Artificial intelligence has largely been a white-collar story.
It drafts emails, summarizes meetings, and crunches spreadsheets. But the first place on Main Street where AI is actively running day-to-day operations is becoming clear.
It’s restaurants.
Across thousands of locations using restaurant technology company Lavu, an AI platform called Marty is already helping operators manage labor, flag fraud, reduce waste, and prioritize what to fix each day. Instead of simply reporting what happened, the system actively guides what to do next — before margins slip.
According to CEO Saleem Khatri, that shift only became possible because the underlying technology finally caught up to the realities of restaurant operations.
“It’s both rules-based and ML, and that’s intentional,” Khatri said. “Pure ML with no guardrails in a restaurant environment would be reckless. You can’t have a model hallucinate a labor recommendation that causes someone to cut staff during a rush.”
Marty blends machine learning trained on years of transaction data with deterministic rules that ensure reliability.
“The foundation is real ML trained on our transaction data, billions of dollars across thousands of restaurant locations over years,” he said. “That gives Marty pattern recognition that no rules engine could replicate.”
But intelligence alone isn’t enough in hospitality.
“The ML is what makes Marty smart. It lets us rank issues by financial impact instead of listing everything that’s out of bounds,” Khatri said. “The rules are what make Marty reliable. In restaurants, reliability beats intelligence every time. You need both.”
An Advantage Most AI Startups Can’t Replicate
The restaurant industry has recently been flooded with AI tools promising optimization. Khatri argues most are missing the one ingredient that matters most: real operational history.
“It is the advantage. Honestly, it’s why we built Marty instead of leaving it to startups,” he said.
“When you’ve processed billions in real transactions across thousands of locations, you don’t need synthetic training data. You have the actual heartbeat of real operations. Every ticket, every modifier, every void, every clock-in, every comp, every discount. Across concepts, dayparts, geographies, seasonalities. That’s institutional knowledge.”
That depth of data lets Marty distinguish between normal behavior and costly problems.
“What that means practically: Marty can tell the difference between a server voiding a steak because the customer sent it back and a server voiding a steak because they’re feeding their friends,” he said. “The patterns are different. The timing is different. The frequency is different. We know that because we’ve seen millions of both scenarios across real restaurants.”
It also means recommendations are grounded in real-world comparisons.
“When we tell a 15-location taco chain that their labor is running high, we’re not comparing them to some theoretical benchmark. We’re comparing them to similar concepts, similar markets, similar revenue profiles,” Khatri said. “You can’t buy 15 years of transaction data. You have to earn it.”
Decisions That Happen Before Service Even Starts
Despite the “real-time AI” label, Marty’s biggest impact happens before the doors open.
“People imagine AI making split-second decisions during service like a self-driving car. That’s not how restaurant operations work,” Khatri said. “The decisions that save money aren’t made in milliseconds. They’re made in the right hour.”
Each morning, operators receive a prioritized briefing that highlights where profits were lost and what to change that day.
“Marty’s primary value is the 18-Hour Alpha, the morning briefing that arrives before service starts,” he said. “We’re analyzing the previous day’s full picture: labor actuals against revenue, void patterns, food cost movements, comp trends.”
During service, the system monitors live data but focuses on lightweight checks designed to surface actionable issues.
“The hard technical problem isn’t latency. It’s relevance,” Khatri said. “Making sure that when we flag something mid-service, it’s actually actionable.”
No Black Boxes, Just Evidence
For small business owners, trust is everything. Marty was designed to show its math.
“Every recommendation comes with the receipt. Non-negotiable,” Khatri said.
“If Marty says ‘Location 5 had a labor overrun of $1,800 yesterday,’ you see exactly why. Scheduled hours versus actual hours by role. Revenue per labor hour compared to your trailing average. Which shifts ran over. Which employees clocked overtime.”
He added, “It’s not ‘our AI detected an anomaly.’ It’s ‘here’s the math, here are the names, here’s what happened.’”
That transparency is central to adoption.
“The only way to earn trust is radical transparency,” he said. “If I can’t explain to a restaurant owner why Marty flagged something in one sentence, we haven’t done our job.”
From Assistant to Autonomy
Today, Marty recommends actions. Over time, it will take on more.
“The long-term vision is full autonomy. But we’re going to earn it in stages, not declare it in a press release,” Khatri said.
He describes the roadmap in phases, starting with surfacing issues and moving toward pre-approved automation like auto-adjusting inventory or staffing.
“Every accurate morning briefing, every legitimate flag, every recommendation that saves money is a deposit in the trust bank,” Khatri said. “We’re not rushing it.”
For now, the results are already visible. Restaurants using the platform are reducing labor overruns, cutting waste, and catching problems earlier — often before managers even realize something is off.
That’s why hospitality is emerging as an unlikely leader in applied AI.
In restaurants, decisions happen hourly, and margins are thin. Software that actively protects profit isn’t a novelty. It’s a necessity.
And increasingly, it’s already running the show.
