Restaurant AI failures: eight public cases and what they teach

The public restaurant AI setbacks of the last three years trace to four gaps: systems that could not cope with real guests, vendor numbers that meant less than buyers assumed, weak security around the AI, and answers nobody checked. McDonald's, Del Taco and Starbucks ended programs. Burger King and Taco Bell changed course.

The eight cases below each come from a company statement, a regulator's order or a major trade or news outlet. We left out viral clips that drew no company response. Each case ends with the check an operator can run before their own launch.

Which restaurant AI projects were paused, ended or drew complaints?

When	Company and system	What happened	Lesson for an operator
Second half of 2023	Del Taco, Presto drive-thru voice ordering	Stopped using the system after testing it from 2022 and announcing a wider rollout early in 2023	A wider rollout can still end within the year. Set a stop rule before you start.
March 2024	New York City's MyCity chatbot for business owners	Told reporters a restaurant could go cash-free and that bosses could take workers' tips; both answers were wrong	Take legal and wage questions to the agency or a lawyer.
June 2024	McDonald's, IBM automated order taking	Ended the test in more than 100 U.S. restaurants and switched it off by July 26, 2024	Test the accents, dialects and odd orders your lane hears.
January 2025	Presto Automation, drive-thru voice	Settled SEC charges that it misled investors about how much of its ordering people did	Ask who touches each order, including people off site.
July 2025	Paradox's Olivia chatbot on McDonald's McHire site	Security researchers got in through an old test account and an API flaw	Ask a vendor how it controls test accounts and passwords.
September 2025	Taco Bell drive-thru voice AI	A prank order for 18,000 water cups went viral; guests also posted about wrong orders and pauses	Cap quantities and send strange orders to a person.
May 2026	Starbucks, AI inventory counting	Ended camera-based counts of milk and other drinks after reports of miscounts and mislabeled products	Run the AI and your old process side by side before switching.
September 2026	Burger King drive-thru voice AI	Said some guests in pilot restaurants drove off rather than talk to a bot	Give guests an easy way to reach a person.

What went wrong at the drive-thru?

McDonald's told franchisees in June 2024 that it would end its automated order-taking test with IBM. Chief restaurant officer Mason Smoot wrote that "there is an opportunity to explore voice ordering solutions more broadly," Restaurant Business reported. The company said the test left it confident voice ordering would be part of its future. CBS News described TikTok videos from 2023 in which the system added nine sweet teas to one order and butter and ketchup to an ice cream order. Two sources told CNBC the system had trouble with different accents and dialects. Our McDonald's AI drive-thru timeline covers what the chain is testing now.

Del Taco began testing Presto's voicebots in 2022 and had decided to stop by the second half of 2023, according to a Presto SEC filing that Restaurant Business reported in February 2024. Presto said the loss would have no material impact on its business.

At Taco Bell, a guest ordered 18,000 water cups for a social media video, and an employee took over the order. In a September 2025 report, Nation's Restaurant News said chief digital and technology officer Dane Mathews was weighing where to use the system, since busier restaurants may do better with a person taking orders. A spokesperson said Taco Bell still sees voice AI as "a core part of our future."

In September 2026, Burger King's vice president of brand technologies, Chakri Somisetti, told a conference that "there are drive offs because guests don't want to talk to a bot." The chain is testing voice AI in 50 to 70 locations and is building a way for guests to choose how they order, Nation's Restaurant News reported. The tracker of chains using AI at the drive-thru lists who is still rolling out.

What did the Presto case show about vendor numbers?

In January 2025 the SEC settled charges against Presto Automation. Presto had claimed its own voice product eliminated human order-taking. The SEC order found that Presto's "non-intervention" rates counted orders finished without restaurant staff, while contracted agents in the Philippines and India still entered many of them. It says a newer version piloted from June 2023 needed a human agent to enter orders about 70% of the time through at least December 2023. Presto neither admitted nor denied the findings.

For a buyer, the lesson is about definitions. Before you compare vendors, ask each one in writing what counts as a completed order, who decides an order was right, and whether anyone outside your restaurant listens to or keys in orders.

What went wrong away from the lane?

Inventory. Starbucks ended its AI counting of milks and other drinks in May 2026 and moved all counts to one process. It said the change "reflects being disciplined about where automation adds value," Nation's Restaurant News reported, citing Reuters reports of miscounts and mislabeled products. Baristas had complained on social media.

Hiring. On June 30, 2025, two researchers reported a flaw in the Olivia hiring chatbot that Paradox runs for McDonald's McHire site. TechCrunch reported they logged in with the password 123456 and said data on up to 64 million applicants was at risk. Paradox said the test account had not been used since 2019, that five candidates' details were viewed, only by the researchers, and that it fixed both problems within a few hours.

Advice. In March 2024 The Markup found New York City's business chatbot telling owners, "Yes, you can make your restaurant cash-free," though a 2020 city law requires businesses to accept cash. It also said employers could take workers' tips. In January 2026 the new mayor's office said it would take the chatbot down.

What should you check before launching restaurant AI?

These checks come from the cases above. Run them on your own menu, staff and busiest hours.

1. Write down what success means. Separate orders the AI finished alone from orders staff fixed, and ask the vendor to report both.
2. Ask who else touches an order. Get a written answer on whether any person off site listens to calls or enters orders.
3. Test your guests' voices. Include the accents, languages, local words and background noise your lane or phone hears on a Friday.
4. Try silly and huge orders. Order 50 drinks, change your mind three times and ask for an item you do not sell. Decide which of these go to a person.
5. Keep a person one step away. Time how long it takes a guest to reach staff, and make sure the crew can switch the AI off in one step.
6. Run it side by side. For counting, scheduling or forecasting tools, keep your old process for several weeks and compare results line by line.
7. Ask about security. Ask how the vendor retires test accounts, whether it runs a bug bounty or other way to report flaws, and what guest or applicant data it keeps.
8. Check any legal or wage answer. Confirm rules on cash, tips, scheduling and allergens with the agency or an adviser, whatever a chatbot says.
9. Set a stop rule. Pick the numbers that would make you pause the trial, and the date you will review them.

The drive-thru pilot worksheet turns these checks into a one-lane trial plan.

How does Maple handle these risks?

Maple sells voice AI for restaurants. Its drive-thru unit stands beside the existing speaker post, lets guests ask for a person at any time and has a one-press switch that returns the lane to the crew. It is in early access with a small number of restaurants and has no public price. For phones, the phone ordering page sets out a trial that counts completed orders, staff handoffs and requests not completed as separate numbers. Ask Maple the same questions on this list.

Published by Maple. This AI-assisted guide combines company statements, an SEC order and reporting from Restaurant Business, Nation's Restaurant News, CNBC, CBS News, TechCrunch and The Markup, all linked above. It reports what those sources said on the dates shown. It does not estimate any company's results or report a test Maple ran.