. Can you turn the frequency down? More on this next, and it is the question most buyers forget.

Our roundup of AI ordering systems works through which Indian tools genuinely do this and which just put the word on a landing page.

The setting nobody talks about: restraint

A system that suggests something on every single order will stop working, and it will take your guest experience with it.

Guests learn fast. If the same prompt appears every time, it becomes furniture, and people scroll past furniture. Worse, at a table where someone is spending carefully, a relentless upsell reads as pressure rather than service. Restaurants run on regulars, and regulars notice.

This is why frequency is a setting worth having. In dineomAI it is a percentage you control, alongside a monthly cap on what the AI is allowed to cost you, so the suggestion stays occasional enough to feel like a recommendation instead of a pop-up. Turning it down usually costs less revenue than operators expect, because a suggestion that lands is worth several that get dismissed.

The rule of thumb from the floor applies unchanged: a good server suggests once, reads the response, and drops it. Software should be configured to do the same.

Where the suggestion actually happens

One structural point, since it decides whether any of this is available to you.

If your billing system records the sale after a server has taken the order on paper, there is no moment for the software to suggest anything. The decision was made at the table, and the POS met it afterwards. This is the difference between a system that records orders and one that takes them, which we pull apart in POS vs ordering system.

AI upselling only exists where the guest and the software meet while the order is still open. That means the guest ordering from their own phone, or a server taking the order on a handheld that prompts them. In dineomAI the guest scans the QR code at the table and orders by chat in their phone browser, in English, Hindi, Kannada, Tamil or Telugu, with nothing to install, and the suggestion happens inside that conversation. For the broader picture of how this category works, start with AI ordering for restaurants.

FAQ

Does AI upselling actually work?

Yes, modestly and reliably. Expect roughly 5 to 10 percent on the average ticket, mainly because the suggestion gets made on every order rather than only when staff remember. The larger figures in circulation come from kiosk research and staff-training data rather than from AI, so treat 30 percent as marketing rather than a target.

How much does AI upselling increase average order value?

A realistic range is 5 to 10 percent. On 1,000 orders a month at a ₹400 average, a 7 percent lift is about ₹28,000 extra per month. The gain scales with how inconsistent your current upselling is: floors that rarely suggest anything see more, well-drilled floors see less.

Is AI upselling annoying for guests?

It can be, if it fires on every order. The fix is frequency control, so suggestions stay occasional, plus relevance, so the pairing makes sense. A suggestion that reads as a recommendation is welcome; the same suggestion repeated every visit becomes noise that guests learn to ignore.

What is the difference between AI upselling and a recommended items list?

A recommended list is static and usually generic. AI upselling picks the pairing from your actual menu and order patterns, places it inside the ordering flow while the guest is still deciding, and can respond to questions. The placement matters as much as the intelligence: a perfect suggestion shown after checkout sells nothing.

Do I need a new POS to use AI upselling?

Not necessarily, but you do need something that meets the guest while the order is still open, either the guest's own phone or a server handheld that prompts. A biller that records the order after it was written on a pad has no moment to make a suggestion, whatever software is attached to it.

What to do next

Before you buy anything, work out your own baseline: take last month's sales, divide by the number of bills, and write the number down. Then ask any vendor a single question, which is what share of orders their system actually shows a suggestion on, and ask to see that report rather than a case study. If they can show coverage, the lift is measurable. If they cannot, you are buying a claim. Book a short dineomai demo if you want to see the suggestion happen inside a real order on your own menu.