The word “agent” is being sold as the next great AI tool, often with a heftier price tag attached. In restaurants and retail, though, recent examples show why operators are asking a basic question: what can these systems actually do once they meet the messy details of daily work?
The concerns are not limited to one type of technology. Computer vision, voice ordering, and broader AI platforms have all faced problems in real restaurant settings. Some systems were switched off, some became the focus of legal action, and one company faced charges after statements about its automation did not match how the service worked.
Restaurant AI Has Hit Some Very Real Problems
Starbucks switched off a computer-vision system for counting inventory before it was a year old. That decision does not prove that computer vision cannot help with inventory, but it does show the distance between selling an AI system and keeping it in use.
Pizza Hut faced a more costly dispute. The chain’s biggest franchisees sued for $100 million, claiming a mandatory AI platform broke down across more than 110 restaurants. A system that fails across that many locations is not a small technical nuisance. It can become a business problem for the operators expected to depend on it.
Voice ordering has also produced uncomfortable examples. An earlier version of Taco Bell’s AI voice-ordering system took an order for 18,000 cups of water. The number alone became a vivid reminder that a system can follow the shape of an order without handling the situation in a sensible way.
That gap matters in restaurants because work rarely follows a clean script. The technology may handle a standard exchange, but unusual requests and physical tasks still sit outside the promise of a fully automated experience.
Khara Mangiduyos, owner of Kalei’s Kitchenette, put that limit in plain language: “Who’s gonna scrape that oil? Not a bot. Not AI.”
When Automation Claims Meet Human Work
In January 2025, the SEC charged Presto Automation over statements that its drive-thru voice system had “eliminated the need for human order-taking.” Regulators found that the company leaned on workers in the Philippines and India to complete most orders.
That case adds a different concern to the discussion. The issue was not only whether a voice system could take an order. It was also how much human work remained behind the system and whether customers or operators received a clear picture of that work.
This is where “agent washing” enters the conversation. The label points to the way the word “agent” can make a product sound more capable than the underlying system. Calling a tool an agent does not settle whether it can manage exceptions, complete physical tasks, or operate without workers stepping in.
The restaurant examples also show why price can shape expectations. When a product carries a heftier price tag because it is marketed as an agent, operators may expect a broader range of work from it. A breakdown, a strange order, or hidden human support can then turn a technology purchase into a trust problem.
Jan Snoeckx captured the needed balance in one sentence: “SCP leaders should prepare for an agentic AI future, but they need to separate meaningful capability from market noise.”
A Different Path for Retail AI
Retail is also moving toward systems that need to work across multiple companies. In January 2026, Google introduced the Universal Commerce Protocol at the NRF retail conference. The standard is openly licensed and was built with Shopify, Etsy, Wayfair, Target and Walmart.
That effort presents a different kind of AI story from a single platform breaking down inside a restaurant. A shared, openly licensed standard focuses on how commerce systems connect across businesses. The participating companies include names from online shopping, retail, marketplaces and home goods, which gives the protocol a broad commercial setting.
Still, a standard does not remove the need to test what an AI system can do. It can help systems work together, but the restaurant examples show that connections and labels are only part of the challenge. The tool must also handle the work it promises to handle.
For operators, the practical lesson is straightforward: ask what the AI does, what it cannot do, and when a person takes over. Ask whether the system counts inventory, takes orders, or supports commerce across companies, then look at the failures that happen outside the usual path.
AI agents may become useful tools in restaurants and retail. The strongest case will come from systems that prove their value in the work itself, not from a bigger price tag or a more exciting name.
Based on
- Agent Washing: Why Some Restaurant Operators Are Wary of Overhyped AI — unite.ai
- Cut Through The AI-Wash: How To Tell Whether AI Is Truly Agentic — forbes.com
- Who Pays When Your AI Agent Buys The Wrong Thing? — forbes.com
- Three Common Misconceptions About AI Agents That Are Holding Organizations Back — forbes.com
