McDonald's ended the drive-thru test before it became the operation

In 2019 McDonald's bought a voice-ordering startup called Apprente and renamed the group McD Tech Labs. Two years later it sold that lab to IBM and announced a global partnership, on terms that were not disclosed. Amelia Lucas, writing at CNBC on 3 July 2024, reported what happened when the work left the lab and met the restaurants.

The technology had already been tried at a handful of sites around Chicago. Moving the lab to IBM produced the larger trial, roughly 100 restaurants. In June 2024 the company told franchisees it was ending the Automated Order Taker test. Lucas writes that the trial fell short of McDonald's standards, and that an early mover in voice ordering was turning to other vendors. The test had run in real restaurants. It had not become the way those restaurants take an order.

Two sources familiar with the technology told CNBC it had trouble with accents and dialects, which hurt accuracy, among other problems. McDonald's declined to comment on accuracy. IBM did not answer a question about it. The sentence the company did send came from Mason Smoot, senior vice president and chief restaurant officer for the U.S. He told franchisees there had been successes, and that McDonald's saw a reason to explore voice ordering more broadly. The lab had been real enough to buy and then to sell. The question a crew would ask, whether the speaker box can handle the way customers talk, did not get a public answer. Ending the trial is a decision restaurants can plan around. Silence on the accuracy mark leaves them with a story and no number.

The same piece shows how far other chains had and had not got. Taco Bell, owned by Yum Brands, was expanding a voice test from five California restaurants to 30, after executives said in May 2024 that consumer feedback had been positive. Wendy's had announced a 2023 test at one company-owned restaurant in Columbus, Ohio, with Google. White Castle planned to put SoundHound's system into more than 100 restaurants by the end of 2024. One site, a step from five to 30, and a plan for 100 are three different distances from a chain that takes every order this way.

Andrew Charles, a T.D. Cowen analyst, told CNBC the early movers tended to be chains with lower sales per restaurant, where a tool that might offset labour is easier to justify. He thought that within 12 to 18 months at least two of the top 25 U.S. chains would take a small trial across the whole system. He pointed at third-party delivery: once McDonald's went with Uber, other chains followed. His view in July 2024 was that McDonald's would not be the company that pulled the others over this time.

Lucas is plain about why the tests got funded. Labour costs were rising before the pandemic. Covid pushed business into the drive-thru. California's 2024 move to a $20 fast-food wage made operators look harder at tools that might cut labour, which until then had mostly automated back-of-house work. A National Restaurant Association survey in the piece found that 16 percent of operators planned to invest in artificial intelligence that year, including voice recognition, with the serious money coming from large chains. Interest of that kind will fund a pilot. It does not name the restaurant operator who owns the result, the accuracy the chain already demands, or the meeting where someone may stop.

Presto Automation shows how the live operation can sit inside a vendor's claim. In securities filings the year before, Presto disclosed that human agents in places including the Philippines and India completed orders. Interim chief executive Gee Lefevre said that practice is common and helps train the system without loading the restaurant's own crew. In May 2024 the company showed a version it called fully autonomous. Lucas notes that the earlier gap in transparency can put operators off. The question to ask, before a signature, is who finishes the order when the model does not, and where that person sits.

SoundHound, described as an early leader, offered its own rates. It said the AI takes more than 90 percent of orders with no person stepping in, against a human accuracy rate of 80 to 85 percent, and that lanes run about 10 percent faster because the system works quicker and tries to add an item to every order. Charles said the same tools might later serve people who are not speaking English. Those are the vendor's figures and an analyst's look ahead, as CNBC reported them. McDonald's, in the same article, had run a restaurant test, declined to discuss accuracy, and ended the IBM partnership. A vendor can publish a rate, and the chain that hands over the food can still refuse to adopt the tool.

Sara Senatore of Bank of America Securities wrote that week that restaurants risk their reputation with this technology, because an inaccurate order still delays and frustrates the guest when a person then takes over. A National Restaurant Association consumer survey from earlier in 2024 found that a majority of baby boomers would rather have fewer technology options when they eat out. The building can reject the demo on its own. Weak Wi-Fi, a noisy highway, and a long menu are the conditions Lucas says make voice ordering harder. A quiet test bay with a short menu is a different product from a highway site at noon.

Ron Shaich, Panera's founder and at the time chair of Cava and chief executive of Act 3 Holdings, told CNBC he would rather follow once voice ordering works than be first. He says he was early to free Wi-Fi in Panera restaurants, to tying the app to the loyalty program, and to self-order kiosks. On voice ordering, he told CNBC that nobody is choosing a restaurant because it has the technology. Guests notice whether the order was right and whether the lane moved.

If you are deciding whether a pilot may leave the team that built it, the July 2024 account is specific enough to use on work that has nothing to do with restaurants.

1. Name the operating owner before the lab is sold or expanded. The hundred-restaurant test at McDonald's came after McD Tech Labs was sold to IBM. That name has to be in the room when the result is called.
2. Write the pass mark in the operation's units before the test starts. If the trial ends and nobody will say whether accuracy was hit, the sites learn that the number stays with the lab.
3. Run the hard version. Accents were part of what sources described at McDonald's. Noise, a long menu, and a weak connection are the conditions the piece says real sites will present.
4. Ask who finishes the job when the model stops, and get it in the contract. Presto's filing is the shape of the question. People offshore can be a deliberate design. They are also a cost the operator should see before go-live.
5. Keep a one-site test, a five-site test, and a plan for 100 sites in the test column. Positive guest feedback, which Taco Bell cited on the way from five restaurants to 30, is not yet a decision that every site will run this way.
6. Put a person on the path for the order the model gets wrong. Senatore's point is that a late handoff still leaves the guest with the delay, and with your brand.
7. Separate why you funded the pilot from why you would run it daily. Wages and a wave of interest explain the budget. Shaich's check is whether the experience beats what the restaurant already does. If it does not, stopping this version, and saying so to the people who would have run it, is an available outcome. McDonald's took it with IBM.

An AI governance session is where the sponsor and the operator write the owner, the pass mark, and the stop rule for one live pilot. The lab can finish a demo without that page. The operation has to live with the answer.