The AI drive-thru reversal: what enterprise buyers should learn from fast food
Fast-food chains abandoned AI drive-thrus in 2024, then quietly scaled them to thousands of lanes. The reversal is a case study in why operational AI programmes fail on integration and measurement rather than model quality, and what to check before your next rollout.
A technology the market wrote off is quietly scaling
By the end of 2024, voice AI in the fast-food drive-thru looked like a closed case. McDonald’s had started the category in 2019 by acquiring a startup called Apprente and folding it into a division named McD Tech Labs. It later sold that division to IBM and then withdrew the technology altogether. Del Taco ended its drive-thru voicebot test. Branded Hospitality, which invests in food-service technology, sold off its early positions in the category.
“The restaurants ended up spending more time smoothing over the problem they created,” cofounder Michael Schatzberg told Wired. “Everyone was like, ‘This is the worst. Stop. Get the people back with the headsets.’”
Two years later the same technology is scaling. Taco Bell announced in July that voice AI runs 890 of its drive-thru lanes, more than a tenth of its US locations. Dairy Queen is rolling out across 25 states with plans to reach every restaurant. Twelve percent of White Castle sites use an agent named Julia, and every new store now opens with her taking orders. McDonald’s has re-entered, testing an English and Spanish system with Google at five locations. Adoption tracked by Intouch Insight moved from roughly 4 percent of US locations in 2024 to 5 percent last year, and early research puts it near 6 percent now.
For anyone running an operational AI programme, the useful material here is not the technology. It is the shape of the first failure, and what was different the second time.
The first attempt failed on operations, not on model capability
The public record of the 2023 rollouts is a series of viral clips: an order that grew to 260 chicken nuggets while customers shouted at the speaker, bacon added to an ice cream, a customer who ordered 18,000 cups of water at a Taco Bell. Those are memorable, and they are also the least instructive part of the story.
The quieter findings mattered more. Intouch Insight’s testing found AI lanes scored lower than humans on both order accuracy and friendliness. The systems struggled with accents and dialects, which chains addressed by keeping a staff member on a headset ready to take over. A fast-food order is a branching decision tree that a trained employee navigates without thinking; for the early agents it was an awkward fit.
Read that as an integration problem rather than an intelligence problem. The models could transcribe and respond. What they could not do was absorb the operational variance of a live restaurant, and nobody had designed the handoff between agent and human before go-live. The cost showed up as staff time spent repairing orders, which is precisely the cost the programme was meant to remove.
This pattern repeats well outside food service. A deployment that performs in evaluation and degrades in production is usually not a model selection error. It is an error in the surrounding process design: who intervenes, when, on what signal, and at what cost.
What changed the second time
Three things, and none of them is a breakthrough model.
First, the chains that continued did so incrementally. White Castle has been working with SoundHound for four years. Chief marketing officer Jamie Richardson describes a strategy built on steady improvement rather than waiting for an agent that never errs. “It wasn’t bad at the start,” he said. “It’s gotten so much better as time has gone on.”
Second, the operational metric moved to the front. Intouch Insight now measures AI lanes as 21 seconds faster than the industry average. That is a number an operator can act on, tied directly to throughput per hour and to labour cost per order.
Third, the vendor market matured around the use case. The current deployments name specific partners for specific chains: Presto with Dairy Queen and Del Taco, SoundHound with White Castle and Panda Express, Nvidia and Omilia with Taco Bell, Google with McDonald’s. Branded Hospitality, the investor that had exited the category, has re-entered through a partnership with Arc, which works with Jack’s and Pollo Campero. “Companies have evolved, technology has evolved, and they’re like, we can fix a lot of those problems,” Schatzberg said.
The chains that stayed in the market through the trough are the ones now scaling. That is worth sitting with before cancelling a programme that has underperformed in its first year.
The measurement problem most programmes repeat
The first wave was justified by press-release claims: faster, more accurate, better at upselling than staff. None of those were instrumented in a way that survived contact with customers, so when the clips went viral there was no counter-evidence to weigh against them and the programmes were withdrawn.
The second wave carries a specific, externally measured figure. The difference is not enthusiasm. It is that somebody defined the metric before the rollout and kept measuring it afterwards.
Most operational AI initiatives arrive at review with the same gap. The business case is written in general terms, the pilot is assessed on whether it felt like it worked, and the first visible failure decides the outcome. A programme that cannot state its baseline cannot defend itself.
Choosing not to deploy is also a position
Not every operator reached the same conclusion. Chick-fil-A consistently tops Intouch Insight’s friendliness ratings and has never signalled an intention to replace drive-thru staff. Dutch Bros and 7Brew, two of the fastest-growing chains in the country, have also stayed out.
“Every brand has to make choices that fit its business,” 7Brew chief marketing officer Nick Chavez said. “For us, hospitality starts with genuine human connection, and that’s something our guests consistently tell us they value.”
That is a defensible strategic decision, made against a clear understanding of what the brand sells. Customers notice either way. One described being served by an automated voice at a Wendy’s as bewildering, and said he worries about what the technology removes from ordinary interaction. Where the interaction itself is the product, automating it subtracts value even when it works.
Abstention, in other words, can be a decision with reasoning behind it rather than a failure to modernise.
What to take into your next deployment review
Four questions, drawn directly from the arc above:
- Where is the handoff between the automated system and a person, and has anyone rehearsed it under load?
- What single operational number defines success, who owns it, and what was the baseline before deployment?
- If the first quarter underperforms, is the plan to withdraw or to iterate? Decide now, in writing, rather than after the first bad week.
- Does automating this interaction subtract from what the customer is actually buying?
The fast-food sector paid for these lessons in public over six years. Most organisations reviewing an AI programme this quarter sit somewhere on the same curve, usually near the point where a first rollout has disappointed and the instinct is to stop.
Book a strategy consultation to pressure-test your deployment against these questions before that decision gets made for you.