Lessons

Why McDonald's Ended Its AI Drive-Thru—and What It Teaches Restaurants About Phone Ordering

Updated 2026 · 8 min read

Here's the part that should make every restaurant operator stop scrolling: the company that ended the test had nearly unlimited resources, one of the most studied menus on earth, and a partner in IBM. And it still hit pause. As reported by CNBC and Restaurant Business in June 2024, McDonald's announced it was ending its automated order-taking (AOT) test with IBM and removing the technology from the more than 100 restaurants where it had been running. If they stepped back, the obvious question is: should anyone else be touching AI ordering at all?

The answer is more interesting than a simple yes or no—and it's good news for restaurants, not bad. Because the reasons that early experiment stumbled are specific, public, and largely solvable. More to the point, most of them barely apply to the place AI ordering actually works best today: not the drive-thru lane, but the phone.

What actually happened—just the public record

It helps to separate the verified facts from the noise. Here is what has been publicly reported, and nothing more.

McDonald's began an AI partnership with IBM in 2021 and tested "Automated Order Taking" (AOT) at drive-thrus.
In June 2024, McDonald's announced it was ending that IBM test and removed the technology from the more than 100 restaurants where it had been running—as reported by CNBC and Restaurant Business in June 2024.
Publicly reported challenges included difficulty with different accents and dialects affecting order accuracy, along with reports of misorders.
McDonald's said it still believes voice ordering is part of its future and would evaluate a future solution.

That last point is the one most coverage glossed over. This was not a company declaring voice AI a dead end. It was a company ending one specific test while saying out loud that it still sees voice ordering in its future. Read that way, June 2024 looks less like a verdict and more like a checkpoint.

A note on tone: there is nothing embarrassing about a large company running a real-world pilot, learning from it, and adjusting course. That is how serious technology actually gets built. The lesson here is technical, not a judgment of McDonald's or IBM.

Why early AI ordering stumbled

If you want to understand whether AI can take your orders well, you have to understand why the hard version of the problem is hard. Several forces stack up against a drive-thru bot at once.

The drive-thru is a brutal place to listen

Start with the audio. A drive-thru lane is one of the worst acoustic environments imaginable for speech: engine idle, road noise, a window cracked into the wind, a kid in the back seat, music from the dash, and a microphone mounted outside in the weather. Even a human worker asks people to repeat themselves there. Ask software to transcribe that cleanly, every time, and you have set it an unforgiving task.

Accents and dialects

Among the publicly reported challenges was difficulty with different accents and dialects affecting order accuracy. This is the real world: customers don't speak in a single tidy accent, and an ordering system that only understands some of them is not serving everyone. Early speech models were simply less robust across the full range of how people actually talk.

Menu complexity

A modern menu is a combinatorial maze—sizes, modifiers, substitutions, combos, "no this, extra that," limited-time items, and regional differences. Mapping a messy spoken sentence to the exact right item and the exact right modifiers is far harder than it looks, and small mismatches show up as the misorders that were publicly reported.

Rigid bots with no graceful exit

Perhaps the most important lesson is structural. A rigid system that tries to complete every order itself, with no clean way to hand a confusing one to a person, turns every edge case into a failure. The fix is not a smarter bot that never needs help; it's a system humble enough to pass the call to a human the moment it's out of its depth. Many early pilots lacked that safety valve.

Taking a message vs. completing the order correctly

Finally, there's the difference between hearing words and getting the order right. A system can transcribe a sentence and still fire the wrong thing to the line. The goal was never "write down what the customer said"—it was "put the correct order in the system." When those two come apart, you get misorders even when the speech-to-text looked fine.

So what's genuinely different now?

Quite a lot, and not just "the AI got better" hand-waving. Several concrete things have changed, and—critically—the phone sidesteps the hardest of the old problems.

Modern speech understanding

Speech models today are meaningfully more robust to natural, messy talk and to a wider range of accents than the tools available when the 2021 test began. They also track context across a conversation, so when a caller says "make that two," the system knows what "that" refers to. That doesn't make them perfect—nothing is—but the floor is much higher than it was.

The phone is a more controlled channel than a drive-thru

This is the quiet, decisive advantage. Most of what made the drive-thru hard is about the environment, not the AI. On a phone call, the caller is usually speaking directly into a handset, in a quieter setting, without wind, engine noise, or a back-seat conversation fighting for the microphone. The signal arriving at the system is simply cleaner. The single hardest variable in the McDonald's experiment—chaotic outdoor audio—is largely removed before the AI even starts.

Completing the order in the POS, not just transcribing

The systems worth using don't hand your staff a transcript to re-key. They place the order directly into the point-of-sale and fire it to the kitchen—either as part of the POS or by integrating deeply with one, such as Square, Clover, Loyverse, Epos Now, or Revel. Grounding the conversation against your real menu and modifiers is what keeps the AI from inventing items, and completing it in the POS is what closes the gap between "heard the order" and "got the order right."

Multilingual by default

Modern voice AI commonly serves English, Spanish, and Chinese among other languages, and can detect the caller's language and switch automatically. The same menu and modifier grounding applies in each language, so a Spanish-speaking caller's order maps to the same kitchen ticket an English-speaking caller's would.

The safety valve the early pilots lacked: graceful human handoff

This is the heart of it. A well-built phone system stays in its lane and transfers to a person whenever:

The caller simply asks for a human—caller preference always wins.
The order is unusually large, a catering job, or from a known VIP who deserves a personal touch.
The request is genuinely unusual or outside what the system can safely complete.

The point of AI here is not to replace your staff or to trap callers in a bot. It's to catch the routine, high-volume calls so your team can give full attention to the ones that need a person. A rigid pilot with no escape hatch makes every hard moment a failure; a graceful handoff turns those same moments into a smooth experience.

Why early pilots stumbled	What good phone ordering does differently
Noisy, outdoor drive-thru audio with wind and engines	A phone handset is a quieter, more controlled audio channel
Difficulty with some accents and dialects (publicly reported)	More robust modern speech understanding across natural speech
Complex menu mapping leading to reported misorders	Conversation grounded on your real menu and modifiers
Transcribing words rather than completing the order	Order placed directly into the POS and fired to the kitchen
Rigid bot with no clean way to reach a person	Graceful transfer to a human for VIP, large, or unusual calls
One-language service in diverse neighborhoods	Multilingual incl. English, Spanish, and Chinese, switching automatically

How to evaluate any AI phone vendor honestly

The McDonald's story is the best buyer's checklist you'll ever get for free, because it tells you exactly which questions separate substance from a slick demo.

What happens after the caller hangs up? If the answer is "we send your staff a transcript" or "we create a ticket someone confirms," that's manual re-entry in a nicer outfit. Insist on completion in the POS.
Is it grounded on my real menu and hours, or a generic script that can invent items and quote prices that don't exist?
When and how does it transfer to a human? There must be a clean escape hatch—for VIPs, large orders, and anything unusual.
What languages, and does it switch automatically?
Can I hear it myself—on a real line? This is the big one. A canned recording proves nothing; the entire McDonald's lesson is that systems behave differently in the messy real world than in a polished clip. You should be able to call a live demo and judge it with your own ears.

That last point deserves emphasis. Don't buy AI ordering from a slide deck or an edited highlight reel. Pick up a phone, call a live line, talk like a real customer—mumble, change your order mid-sentence, ask for a human—and listen to how it actually handles you. You can call our live demos at /#try; they're real lines, not canned recordings.

An honest note: no system is perfect, and that's the point

It would be easy to end with "and that's why our AI never makes a mistake." It would also be false. No ordering system, human or AI, is flawless—people mishear orders too. The right goal isn't a system that never stumbles; it's a system that knows when it's stumbling and hands the call to a person before a small confusion becomes a wrong order. The graceful human handoff isn't a fallback bolted on for emergencies. It is the design. That's the single biggest lesson from June 2024: the safety valve is the feature.

Read alongside McDonald's own statement that voice ordering is still part of its future, the takeaway for an independent restaurant is encouraging. The hardest part of that experiment—chaotic drive-thru audio—doesn't apply to your phone. The solvable parts—speech robustness, menu grounding, POS completion, and a clean human handoff—are exactly what a good phone system is built around today.

Hear AI phone answering that completes the order—then judge for yourself

KwickPhone answers every call and places it natively into your POS, or bolts onto the ordering system you already run—with a graceful handoff to your team whenever a caller wants a person. Don't take our word for it: call our live demos at /#try.

Book a demo

Frequently asked questions

Did McDonald's stop using AI for ordering?

As reported by CNBC and Restaurant Business in June 2024, McDonald's announced it was ending its automated order-taking test with IBM and removing the technology from the more than 100 restaurants where it had been running. McDonald's also said it still believes voice ordering is part of its future and would evaluate a future solution.

Why did the McDonald's AI drive-thru struggle?

Publicly reported challenges included difficulty with different accents and dialects affecting order accuracy, along with reports of misorders. The drive-thru is also a hard audio environment—engine noise, multiple passengers, and a moving vehicle all work against clean speech understanding.

Is the phone an easier channel for AI ordering than the drive-thru?

In general, yes. A phone call is a more controlled audio channel than a noisy drive-thru lane: the caller usually speaks directly into a handset in a quieter setting, without wind, engine noise, or back-seat conversation competing for the microphone. That gives modern speech understanding a cleaner signal to work with.

What makes AI phone ordering work today?

Modern speech understanding, a cleaner audio channel, completing the order directly in the POS instead of just transcribing it, multilingual support including English, Spanish, and Chinese, and—crucially—a graceful transfer to a human whenever the caller prefers a person or hits a large, VIP, or unusual order. That human safety valve is the part rigid early pilots lacked.

How do I evaluate an AI phone vendor honestly?

Ask what happens after the caller hangs up—does it complete the order in your POS or just send a transcript? Confirm it is grounded on your real menu and hours, ask how and when it transfers to a human, and insist on hearing it yourself by calling a live demo line rather than listening to a canned recording. No system is perfect, so the clean human handoff matters.