McDonald's
Transactional voice AI: speech recognition and understanding replacing the employee at drive-thru order taking, in an uncontrolled acoustic environment
In June 2024, McDonald's ended its AI voice ordering test with IBM across more than 100 drive-thrus, with accuracy capping at about 85%; the chain nonetheless pursues the voice AI ambition via Google Cloud.
Objective
Reduce labor costs and wait times at the drive-thru (the majority channel for US sales), standardize order taking and free up the teams.
The deployment
McDonald's acquired the voice AI startup Apprente in 2019, tested automated voice order taking in Chicago drive-thrus, then sold its McD Tech Labs subsidiary to IBM in late 2021 while signing a partnership to industrialize the technology. The system was deployed in more than 100 US drive-thrus between 2021 and 2024. The customer spoke to an AI at the kiosk, with an employee stepping in only as backup. Reported accuracy capped at around 85%, with human intervention on about one order in five. In 2023-2024, viral TikTok videos documented absurd orders: nine sweet teas instead of one, hundreds of McNuggets added while customers shouted stop. On June 13, 2024, McDonald's told franchisees the IBM partnership would end and the system would be shut off no later than July 26, 2024 - while reaffirming that drive-thru voice AI would be part of the future of its restaurants.
Results Proof C
No published financial figures: the facts rest on an internal franchisee memo revealed and confirmed by major business press (CNBC, CBS News, Restaurant Business) and on McDonald's official statement - solid, convergent press sourcing, but without a public financial document.
How it works
Documented architectureThe stack in detail
- outil Apprente (voice AI) ASR/NLU startup acquired by McDonald's in 2019, becoming McD Tech Labs then sold to IBM in late 2021
- plateforme IBM Automated Order Taking (AOT) voice order taking system industrialized by IBM and deployed in more than 100 US drive-thrus between 2021 and 2024
- llm Briques ASR/NLU IBM Watson pre-LLM generation speech recognition and understanding; reported accuracy of about 85% at the drive-thru, human recovery on about one order in five
- infra Bornes de drive-thru equipees microphones and speakers in an uncontrolled acoustic environment (noise, accents, two lanes), the worst case for ASR
Post-mortem
GraveyardWhat happened sourced
2019: acquisition of Apprente, first tests in Chicago. Oct. 2021: sale of McD Tech Labs to IBM + AOT partnership. 2021-2024: test in 100+ drive-thrus. 2023-2024: viral TikTok videos of erroneous orders, picked up by national press. June 13, 2024: internal memo to franchisees announcing the end of the IBM test; shut-off no later than July 26, 2024. McDonald's then announces an expanded partnership with Google Cloud on generative AI.
Reason for failure sourced
No official quantified admission, but a documented body of evidence: accuracy of about 85% with human intervention on roughly 1 order in 5, inability to handle accents, ambient noise and multiple voices, and the reputational damage of the viral videos. McDonald's statement: the test aimed to check whether automated voice could simplify operations for crew; the company ends the IBM partnership while wanting to explore voice ordering solutions more broadly - a disavowal of the execution, not of the ambition.
Cost sourced
Not published. Identifiable costs: acquisition of Apprente (estimated tens of millions of dollars), 3 years of piloting across 100+ sites, resale to IBM, and a real reputational cost (TikTok videos accumulating tens of millions of views, becoming the public symbol of failed voice AI). For IBM: loss of the most visible showcase of its AOT offering.
Warning signs inferred
An accuracy of about 85% at the drive-thru means 15% friction on the most frequent action of the main channel: the tolerance threshold (>95%) was known and not reached after 3 years. The drive-thru environment (noise, accents, two lanes) is the worst case for ASR - choosing the hardest terrain first was risky. The virality was predictable: every failure happens in front of a customer with a smartphone. And the sale to IBM as early as 2021 suggested that McDonald's no longer wanted to carry the development in-house.
Lessons in hindsight inferred
(1) For a transactional AI in a physical environment, the required accuracy rate is near human level from day 1: at 85%, automation costs more (recovery + reputation) than it brings in. (2) Testing a voice system in its most hostile environment, with high public visibility, maximizes reputational risk - start with channels where failure is private (app, phone). (3) Make human interventions per order an explicit stop criterion for the pilot. (4) McDonald's exit (end of the partnership, not of the technology) is a model of controlled backpedal that preserves the future option.
Yes, clearly. McDonald's reaffirmed that a drive-thru voice solution will be part of its future and partnered with Google Cloud; Wendy's expanded FreshAI, Taco Bell/Yum! deployed voice AI across hundreds of drive-thrus in 2024-2025. The failure condemns an execution (pre-LLM 2019-2021 technology, hostile environment, insufficient accuracy), not the pattern - a case of too early with the wrong technology generation.
How your customers perceive this type of use
Sourced studiesLes consommateurs n'acceptent pas les chatbots par defaut : 64% prefereraient que les entreprises n'utilisent pas d'IA dans leur service client (Gartner, 2024) et pres d'un utilisateur sur cinq du service client par IA n'en retire aucun benefice (Qualtrics, 2025). L'acceptation se construit sur trois conditions mesurees par Salesforce : savoir qu'on parle a une IA, pouvoir escalader vers un humain, comprendre la logique de l'agent.
Acceptance conditions
- Etre informe qu'on parle a une IA et non a un humain (pres de 75% le demandent, Salesforce 2024)
- Un chemin d'escalade clair vers un agent humain (45% plus enclins a utiliser l'agent IA, Salesforce 2024)
- Une logique de l'agent clairement expliquee (44% plus enclins, Salesforce 2024)
Red lines
- Rendre l'humain injoignable : c'est la premiere inquietude des consommateurs sur l'IA dans le service client (Gartner 2024) et 50% craignent que l'IA les coupe du contact humain (Qualtrics 2025)
- Remplacer le service client par l'IA sans alternative : 53% envisageraient de partir chez un concurrent (Gartner 2024)
Sources: Salesforce 2024 · Gartner 2024 · Qualtrics 2025
The same trap, elsewhere
Sources
- S1 McDonald's ends AI drive-thru orders - for now - CBS News Established press archive pending
- S2 McDonald's Abandons AI for Drive-Thru Orders - What's Next for Fast-Food CX? - CX Today Secondary archive pending
- S3 McDonald's is ending its drive-thru AI test - Restaurant Business Established press archive pending
- S4 McDonald's to end AI drive-thru test with IBM - CNBC Established press archive pending
An error, newer info, a source?
This page lives on its accuracy. If a figure has moved, if the deployment has changed, or if you have a higher-quality source, tell us. Every sourced correction is verified before publication.