The McDonald's drive-thru is the agent UX you should study

McDonald’s is piloting an AI drive-thru system called ArchIQ in five US restaurants. The system greets you, takes your order in English or Spanish, recognizes returning customers, and answers “can I get my usual?” with the right combo. The franchise running the demo says 90 percent of orders so far have completed without a human escalation. The corporate framing, from a recent CEO memo, is that “more of the customer journey becomes automated” and that “the bar for hospitality only goes up” as a result.

Set aside the politics and the labor questions for a minute, because there’s a less obvious story here. This is the largest live, voice-first, multilingual agent UX rollout currently running on US soil, sitting at one of the highest-volume customer touchpoints in the country. Whether you ever buy a Big Mac or not, this pilot is going to teach the rest of the industry more about real-world agent UX in six months than every keynote demo combined.

A few things to watch for, all of which generalize to enterprise agents.

Memory is the feature, not the model. The headline isn’t that the assistant uses an LLM. It’s that it remembers what you ordered last time. The hard engineering is the identity link, the consent flow, the retention policy, and the graceful failure when the memory is wrong. Anyone shipping a customer-facing agent is going to live or die on this, not on benchmark scores.

Multilingual is now baseline. The pilot ships English and Spanish on day one. Any enterprise agent that lands in 2026 with “English only, others coming soon” is going to look dated inside a quarter. Plan for that in your data, your evals, and your governance from the start, not as a sprint at the end.

The escalation path is the product. 90 percent without-escalation is a great number, and it also means 1 in 10 customers needs a human. The interesting design question is what that handoff feels like: how quickly the human is in the loop, what context they get, whether the agent has already half-solved the problem or made it worse. For SI partners building customer-facing agents, the escalation UX is usually where the deal is won or lost.

The brand defines the failure mode. When the agent mishears your order, the customer experience is McDonald’s, not “the LLM vendor.” Every enterprise that ships an agent inherits the failure mode of the slowest, weirdest part of the pipeline. Plan for that politically, not just technically.

Microsoft’s agent platform thesis from Build keeps insisting that the system around the model matters more than the model. McDonald’s just put that thesis on the menu, literally. Watch the pilot. The teams that learn from it will ship better enterprise agents than the teams that ignore it because it isn’t enterprise.