The voice-ordering system that was switched off after viral mistakes

What happened

In 2021, McDonald's began testing an automated order taker, built with IBM, that used voice recognition to take drive-thru orders. Over about two years it ran at around 100 restaurants in the United States. In June 2024, McDonald's said it was ending the partnership on this technology and would switch it off in participating restaurants no later than July 26, 2024.

The decision followed a stream of videos on social media. Customers recorded orders that went wrong: one widely reported example had the system add hundreds of chicken nuggets, reported as 2,510 meals costing $264.75, and others showed unwanted items such as bacon added to ice cream. In some videos, staff stepped in to finish an order the system could not understand. Trade press, citing sources, also reported that the system struggled with different accents and dialects.

McDonald's did not frame it as the end of voice ordering. It said the test had given it confidence that a voice-ordering solution would be part of its future, and that it would evaluate options.

Why it happened: a design reading

The public reporting doesn't give technical detail, so this is an outside reading of where voice ordering is hard:

• A noisy, open-air channel. Drive-thru audio has engine noise, wind, speaker distortion and several people talking.
• Accent and dialect diversity. Recognition models perform unevenly across speakers unless they are tested and tuned on them.
• Domain vocabulary. Menu items, modifiers and brand names are unusual words.
• High-consequence mistakes with weak checks. Absurd quantities and odd combinations reached the order.
• Social amplification. Failures in public service are easy to film and share.

Design lessons

1. Test recognition per accent and locale, with real audio from the setting, and measure the gap between groups (module 20).
2. Bias recognition toward the menu, using custom vocabulary and phrase hints.
3. Validate orders with business rules. Quantities, combinations and totals have sane ranges. Flag the outliers and ask (module 10).
4. Confirm before committing, especially numbers and totals, by reading back in chunks (module 19).
5. Give people a smooth takeover: a visible, immediate way to switch to a human, with the partial order preserved (module 11).
6. Run staged rollouts with kill switches and monitoring for anomalies (module 16).

How to build it better

The key moves: a range check that triggers a confirmation, a human option, and a recovery that doesn't blame the customer.

Try it

Write the ten most common order mistakes you would expect in a noisy environment. For each, define the guardrail (range check, confirmation, human takeover) and the test, including the same order spoken in three accents.