By Saleem S. Khatri, CEO of Lavu and Founder of Marty - 4.28.2026
The restaurant industry has an AI-washing problem. Vendors are slapping “AI-powered” onto tools that haven’t materially changed in years, and the hype cycle is producing more confusion than value. Most early AI pilots across industries have failed to deliver measurable ROI, and restaurants are no exception. The skeptics aren’t wrong.
But here’s where I part ways with the doom narrative: the answer isn’t to reject AI. It’s to get better at telling the real thing from the fake. And to do that, you have to start with the problem AI is supposed to solve in the first place.
Restaurants Don’t Die from One Big Thing. They Die from a Thousand Papercuts.
A restaurant doesn’t collapse because of a single catastrophic event. It bleeds out slowly. A few hundred dollars in labor drift on Tuesday. A void pattern nobody catches on Thursday. An 86’d item running three days that silently kills $860 in upsells. A vendor charging 31 percent more than the supplier next door for the same product. None of these show up as a five-alarm fire. They show up as margin erosion that compounds week after week until someone looks at the P&L and wonders what happened.
Here’s a scene that plays out every week somewhere: a location manager schedules for a 140-cover lunch. Rain hits. Covers drop to 72. Labor stays. No one notices until payroll closes. That’s not a crisis. It’s a papercut. But stack enough of them across enough locations and you’re looking at a margin problem that no menu price increase can fix.
The reason these papercuts go undetected isn’t that operators don’t care. It’s that the systems they rely on were never designed to talk to each other. Your POS sees revenue but misses who made the food and how long it took. Your labor platform tracks hours but can’t connect staffing to what actually sold during those shifts. Your accounting system tracks vendor payments but has no idea what got comped, voided, or wasted. Each one tells a partial truth. The cash you’re losing lives in the gaps between them, and no single system can see it.
This is the fundamental problem. And it’s why AI in restaurants needs to be a painkiller, not a vitamin.
What a Painkiller Looks Like
A vitamin is nice to have. It promises general wellness over time. You take it and hope it’s doing something. A painkiller solves a specific problem you’re feeling right now. In restaurant tech, the vitamins are the products that offer “better visibility” and “more insight” and “AI-powered analytics.” They sound good in a demo. They don’t stop the bleeding.
A painkiller is an intelligence layer that sits on top of every system you already run, reads all of them simultaneously, and tells you where your cash is hiding. Not in theory. In exact dollars, by cause, by location, by the person responsible, with one specific action to take. It’s neutral. It doesn’t care about politics or tenure or who’s friends with whom on the line. It parses signals from data your team can’t cross-reference manually because the volume is too high and the systems are too fragmented.
When it works, it does three things.
First, it saves you time and makes you money. Not eventually. Immediately. Within 48 hours of connecting to your existing data, it should be surfacing recoverable dollars that were invisible yesterday. If it can’t, there’s nothing to find, or the product doesn’t work. Either way, you know fast.
Second, it reduces cognitive load on the people who need it most. Restaurant operators lead with their hearts. They got into this business because they love hospitality, not because they love parsing labor analytics at midnight. The right AI takes the analytical burden off their plates so they can focus on the guest, the team, the experience. It doesn’t add another platform to learn, another dashboard to check, another report to run. It delivers the answer and gets out of the way.
Third, it predicts the future. Restaurant data is a treasure trove of patterns that no human team can process in real time. Weather impact on covers. Day-of-week labor curves. Seasonal menu performance. Promo cannibalization. The same data that tells you what happened yesterday, when read by the right system, tells you what’s about to happen tomorrow. That’s not a dashboard feature. That’s a fundamentally different way to operate.
If the AI product you’re evaluating doesn’t do all three, it’s a vitamin. You need the painkiller.
The Checklist: How to Tell If It’s Real
The operators who learn to spot the difference won’t just avoid getting burned. They’ll find the tools that actually protect their margins. Here’s what to look for.
Does it connect to your data, all of it, at once? Real AI reads your POS, labor, inventory, invoices, and delivery platforms simultaneously. Not one feed at a time. All of them, cross-referenced, because the insights live at the intersections. If you can’t ask it a question about your Tuesday lunch rush and get a real answer grounded in your numbers, it’s not AI. It’s a dashboard with a logo.
Does it get smarter over time? On day 90, it should know things about your business that it didn’t know on day one. It should be catching patterns that compound: an 86’d item running undetected for three days across multiple locations, a labor-to-revenue correlation that’s barely better than random scheduling, a void pattern concentrated on a single login that nobody flagged because it was spread across shifts. Ask the vendor to show you something the system learned. If they can’t, the “intelligence” is artificial in the wrong sense of the word.
Does it tell you something you didn’t already know? A chart of yesterday’s sales is reporting. Real AI catches what you missed. Your food cost crept 2 percent this week and here’s the specific vendor pricing spread that caused it. A single generic POS login is processing three quarters of a million dollars a month with every override untraceable. Your “Did Not Like” comp reason was used 111 times in 14 days and the kitchen was never told. If the product never surfaces something that surprises you, what you have is a dashboard in a trench coat.
Does it reduce decisions, not add them? Think about what your morning should look like. You walk in, and before you’ve poured coffee you know: where you’re losing money today, exactly how much, the specific cause, and one action to take. Every recoverable dollar ranked by value, highest first. No logging into three platforms. No cross-referencing spreadsheets. The insight finds you, not the other way around. There’s a meaningful difference between “Here’s a labor analytics dashboard” and “Location 4 has nine staff on a Tuesday lunch projecting 40 percent below Monday. Cut two servers by 11:30 and save $485.” The first one creates work. The second one eliminates it.
Can the vendor explain what they actually built? Did they build a model trained on restaurant operational data, or did they plug into someone else’s general-purpose technology? A generic AI wrapper is not proprietary intelligence. A general-purpose language model doesn’t know what a void spike at 3 PM on Bar 2 means, or why overtime being negatively correlated with revenue at half your locations should alarm you. Ask directly: “What data did you train on? What happens to my product if OpenAI raises prices 10x tomorrow?” If the answer involves the phrase “we leverage cutting-edge LLM technology,” they’re reselling someone else’s product at a markup.
Does it pay for itself, clearly and measurably? There’s a trend right now that should concern every operator: vendors raising prices to cover their own AI investments, then passing that cost to the customer with a vague promise of “better insights.” If an AI product increases your monthly software spend, it must increase your operating margin by more than it costs. That means you should be able to answer concrete questions: How much food cost did this reduce? How much labor waste did it prevent? What’s the measurable ROI per location? If a tool identifies $2.3 million in hidden labor waste across five locations in 48 hours, that’s measurable. If it catches a $33.9 million annual delivery commission structure that was never negotiated, that’s measurable. AI doesn’t need to be cheap. It needs to be accretive. The standard isn’t “Is it affordable?” The standard is “Did it pay for itself?”
The Bottom Line
Most AI pilots aren’t failing because AI doesn’t work for restaurants. They’re failing because most of what’s being sold as AI isn’t AI, and most of what is AI isn’t solving the actual problem. The actual problem is that restaurants are bleeding cash through the gaps between disconnected systems, one papercut at a time, and nobody is connecting the dots.
The operators who demand more will find it. Connect to the real data, all of it. Learn over time. Surface what everyone missed. Deliver the answer, not more data. Explain what’s under the hood. Prove the ROI in dollars. And above all, solve the pain that’s happening right now, not the pain that might matter someday. That’s the difference between a painkiller and a vitamin. And in this margin environment, you can’t afford to swallow another vitamin.
The future of restaurant technology isn’t another system. It’s the intelligence that connects the systems you already run.
Saleem S. Khatri got his start in restaurants scooping ice cream. Today he is CEO of Lavu and Founder of Marty. A serial entrepreneur who has built and sold startups, including Y Combinator-backed companies, he has spent the past decade building restaurant technology platforms now operating in over 75 countries. His team has identified millions in recoverable cash hidden in restaurant operating data. He advises multi-unit operators on turning disconnected systems into margin protection. Connect with him on LinkedIn at or on X at @saleemskhatri.
Are you an industry thought leader with a point of view on restaurant technology that you would like to share with our readers? If so, we invite you to review our editorial guidelines and submit your article for publishing consideration.
