McDonald's AI drive-thru fail: why did it end the IBM test?

Quick answer In June 2024, McDonald's told U.S. franchisees it was ending its AI drive-thru voice-ordering test with IBM. The automated order taking system, tested in more than 100 restaurants since 2021, was to be shut off no later than July 26, 2024, after viral videos showed it garbling orders, from bacon on ice cream to hundreds of dollars of nuggets.

What went wrong at the McDonald's AI drive-thru?

You pull up to the speaker. You say your order. A voice, not a person, takes it down. On the screen, your order starts to grow.

That was the scene in a run of videos that went viral online. McDonald's had been testing automated order taking, or AOT, built with IBM, at the drive-thru. And customers were filming it.²

In one TikTok with 30,000 views, a young woman tries again and again to convince the AI she wants a caramel ice cream. It keeps adding stacks of butter to her order instead. In another, with 360,000 views, a customer says her order got mixed up with someone else's and nine orders of tea landed on her bill. In another popular clip, two people laugh as hundreds of dollars' worth of chicken nuggets pile up on their order. The New York Post reported a customer who had bacon added to their ice cream.²

Funny in a video. Less funny if you are the crew fixing every order, or the customer who does not check the screen.

How did McDonald's end the test?

On Thursday, June 13, 2024, McDonald's sent franchisees an email. Restaurant Business obtained a copy and reported it the next day. The test would end without any expansion.¹

After a thoughtful review, McDonald's has decided to end our current partnership with IBM on AOT and the technology will be shut off in all restaurants currently testing it no later than July 26, 2024.

Mason Smoot, chief restaurant officer, McDonald's USA, message to franchisees as reported by Restaurant Business

Smoot also wrote: "While there have been successes to date, we feel there is an opportunity to explore voice ordering solutions more broadly."¹

McDonald's had been testing drive-thru voice AI since 2021, the same year it sold its McD Tech Labs unit to IBM.¹ CNBC reported the test ran in more than 100 restaurants and that the global partnership began in 2021.³

McDonald's did not give up on the idea. It said its work with IBM had given it confidence that "a voice-ordering solution for drive-thru will be part of our restaurants' future," and that it would make "an informed decision on a future voice ordering solution by the end of the year."¹ IBM defended the system as "proven to have some of the most comprehensive capabilities in the industry, fast and accurate in some of the most demanding conditions," and said it would keep working with McDonald's on other projects.¹

Why it happened

Speech is hard, and drive-thru customers do not all sound alike. CNBC reported, citing two sources, that the system had trouble interpreting accents and dialects, which hurt order accuracy. A BTIG analyst estimated accuracy was still "in the low-to-mid 80% range." McDonald's declined to comment on accuracy.³

Put that number in context. At 80 to 85% accuracy, roughly one order in every five to seven would need fixing, which means a person still has to keep an eye on the screen.

The caveats matter here. This was a limited test, in a little over 100 restaurants, and McDonald's said there had been "successes to date." The viral clips show the worst moments, not an average day, and the accuracy figure is an outside analyst's estimate. McDonald's framed the ending as a change of partner, not a verdict against voice AI.¹

But the pattern is familiar. In the clips, the AI does not crash. It keeps going, confidently, while getting the order wrong, and nothing appears to stop a wildly unusual order, like hundreds of dollars of nuggets at a single car, from piling up without a second look.

What this means for your business

You may not run a drive-thru. But if an AI agent takes orders, answers customers on WhatsApp or edits carts and invoices for you, the same failure shows up: a confident system that is right most of the time and wrong in public.

• Sanity-check the unusual. An order ten times the normal size, a discount nobody approved, a refund on a fresh order. Those should pause for a person, not go straight through.
• Let customers reach a human fast. The frustration in those videos came from customers stuck correcting a machine. Make the handoff easy.
• Measure accuracy yourself. Vendor claims and viral clips are both bad data. Track how often a human has to fix the agent's output.
• Start narrow, then widen. McDonald's tested in a little over 100 restaurants, not all of them. Give a new agent one job and let it earn more. Our guide to why AI needs a confirm button covers where to put the checkpoints.

Customer-facing AI has gone wrong in court too: read how Air Canada was held to its chatbot's answer.

How ZNOS handles this

In ZNOS, AI employees each own a job, such as Jason on sales, Raghu on ops and Sophia as executive assistant, and they work from the company's live data through the Brain. You can meet the team here.

The way we run it, autonomy is earned gradually. Routine decisions are handled. Risky ones, including money, customers, anything published under your name and production changes, wait for a human approval. The Brain watches orders, invoices and shipments around the clock and raises plain-English alerts, in-app, by email or on WhatsApp, when something looks wrong.

For customer conversations, the WhatsApp Business Agent hands off to a human, and you set "never-say" phrases it will not use. An Audit Log records who changed what, so when an order looks strange you can trace how it got that way.

The McDonald's lesson is not that voice AI is useless. It is that an AI talking to customers needs a person close behind it, and a way to flag the order that does not look right.