AI Takes the Drive-Thru: Fast Food as Labor Displacement Proving Ground

Let's sit with that 90% for a moment.

If you operate 3,000 drive-thru locations and process even 200 orders per location per day, you're running 600,000 transactions daily. A 10% error rate is 60,000 wrong orders every 24 hours. That's not a rounding error — that's a full-time cleanup operation. The accuracy claim is doing heavy lifting, and it's doing it by redefining what accuracy means: Presto measures a composite of greeting quality, upsell attempts, and order entry — which lets a system that misheard your Blizzard flavor still score well because it remembered to ask if you wanted a medium.

This is DQ's second serious pass at this. Before them, McDonald's ran a two-year experiment with IBM that ended in June 2024 after the system put bacon in a customer's ice cream and generated hundreds of dollars in unsolicited McNuggets. McDonald's is now working with Google Cloud on a new attempt. Taco Bell scaled to 500 locations before someone discovered you could crash the system by ordering 18,000 cups of water. There were viral videos. There was a pivot announcement. Both are back now with human override layers — not because the AI works, but because "AI-assisted" is easier to defend than "AI-replaced-and-failed."

Which brings you to the actual story, which the accuracy numbers are designed to obscure: this isn't a customer experience initiative. It's a labor displacement pilot. The drive-thru window is one of the last customer-facing jobs in fast food that hasn't been automated away. Order-taking, upselling, greeting — hand those to a voice model and you've eliminated a meaningful chunk of the labor cost at each location. That's not speculation. It's the math Presto's investors are running, and it's the math franchisee operators are implicitly running every time they describe "improved efficiency."

The companies don't say "we're replacing workers." They say "we're improving the customer experience." They measure satisfaction scores instead of staffing levels. They report accuracy rates without disclosing how accuracy is measured. The technology is fine for what it does — it's a voice assistant with context about a specific menu. The framing is the story. When accuracy is defined to include upsell attempts, what the system is being optimized for becomes clear. It's not trying to get your Blizzard right. It's trying to sell you a larger Blizzard.

Three cycles in, the pattern is consistent: announce, deploy selectively, discover failure modes, insert humans as backup, rebrand as "human-AI collaboration," continue calling it automation. At no point does anyone lose a job on paper. At no point does anyone get hired for the job that isn't there anymore.

The drive-thru AI will improve. The accuracy will creep toward 95%. The error modes will get weirder and rarer. And at some point, the humans currently serving as the override layer will become hard to justify on the labor report.

That's not dystopia. That's just the lifecycle of a pilot program.