McDonald's spent nearly five years building an AI to take drive-thru orders, then switched it off

What McDonald's switched off in July 2024

In September 2019, McDonald's agreed to buy Apprente, an early-stage voice-AI company, and folded its engineers into a new in-house group called McD Tech Labs, aimed at faster and more accurate order taking at the drive-thru.¹ Two years later it sold that group to IBM and let IBM scale the system, which the two companies called automated order taking, across more than 100 US restaurants.² In June 2024, McDonald's told its franchisees to turn the system off. Every test restaurant removed it no later than July 26, 2024.³

Taking an order at a drive-thru sounds like the easy case for a talking computer. The menu is fixed, the same items repeat all day, and the customer is trying to be understood. McDonald's had five years, the company that built the technology, and then IBM's scale behind it. It still could not make the system accurate enough to keep.

From Apprente to IBM and back off the menu

McDonald's did not build the order-taker itself. It bought the pieces, assembled them in-house, and then sold the whole effort to a company with more experience running customer-service AI.²

1. Sep 2019

   McDonald's buys Apprente¹

   The Apprente team becomes the founding staff of McD Tech Labs, a new group inside McDonald's Global Technology, chartered to build voice ordering for the drive-thru first and mobile and kiosks later.

2. Spring 2021

   The system goes live at 10 Chicago restaurants⁴

   The first public test puts automated order taking on the speakers at ten Chicago-area drive-thrus.

3. Oct 27, 2021

   McDonald's sells McD Tech Labs to IBM²

   A joint statement moves the team to IBM's Cloud & Cognitive Software division. Both companies say the testing so far "has shown substantial benefits to customers and the restaurant crew experience."

4. 2021–2024

   IBM scales the system past 100 restaurants⁵

   The test runs for about two years across more than 100 US drive-thrus. Later reporting says the system struggled with accuracy across accents and dialects, and that franchisees found it expensive to run.

5. Jun–Jul 2024

   McDonald's ends the test³

   Mason Smoot, chief restaurant officer for McDonald's USA, tells franchisees the IBM test will end with no expansion. Automated order taking is switched off in every test restaurant no later than July 26, 2024.

Three things have to work between the speaker and the fryer

A voice order-taker is not one program. It is a chain of three, and an order survives only if each link holds.

1. Turn sound into text

   The first stage is automatic speech recognition: the part that turns the audio coming off the speaker into written words. A modern recognizer like Whisper does this by predicting the most likely words for the sound it hears. This is the piece most people mean when they say the AI is "listening."

2. Turn the text into an order

   The second stage reads those words and decides what was actually ordered: which menu items, how many, which changes, and what to ignore. "Lemme get, uh, two of the number three, no pickles, and a large, actually a medium, Coke" has to become clean line items on a kitchen screen.

3. Decide when to call a human

   The third stage is the fallback. When the system is unsure, it hands the order to a crew member who listens and fixes it. How often that happens is the real measure of whether the automation is doing the job or a person is.

Each stage can fail, and the drive-thru attacks the first one hardest. A drive-thru is a noise machine: idling engines, traffic on the road behind, wind across the microphone, a passenger talking over the driver, and menu names said forty different ways. Speech recognition degrades in exactly these conditions. The clearest published measure of that comes from the Whisper speech model, whose developers scored it on recordings with noise mixed in at controlled levels.⁶

Signal-to-noise ratio, or SNR, is how loud the speech is next to the noise around it, measured in decibels. Word error rate, or WER, is the share of words the recognizer gets wrong. As the ratio drops below about 10 decibels, WER climbs steeply, and even the best model in the test loses the words.⁶ A car idling at a menu board, with traffic behind it, pushes the ratio down toward that range.

McDonald's never published how accurate its own system was. The figure that gets repeated, about 85 percent, comes from reporting rather than from McDonald's or IBM, and neither company released a measured number.⁴ Taken at face value, 85 percent accuracy means a crew member had to step in on nearly one order in five.⁴ The same reporting that carried that figure said the system had the most trouble with different accents and dialects.⁵

Why no one can say whether it worked

McDonald's did not describe the ending as a failure.

Set that against the public record, and the record cannot confirm it or refute it, because there is no accuracy number in the record either. What the public saw instead were the failures customers filmed. In one, bacon was added to an ice-cream order. In another, a request for vanilla ice cream and water came back with several ice creams, packets of ketchup, and two servings of butter.⁸ These clips are real and were carried by mainstream outlets. But they are user-posted videos with no verifiable original poster, restaurant, or date. They show the behavior was filmed and went viral. They do not measure how often it happened.⁸

McDonald's says the test succeeded well enough to keep pursuing the idea. The visible evidence is a run of filmed mistakes and a reported accuracy in the low-to-mid 80s, under the roughly 95 percent that competitors treat as the bar for going fully automated.⁹ Two things would separate the accounts: a measured order-accuracy rate over the pilot, and the stated reason for stopping. McDonald's gave neither, so the question stays open.

The same order still routes to a person

The McDonald's system needed a crew member on some share of orders. Across the industry, the human is often further away and doing more of the work.

The plainest documentation of that comes not from McDonald's but from a competitor. In January 2025 the Securities and Exchange Commission settled a case against Presto Automation, which sold AI voice ordering to chains including Carl's Jr., Hardee's, and Del Taco.¹⁰

The order a driver spoke into a Presto board was, most of the time, typed in by a person on the other side of the world. Presto had publicly claimed its system "eliminat[es] human order taking," which the SEC found false.¹⁰ This is a different arrangement from McDonald's, whose humans were the crew at the counter rather than remote agents. It makes the same point anyway: an "AI" drive-thru can be mostly a person, and the label on the speaker does not tell you which.

The task did not disappear when McDonald's stopped. White Castle runs SoundHound's voice AI at more than 100 drive-thru lanes, many of them around the clock, and SoundHound claims a 90 percent order-completion rate without human help.⁹ Whether that holds in the same engine-and-traffic noise that pushed McDonald's system into the low 80s is the question no press release answers. The customer at the speaker still cannot tell whether a computer or a person is taking the order.