Why McDonald's shut down its AI drive-thru — the pitfalls of voice AI adoption
Contents
McDonald's ran a pilot with IBM starting in 2021 to automate voice ordering at the drive-thru, but shut it down after falling short of its accuracy target and order-error videos went viral. It's a case where a polished demo collapsed in the real world.
At a glance
| Item | Details |
|---|---|
| Company/organization | McDonald's (US) |
| Industry | Quick-service restaurants / franchising |
| AI technology deployed | IBM voice-recognition-based Automated Order Taking (AOT) |
| Deployment | About 100 locations, run for roughly 2.5 years |
| Key outcome | Pilot ended for missing accuracy targets; that partnership with IBM was discontinued |
Background — what went wrong
The drive-thru accounts for a large share of sales, but chronic understaffing and slower service during peak hours were persistent problems. McDonald's expected that if voice AI took orders, staff could focus on cooking and packaging, and the line would move faster.
That expectation held up in demos — quiet settings, standardized pronunciation, a limited range of questions. The problem was that a real drive-thru is nothing like a demo.
How it was rolled out
According to Nation's Restaurant News, McDonald's developed Automated Order Taking (AOT) technology with IBM and piloted it at about 100 US locations for nearly 2.5 years, targeting an order accuracy rate of 95% or higher. The model kept learning from real order data, but the project stalled at the stage of deciding whether it had hit that target ahead of a nationwide rollout.
Results & outcome
- Accuracy: low 80s percent (per an industry analysis report cited by restaurantonline.co.uk, 2022) — short of the 95%-plus target.
Meaning: accuracy in the low 80s means one or more out of every ten orders is wrong. If a human has to step in every time to fix it, the rework cost can outweigh the labor savings from automation.
Viral videos also chipped away at trust. According to Fast Company, a video of the system repeatedly adding butter to a caramel ice cream order racked up tens of thousands of views on social media, alongside videos of bacon piled onto ice cream and hundreds of dollars' worth of chicken nuggets added to a single order. These cases hurt perceived trust far more than the actual error rate did.
McDonald's notified franchisees the pilot was ending, and as Seeking Alpha reported, the specific partnership with IBM ended, though the company said it had not given up on voice ordering itself.
Key factors — why the demo and the real world diverged
- Ambient noise and accents: In environments where engine noise overlaps with speaker audio quality, recognition errors rise sharply whenever accents or speaking speed deviate from the standard.
- Long-tail orders: Variations like "no sauce" suffer from a lack of training data — the more common an order, the more accurate the system is, and the rarer the combination, the more likely it fails.
- Asymmetric cost of errors: A chatbot error just means re-entering text, but at the drive-thru, the wrong food is actually handed over, and the risk of that moment going viral on video is high.
Analysis: average accuracy may look fine, but if it drops sharply in specific conditions — noisy peak hours, locations with particular accents — trust collapses at that specific location.
Limits & challenges
Results from other chains in the same period were mixed. According to restaurantonline.co.uk, UK Popeyes reported 97% accuracy in its early 2024 pilot and expanded to more locations. By contrast, according to restaurantdive.com, Wendy's FreshAI had a hands-off resolution rate in the mid-80s percent as of 2024 — similar to the level McDonald's judged insufficient. Results can diverge widely for the same technology depending on menu complexity and evaluation criteria.
Applying this to your organization
- Large enterprises and franchise headquarters: Fix the target accuracy and measurement criteria before a nationwide rollout, and test separately by location and time slot — that's the only way to find weaknesses hidden in the average.
- Small and medium locations: Check the conditions under which a vendor's claimed accuracy was measured, and test it yourself during your busiest hours and in your hardest-to-recognize accents.
- Any organization evaluating AI adoption: The more an error reaches the customer directly, the more you need to design an immediate human-intervention escape hatch from the pilot stage onward.
References
Contents
Related posts

Common patterns among companies that rehire after AI layoffs
More companies are cutting staff with AI, then refilling the same roles months later. Using Klarna, IBM, Ford, and Commonwealth Bank cases, we examine the shared pattern of headcount-first cuts without redesign—and the hidden costs of rehiring.
The Air Canada AI chatbot misinformation case — who is responsible?
Air Canada denied liability for wrong bereavement-fare advice from its chatbot, but a tribunal ruled the bot was part of its website and held it liable.

AI productivity paradox — work feels faster, so why don’t the numbers move?
There’s ample evidence that AI speeds up specific tasks, so why don’t company- and economy-wide productivity stats show it? From Solow’s paradox onward, we examine why this gap persists using recent CEO and research data.

Call center AI adoption, why does it end in rollback — failure patterns and how to avoid them
Based on a recent survey showing 74% of companies roll back call-center AI chatbots, we contrast Klarna’s failure without escalation design and Carmoola’s success with preset criteria, distilling three recurring causes.