Same Problem, Same Technology, Opposite Outcomes: AI in the Drive-Through
Why Wendy’s AI Drive-Through Worked and McDonald’s Didn’t
Two of the largest quick-service restaurant operators in the world tested AI-powered drive-through ordering at the same time. One is scaling to 500+ locations. The other pulled the plug entirely. The technology was roughly equivalent. The difference was operational.
I keep coming back to this story because it captures the differences I see repeatedly in consumer-facing businesses trying to do something meaningful with AI. Success has nothing to do with who has the best model or the biggest technology partner. It comes down to who sets the programme up properly, governs it sensibly, and resists the urge to scale before the evidence supports it.
The setup
Wendy’s operates over 7,000 restaurants, the vast majority franchised. Drive-through accounts for 70-80% of sales across the QSR sector. The lane is the business.
The operational tension is familiar to anyone who has run high-volume consumer operations: speed, accuracy, and labour pulling against each other constantly. Staff turnover in QSR runs at 100-150% annually. Every second added to a transaction costs revenue. And the ordering process, despite decades of incremental improvement, remained fundamentally manual. One crew member listening, interpreting, keying the order, confirming, and moving on. At peak hours, the pressure on this single interaction drives errors, slower service, and frustrated customers.
In 2023, Wendy’s CEO Kirk Tanner and CIO Kevin Vasconi committed $53 million to a digital-first strategy. Worth noting: the investment was committed before the first pilot went live. The pilot’s job was to validate the deployment approach, not to justify the spend.
The same year, McDonald’s was running its own AI drive-through programme with IBM. Two operators, same problem, same window of time.
What happened
Wendy’s partnered with Google Cloud to build FreshAI, a system using large language models to understand natural speech at the drive-through. Customers speak normally, including the pauses, corrections, and stacked customisations (”no pickles, extra onion, but only on one of the sandwiches”). The system processes orders in real time and confirms them on the digital menu board as it goes.
They started with four restaurants in Columbus, Ohio. Not forty. Not four hundred. Four. They expanded to dozens, then hundreds, adding capabilities as they went: dynamic menu boards adjusting by time of day and weather, multilingual support detecting the language being spoken, and automatic escalation to a crew member when the system’s confidence drops below a threshold. The escalation piece is the most important design choice in the entire programme.
McDonald’s took a different approach. IBM built an Automated Order Taking system and deployed it across roughly 100 locations simultaneously. In June 2024, McDonald’s announced it was pulling the technology from every test restaurant. The system had struggled with accents, background noise, complex modifications, and edge cases at rush hour. Customers posted videos of the system adding hundreds of chicken nuggets to their orders. McDonald’s called it “a necessary reset, not a failure.”
Why one worked
For me, this comes down to three things.
First, the scaling discipline. Four locations gives you room to learn, iterate, and fix problems before they become systemic. One hundred locations means you are firefighting across a network while simultaneously trying to improve the product. Wendy’s had room to get it wrong and fix it before anyone noticed. McDonald’s failures went viral.
Second, the fallback design. When FreshAI’s confidence drops, the order hands off to a crew member automatically. The customer never knows the handoff happened. This single design choice meant accuracy problems never became customer problems. McDonald’s system lacked this graceful degradation. When it failed, it failed in front of the customer, and then in front of the internet.
Third, the framing. Wendy’s positioned FreshAI as a tool to make crew members’ jobs easier, not a system replacing them. As Vasconi put it: “We’re using AI to make the experience better for our customers and easier for our crew members.” This framing shapes how frontline staff experience the technology, turning crew members into allies of the system rather than people threatened by it. I think this gets underestimated in most AI programme planning: it is a change management decision as much as a technology one.
The numbers
The results at Wendy’s tell a straightforward story. Transactions are 22 seconds faster on average. This sounds modest until you do the arithmetic: at a busy drive-through processing 200-300 cars a day, 22 seconds per transaction adds up to 70-100 extra minutes of capacity. Across 500 locations, thousands of additional orders served daily without adding staff or extending hours. The majority of orders now complete without human intervention.
An operational improvement, well governed, scaling on evidence. Which, in my experience, is what most successful AI programmes look like.
What this means for you
If you are leading an AI programme in a consumer-facing business, the Wendy’s and McDonald’s comparison gives you a clean framework for the conversations you need to have with your board.
Commit the investment before the experiment starts, so your team is validating how to deploy rather than fighting for budget while trying to prove value. Start with the smallest number of locations generating useful data, and resist pressure to scale for optics. Design the fallback before you design the feature, because in a customer-facing environment, how it fails determines whether you get the chance to make it succeed. And frame the technology around the people who will use it every day, not around the cost line you hope to remove.
What separated these two programmes was operating discipline.
I publish similar breakdowns and tools to succeed with at transformationplaybook.ai/blog. How consumer-facing businesses are applying AI operationally, what’s working, and what isn’t.