The 9Q Grid
← Case Library

McDonald's (AI Order Taking)

Q3
Retail (F&B) · Spear · The Boomerang
Researched as of October 2026
ShareLinkedInX
AdaptAugmentAssist
Q3Automator
Q6Optimiser
Q9Conductor
Q2Tactician
Q5Partner
Q8Explorer
Q1Operator
Q4Weaver
Q7Navigator
StructuredDynamicComplex
1/9 cells occupied
Shape Read

The shutoff is accepted from the press record. No McDonald's or IBM page states the end. The occupied cell is the drive-thru test itself. McDonald's and IBM built a genuinely functional Q3 Automator for drive-thru voice ordering, one that handled thousands of canonical, English-language, unmodified orders a day with real POS integration and real throughput gains. But the system was marketed from its 2019 Apprente origins in Q5/Q6/Q7 language — multilingual, multi-accent, multi-item, context-aware conversation — while the deployable product only ever cleared the low-to-mid 80s on accuracy against a 95% internal bar. The gap between claim and capability was not primarily rhetorical inflation; it was structural, because the drive-thru's acoustic, phonetic, and semantic environment is inherently dynamic in a way a Q3 system cannot absorb without a human partnership layer, and McDonald's explicitly suppressed that layer by training crew not to intervene. Two and a half years of held-line testing let the mismatch surface publicly through viral failure videos rather than silent, unbounded scaling, and McDonald's terminated the IBM partnership system-wide in June 2024, reverting to crew-led ordering while a Google Cloud and SoundHound-Acrelec successor stack was already being staged. The episode is best read not as an AI failure but as a cell-identification failure: a competent Q3 system asked to operate at a Q5/Q6 boundary it was never architected to hold.

Trajectory: McDonald's had already pre-positioned a Google Cloud partnership before the IBM exit and is staging a successor voice-AI channel via SoundHound-Acrelec, aiming for a Q3+Q5 'assistant, not replacement' architecture similar to Wendy's FreshAI.

Inside the company · not a cell

Nothing quoted yet. A rate or a saving will sit here when it is on the record. It does not light the grid.

Initiative Map
CellInitiativeConfidenceEvidence
Q3AutomatorAOT voice ordering at drive-thru — happy-path orders on standard menuDrive-thru voice ordering (canonical orders)✅Roughly 85% order accuracy in the early Chicago test (2021); about 80% of orders machine-handled with 20% requiring crew intervention; throughput of thousands of orders daily across roughly 100 stores by 2024.As of April 2026
Q3AutomatorAOT voice ordering — accents, dialects, non-canonical phrasingSpeech recognition across regional accents and dialects✅Anonymous CNBC sources reported issues interpreting different accents and dialects affected order accuracy; CEO Kempczinski had flagged 'dialect permutations' in 2021 as a reason broad rollout would be hard.As of April 2026
Q3AutomatorAOT voice ordering — non-canonical menu requests, modifications, off-menu phrasingOrder modification and off-menu phrasing handling✅Viral failures included 9 sweet teas added when a customer wanted a Diet Coke removed, an ice cream cone topped with bacon, 'no caramel' misread as wanting cream packets, a 260-McNugget order, and $200+ chicken-nugget bills; reporting summarized that 'once your order deviated somewhat from the main menu… AI really had a problem.'As of April 2026
Q3AutomatorAOT voice ordering — ambient noise, multi-speaker car interiorsAcoustic environment / ambient noise processing✅CNBC reporting found AOT 'experienced difficulty distinguishing orders from background noise or nearby conversations,' with videos showing the system capturing passengers' side conversations and adding them to the order.As of April 2026
Q3AutomatorAccuracy target vs. realityPerformance measurement and adoption gating✅June 2022 BTIG checks across 24 Illinois stores found accuracy in the low 80% range, well below the 95% target; the gap never closed across 2.5 years of testing across roughly 100 restaurants.As of April 2026
Q3AutomatorIBM commercial framing of AOT (June 2024 termination statement)Vendor commercial communications✅McDonald's shut the system off across 100+ restaurants without expansion the same week; IBM said it was in discussions and pilots with other QSR clients but announced no major replacement wins, while the QSR voice-AI market shifted to Google, SoundHound, and Nvidia instead of IBM AOT.As of April 2026
Q3AutomatorSuccessor architecture (post-IBM): Google Cloud / SoundHound-AcrelecSuccessor platform / architecture transition🔶December 2023 Google Cloud multi-year partnership for edge and GenAI was rolled out before the IBM termination was announced; the July 2025 SoundHound-Acrelec partnership covers 25,000+ drive-thrus including McDonald's, built on LLM-based generative AI rather than the bespoke neural-NLU stack from Apprente.As of April 2026
Q3AutomatorYum Brands voice AI (comparator, not subject)Industry comparator — drive-thru and call-center voice AI◻️Operating accuracy data is scant in public reporting; scale-up announcements have preceded any disclosed evaluation methodology.As of April 2026
Published record

The lines this reading rests on. A date is the source date, when the page or the line names one.

After a thoughtful review, McDonald’s has decided to end our current partnership with IBM on AOT and the technology will be shut off in all restaurants currently testing it no later than July 26, 2024.

McDonald’s is ending its two-year-old test of drive-thru, automated order taking (AOT) that it has conducted with IBM and plans to remove the technology from the more than 100 restaurants that have been using it.

McDonald’s confirmed Monday that it decided to end a global partnership with IBM, which has been testing this artificial intelligence technology at select McDonald’s drive-thrus since 2021.

Strength
Inside its true envelope — canonical English orders on the standard menu in low-noise conditions — AOT delivered real Q3 throughput, POS integration, and speaker-post latency gains across 100+ restaurants.
Fragility
The system was pitched and scaled on Apprente's 2019 Q5/Q6/Q7-coded capability language ('complex, multilingual, multi-accent, multi-item conversational ordering') while the deployable stack remained a narrow Q3 automator, leaving no architectural slack when real customers deviated from canonical phrasing.
Failure mode
Accents, dialects, ambient noise, order modifications, and off-menu phrasing pushed the required work into Q5/Q6 territory the Q3 system could not reliably resolve, producing viral, confidently-wrong outputs (9 unwanted sweet teas, bacon-topped ice cream, 260-McNugget orders) that never closed the gap between low-80s actual accuracy and the 95% adoption gate.
Governance ceiling
McDonald's held a disciplined aggregate 95% accuracy gate for 2.5 years but never built a per-segment gate (by accent, dialect, or order-modification complexity), and it structurally suppressed the missing Q5 human-partnership layer by training crew not to intervene, converting an engineering gap into a metric-preserving policy choice.
← Back to the Case Library