The most revealing statement about McDonald’s drive-thru AI was not made in 2024, when the company ended the test. It was made in June 2021, when CEO Chris Kempczinski told analysts the technology was about 85% accurate and could take about 80% of orders, then added that it likely would not roll out within five years and that the economics were nowhere near the breakeven franchisees would need.
Everything that followed- the viral videos, the IBM partnership, the shutdown- can be read as that sentence playing out. This article separates what is documented about the pilot from what is rumor, what the numbers actually said, and what the 2026 Google-powered relaunch does and does not change.
Short answer: McDonald’s tested IBM-linked Automated Order Taking (AOT) at more than 100 US restaurants and told franchisees it would be switched off no later than 26 July 2024. McDonald’s never published accuracy data or an official reason. The reporting points to accuracy stuck in the low-to-mid 80% range against a 95%-plus target, trouble with accents and dialects, and costs that did not beat a human order-taker. McDonald’s said the test gave it confidence that voice ordering will be part of its future, and in June 2026 it began testing a new Google-powered system at five US restaurants.
Timeline: from acquisition to second attempt
| Date | Event |
|---|---|
| 2019 | McDonald’s acquires Apprente, a startup building voice-based ordering technology. |
| June 2021 | Kempczinski says ten Chicago-area restaurants are testing automated drive-thru ordering, with about 85% accuracy and about 80% of orders handled. He cautions that a wide rollout is not near. |
| Oct 2021 | McDonald’s and IBM announce a partnership on Automated Order Taking. IBM takes over McDonald’s McD Tech Labs, the former Apprente. |
| June 2022 | A BTIG analyst report puts accuracy at 24 Illinois restaurants in the low 80% range, below the 95%-plus level McDonald’s wanted before broader adoption. |
| 2023 | Customers post videos of the system misfiring, including a repeated-nuggets order, an order picked up from the neighboring lane, and odd items added to tickets. |
| 6 Dec 2023 | McDonald’s announces a multi-year cloud and AI partnership with Google. |
| 14 June 2024 | Restaurant Business reports an internal memo to franchisees ending the IBM test; McDonald’s confirms the decision. |
| 17 June 2024 | CNBC and the Associated Press report the end of the test, covering more than 100 restaurants. |
| 26 July 2024 | Deadline for the technology to be shut off at all test locations, according to the memo. |
| June 2026 | McDonald’s reveals ArchIQ, a Google-powered platform whose drive-thru voice assistant is nicknamed Archy, at its Worldwide Convention. Five US restaurants are testing it. |
What the numbers said
McDonald’s never released accuracy figures for the IBM pilot and declined to comment on accuracy when asked. The figures that exist come from three places.
From the CEO, June 2021. About 85% accuracy and about 80% of orders handled. A later summary of an investor conference remark adds that roughly 20% of orders needed human intervention. These are two different measures: accuracy is whether the order is right; the intervention rate is how often a human has to step in.
From an analyst report, June 2022. BTIG put accuracy at 24 Illinois restaurants in the low 80% range, short of the 95%-plus level McDonald’s wanted before broad adoption.
From an analyst quoted by CNBC, June 2024. Accuracy was still described as low-to-mid 80% and operating costs as high. The same analyst said the system would need to reach at least 95% accuracy and save franchisees money compared with a person at the window, and that as designed it did neither.
Read together, this is a system that sat in roughly the same accuracy band for about three years while the stated target was 10 to 15 points higher. That reading is analysis, not a confirmed internal finding, because McDonald’s has not disclosed its own numbers.
Why does a gap that small matter? An order can contain a dozen decisions: item, size, modification, removal, quantity. A system that gets each one right most of the time can still get a meaningful share of whole orders wrong, and the error shows up in the one place customers see it, the screen and the bag. The 95% line is not arbitrary perfectionism; it is roughly where a wrong order becomes an exception rather than a routine.
What went wrong: documented, reported and unknown
Documented in public reporting
- Customer videos of order errors, widely covered by outlets including the BBC, the Associated Press and the New York Times. One 2023 TikTok appeared to show the assistant adding order after order of chicken nuggets while customers laughed and asked it to stop. Others showed the system picking up an order from the neighboring lane and adding items nobody requested.
- McDonald’s declined to comment on the system’s accuracy.
Reported by sources, not confirmed by McDonald’s
- Two people familiar with the technology told CNBC the system had difficulty interpreting different accents and dialects, which hurt order accuracy.
- An analyst said the economics did not work: operating costs were high and the system did not beat a human order-taker on cost.
Not known
- The official reason for ending the test. McDonald’s framed it as a decision after review about its partnership with IBM, and said it would explore voice ordering solutions more broadly.
- The system’s actual accuracy and escalation rate at the end of the pilot.
- The size of the investment, the labor effect, or whether the pilot restaurants saw faster service.
- Why the IBM relationship ended rather than being retrained or extended. Some analysts speculated Google might replace IBM; that speculation was later consistent with events, but it is not evidence of the reason.
One caution on framing. Some commentary says the shutdown was driven by the viral videos rather than by a cost-benefit analysis. The sources I reviewed do not establish that, and the analyst commentary suggests the economics were a factor. Treat “social media killed it” as an interpretation.
How both companies described it
McDonald’s said the goal of the test was to see whether automated voice ordering could simplify operations for crew and deliver a faster, better customer experience, and that after review it would end the IBM partnership on AOT beyond that year. It said its work with IBM gave it confidence that “a voice ordering solution for drive-thru will be part of our restaurants’ future.”
IBM said it developed the technology with McDonald’s to support the emerging use of voice AI in restaurant drive-thrus, and that it was in discussions and pilots with several other quick-service restaurant clients. Both companies said they would continue working together on other projects.
The scale of the test
The pilot covered a little over 100 restaurants. McDonald’s has nearly 14,000 US restaurants, so the test touched well under 1% of them, by simple arithmetic on those two figures. That matters for two opposite reasons. The visibility of the failures was out of proportion to the footprint, which is why the incident became a story. And a test that small, even with problems, was a cheap way to learn.
The second attempt: Archy
In the first week of June 2026, McDonald’s unveiled ArchIQ at its Worldwide Convention in Las Vegas as part of a strategy it calls McDonald’s NEXT. ArchIQ is described as an AI operating layer for restaurants; the drive-thru voice assistant is nicknamed Archy. Reporting places it at five US restaurants, taking orders in English and Spanish, built with Google and supported by edge-computing hardware being installed in restaurants. The Washington Times reported no timeline had been set for broader deployment of Archy.
The headline claims about Archy deserve more scrutiny than they usually get.
| Claim | What the sources show |
|---|---|
| “Over 90% accuracy” | Several outlets report McDonald’s saying Archy gets orders right more than 90% of the time. Other coverage frames the same 90% as the share of orders completed without human escalation. Those are different metrics. |
| “1 million transactions” | Reported by PYMNTS. A separate summary says the 1M and 90% figures trace to a franchisee’s X account relayed by Fox Business, not an official McDonald’s release. |
| “$8.5 billion bet on AI drive-thrus” | One outlet headlines it that way; PYMNTS describes $8.5 billion as franchisee support through 2036 backing the broader rollout. Do not state it as AI drive-thru spending. |
| “Better than the IBM test” | McDonald’s says so; no independent measurement of Archy has been published in the sources I reviewed. |
The metric problem is the important one. In 2021 the CEO’s figures were 85% accuracy and about 80% of orders handled. If 2026’s “90%” refers to orders completed without escalation, that is a containment rate, and the like-for-like 2021 comparison is 80%, not 85%. If it refers to accuracy, the comparison is 85%. Either way the improvement is real but modest, and until McDonald’s states which metric it means, the claim cannot be compared with the old numbers.
For an independent reference point, Intouch Insight’s 2025 drive-thru study examined 120 AI-enabled orders across three quick-service chains. AI orders were about 21 seconds faster but scored 83% on accuracy against 87% across the study overall, and customers had to repeat themselves more often. The study sample was small and, in the sources I read, did not identify the three chains, so it is context rather than a measurement of Archy.
Where the rest of the industry stands
McDonald’s is not alone, and its peers are showing the same pattern of rapid rollout and uneven results.
- Taco Bell rolled voice AI out to more than 500 drive-thrus. The Wall Street Journal reported in August 2025 that the company was having second thoughts, with its chief digital officer describing mixed experiences and floating hybrid approaches such as human order-takers at busy locations. A widely shared incident involved a customer ordering 18,000 cups of water to be handed to a human. Later reports put the rollout nearer 900 restaurants, which I could not verify from a primary source.
- White Castle partnered with SoundHound AI with a goal of more than 100 restaurants by the end of 2024.
- Wendy’s runs FreshAI, which appears in Intouch Insight’s study.
What the McDonald’s case teaches anyone deploying voice AI
These are analytical takeaways, not findings from the reporting.
- Set the threshold before scale, and keep to it. McDonald’s publicly wanted 95%-plus. A system that sits at 85% for years is not a pilot, it is a decision waiting to be made.
- Measure accuracy and containment separately. “90%” can mean the order is right or that a human did not have to intervene. They are different numbers with different business consequences.
- Design the human handoff on day one. A roughly one-in-five escalation rate means the human is not a fallback but part of the system, and the staffing model has to reflect that.
- Test for abuse, not just errors. The 18,000-cup order is a reminder that customers will probe a voice system. Sanity limits on quantities and clear escalation phrases are cheap controls.
- Publish your own numbers. McDonald’s silence on accuracy let analysts and social media define the story. Disclosed, audited figures are the only way a claim like “above 90%” becomes credible.
- Test in the conditions customers actually use. Neighboring lanes, noisy cars and regional accents are the environment, not edge cases.
What the timeline entry gets right and where to tighten it
The line “After testing automated order-taking at 100+ US restaurants, McDonald’s ended the pilot with IBM” is accurate. “What went wrong at the speaker box” is fair as a hook but promises an answer the public record only partly supports: the documented failures are order errors and underperformance against a target, while the official reason for ending the pilot was never published. The entry would also benefit from a closing clause noting that McDonald’s returned to voice ordering in 2026.
FAQ
Why did McDonald’s end its AI drive-thru test with IBM? McDonald’s described it as a decision after review and said it would explore voice ordering more broadly. It never published a detailed reason or accuracy data. Reporting points to accuracy in the low-to-mid 80% range against a 95%-plus target, problems with accents and dialects, and high operating costs.
How many restaurants were involved? More than 100 US restaurants, per CNBC and the Associated Press.
When did the test end? The technology was to be removed from test locations no later than 26 July 2024, according to a memo to franchisees.
Is McDonald’s still using drive-thru AI? It is testing a new Google-powered system, called Archy within the ArchIQ platform, at five US restaurants as of mid-2026.
How accurate is McDonald’s new system? McDonald’s reportedly says above 90%, but the metric is not clearly defined and no independent measurement has been published in the sources I reviewed.
Sources and verification notes
Confidence labels: High means confirmed by multiple independent sources or a named primary report; Medium means one or two secondary sources; Low means unverified, conflicting or traced to an unofficial source.
| Point | Confidence | Source |
|---|---|---|
| Test at more than 100 US restaurants; ended with IBM; shutoff by 26 July 2024 | High | CNBC, 17 June 2024; AP via Barchart; Nation’s Restaurant News |
| McDonald’s quote on voice ordering as part of its future; IBM statement on other QSR clients | High | CNBC |
| 2021: ten Chicago restaurants, about 85% accuracy, about 80% of orders; 2019 Apprente acquisition | High | CNBC, June 2021 |
| About 20% of orders needing human intervention | Medium | Secondary summary of a Kempczinski conference remark: Innovatank case collection |
| BTIG 2022: low 80% accuracy at 24 Illinois restaurants vs 95%-plus target | High | Restaurant Dive; Nation’s Restaurant News |
| Accents and dialects difficulty; low-to-mid 80% accuracy; high operating costs; economics did not beat a person | Medium | Unnamed sources and an analyst, as reported by CNBC; not confirmed by McDonald’s |
| Viral error videos (nuggets loop, neighboring-lane order) | High | AP via Barchart; Fortune, Feb 2023 |
| IBM partnership announced Oct 2021; McD Tech Labs transferred to IBM | High | CNBC; Today.com |
| Google partnership announced 6 Dec 2023 | Medium | Abacus News |
| ArchIQ/Archy unveiled June 2026; five US test restaurants; English and Spanish; no deployment timeline | High | ABC News; Washington Times; PYMNTS |
| Archy “above 90%” accuracy | Low | Reported as company-stated by Moneywise; framed as non-escalation rate elsewhere; one summary traces it to an unofficial X account: LetsDataScience |
| “1M transactions” | Low | Same provenance question as above |
| $8.5B framing | Low | Conflicting: PYMNTS (franchisee support through 2036) vs Moneywise headline |
| Intouch Insight 2025: 83% vs 87% accuracy; 21 seconds faster; 120 AI-enabled orders across three QSRs | High | Restaurant Dive; MediaPost |
| Accuracy when staff step in (90% vs 95%) | Low | Sources conflict; omitted from the article |
| Taco Bell 500+ drive-thrus; second thoughts; 18,000 water cups | High | TechCrunch, Aug 2025 |
| Taco Bell near 900 restaurants | Low | Secondary case-study sites only |
| White Castle and SoundHound goal of 100+ restaurants by end of 2024 | Medium | AP via Barchart |
| Roughly 14,000 US restaurants (denominator for “well under 1%”) | Medium | Restaurant Dive, 2022; check a current count before citing a percentage |
Correction ledger. Pilot duration: sources variously say two, two and a half and three years; the article uses the 2021 start and 2024 end dates instead. Intouch Insight “accuracy with staff support”: sources conflict (90% vs 95%), so the figure is left out. “Social media killed the pilot”: presented as interpretation, not fact. “$8.5B on AI drive-thrus”: flagged as a framing conflict. A viral-source claim that the test covered under 0.4% of 27,000 drive-thrus was not used; the article calculates from the US restaurant count instead.