For those in the restaurant industry, it seems like AI rollouts have hit a rocky road.
In the Spring, Starbucks switched off a computer-vision system for counting inventory before it was a year old. In that same month, Pizza Hut’s biggest franchisees sued the chain for $100 million, claiming a mandatory AI platform broke down across more than 110 restaurants. Reading these headlines can draw an easy conclusion: AI isn’t ready for this business.
It doesn’t help that the other AI an operator meets is mostly hype. At this year’s National Restaurant Association Show, the floor was riddled with AI claims, including one vendor who was selling a hood-degreasing service that somehow ran on artificial intelligence. Khara Mangiduyos, who owns Kalei’s Kitchenette in San Diego, wasn’t buying it. “Who’s gonna scrape that oil?” she asked a Restaurant Business reporter. “Not a bot. Not AI.”
Both the vendor tactic and the failure headlines build a reasonable double-layered distrust of the technology itself. But if you dig deeper, the AI success currently scaling the industry looks nothing like either the big-name flops or the AI-washed items.
Getting It Wrong Is Part of Shipping. Buying the Wrong Thing Isn’t
The systems that fail most visibly are usually the ambitious ones, aimed at the hardest and least predictable jobs. Often building new technology means putting it into the world, watching it stumble, and fixing it.
Taco Bell is a recent example. The chain recently expanded an AI voice-ordering system to more than 890 drive-thrus, about two years after an earlier version went viral for taking an order of 18,000 cups of water. But these failed rollouts can come with a hefty price tag that mostly enterprise-level businesses are equipped to cover. It’s a risk not everyone can absorb.
Occasionally, the gap between claim and reality is not a stumble but a fabrication. In January 2025, the SEC charged the drive-thru voice company Presto Automation over statements that its system had “eliminated the need for human order-taking.” In fact, regulators found the company leaned on workers in the Philippines and India to complete most orders.
Presto is an extreme case of the same instinct behind the growing problem known as agent washing. In its case, the “AI” was partly people. Agent washing specifically means marketing something as an autonomous AI agent when it isn’t.
The deception is about the level of AI capability. That is the environment Gartner is warning buyers to navigate, in a market where the word “agent” is sold as the best new AI tool out there. Arguably, with a heftier price tag.
“SCP leaders should prepare for an agentic AI future, but they need to separate meaningful capability from market noise,” said Jan Snoeckx, a senior director analyst in Gartner’s supply chain practice. “The priority today is not full autonomy, but building the operational discipline, architectural flexibility and decision frameworks that allow agentic AI to scale as the technology matures.”
The problem is not that ambitious AI sometimes fails; it’s that the failures shape public perception of the entire category, including the modest tools that were never trying to do the hard thing in the first place.
This is where agent washing tips into what’s often called “jumping the shark,” the point where a trend pushes a gimmick so far that the audience stops buying in. When enough vendors slap “AI” on enough products, the label inflates until it lands on something absurd, like a degreaser.
Gartner tells buyers not to take a vendor’s word that a tool is autonomous. A lot of what gets sold as an agent can’t actually work on its own. It follows a script and calls it intelligence.
Instead, use plain automation for repetitive work, save the agents for high-volume jobs where a mistake is cheap, and don’t hand full autonomy hard problems just yet.
The AI That Often Works Isn’t That Glamorous
The tools quietly delivering for restaurants are the ones aimed at a single, bounded job.
Mariano Jurich, a senior product leader at Making Sense who advises mid-market and private-equity-backed companies on which AI to actually buy, says the trouble starts with the word itself in an industry that never agreed on what an “agent” is.
“If you can describe what the tool does on an if-then-do type of sentence, then it’s probably not agentic 100 percent,” he says. That isn’t a knock on the simpler tools. For most of what some teams need, he adds, “agentic is probably an overkill”. If the job is repetitive and rule-bound, then automation handles it without the cost or the risk of turning a machine loose to decide on its own.
The risk exists when introducing an agentic model to manage customer-facing inquiries or financial decisions, without proper human oversight when it counts. It will often improvise, which is the last thing you want anywhere near a customer or a check.
The Stockholm coffee shop that let an AI run the back office is a good picture of what that looks like as part of an experiment. In the spring, the research lab Andon Labs handed a Gemini-powered agent named Mona control of everything but the espresso machine, ordering, permits, hiring, supplier contracts, and let it make its own calls. It bought 3,000 rubber gloves and 6,000 napkins for a tiny café, stocked canned tomatoes nothing on the menu uses, and messaged baristas on Slack after hours. The lab put the mistakes down to the agent’s short memory.
James Tice worked restaurant floors before transitioning to tech, and now, as head of growth at Tab Commerce, a finance platform used by some of the biggest names in the restaurant industry, he spends his time connecting with operators to understand what works and what doesn’t, just yet.
“‘Keep a human in the loop’ has been said so often in relation to AI, in so many industries, that it has started to sound like filler,” says Tice, who believes that many tech companies struggle to understand how operators work.
The AI that clears that bar, he says, is the kind nobody writes about: reconciling an invoice, catching a vendor overcharge, closing the books faster. “Operators care about finances specifically. They need the busywork gone so they can get back on the floor.”
Adding that the efficiency isn’t the goal; it’s what the efficiency buys. “It’s about increasing human time,” Tice says. “It’s not about cutting out the guests or cutting out the experience. It’s about overemphasizing the experience.”
Which is why he thinks Silicon Valley often measures the technology wrong. “There’s no such thing as time savings in a restaurant. There’s just time allocation.”
A tool that hands an operator back thirty minutes has done only half the job if it can’t answer the next question: “What are we effectively taking time away from, and where are we effectively putting time into?” A good one, he adds, “needs to take away $1 time and give back $10 time. But it also needs to find the $10 time, just as it needs to find the $1 time.”
For Tice, all of that reclaimed time points somewhere specific, and it’s the one thing he’s adamant no software should touch. What a guest is buying was never really the food. “You spend 50 bucks on a plate of food, you’re not buying $50 of food, you’re buying $50 of experience,” he says.
He argues, a room where another person serves you only grows more valuable, “the importance of the restaurant still remains the same.” The tools that try to strip the person out of that, in his view, have misread the entire business. “Anytime AI tries to take away from that, it’s says it doesn’t understand our industry.”
Neither Jurich nor Tice are arguing against AI, just understanding which AI belongs where.
“AI is just another tool that you have in your toolbox to solve a problem,” Jurich says. “And that’s why it should be treated as such.”
