McDonald's Drive-Thru AI

What They Said

In 2019, McDonald’s acquired Apprente, a voice AI startup, for an undisclosed amount (estimated $100M+), and partnered with IBM to deploy automated voice ordering across its drive-thru network. The vision: AI that could take orders faster and more accurately than humans, reducing wait times and labor costs across 14,000 US locations. CEO Chris Kempczinski called it “a critical part of our accelerating the arches strategy.”

By 2022, the system was deployed in over 100 test locations, with plans for nationwide rollout.

What Actually Happened

In mid-2024, McDonald’s announced it was ending the IBM partnership and removing the AI ordering system from all test locations. Social media had become a graveyard of viral videos showing the system adding hundreds of dollars of food to orders, unable to understand basic modifications (“no pickles”), and hilariously misinterpreting accents, background noise, and children’s voices.

But the viral failures weren’t the real reason for the pullback. The fundamental problem was that drive-thru ordering is one of the most acoustically hostile environments imaginable: engine noise, wind, multiple passengers talking simultaneously, menu items with similar names, infinite modification combinations, and customers who change their minds mid-sentence. The AI achieved 85% accuracy in controlled testing — and roughly 65% accuracy in real-world conditions. That 20-point gap meant one in three orders had errors.

The Root Cause

Demo conditions and production conditions are different universes. McDonald’s AI worked in quiet labs with clear speakers ordering standard menu items. It failed in drive-thrus with diesel trucks idling, kids screaming, and customers ordering “a McChicken but like, without the mayo and can you add Big Mac sauce, actually wait, make that two.”

The change management failure was equally damaging. Drive-thru employees, instead of being freed from order-taking, became full-time AI babysitters — listening to every AI interaction, correcting errors in real-time, and apologizing to frustrated customers. The labor savings evaporated. In many locations, the AI system actually increased labor costs because correcting AI errors took longer than just taking the order.

The Pattern to Watch For

Voice AI in uncontrolled environments is the hardest deployment scenario in enterprise AI, and most vendors dramatically undersell the gap between demo accuracy and production accuracy. If your vendor shows you a voice AI demo in a conference room, ask them what the accuracy is with background noise at 70 decibels, three people talking simultaneously, and regional accents the model wasn’t trained on.

The broader pattern: any AI system that works between humans (customer-facing, in a physical environment with real-world variability) will face an accuracy gap between lab and production that can easily be 15-25 percentage points. Plan for that gap, or plan for failure.

What You Should Steal

McDonald’s eventually pivoted to a smarter approach: AI for order confirmation and upselling (text-based, on the drive-thru screen) rather than voice-based order taking. The AI reads back the order for confirmation and suggests add-ons based on order patterns — a lower-risk, higher-value application of the same technology. The lesson: when your AI can’t handle the full workflow, find the piece of the workflow where it adds value without being the single point of failure.