AI Voice Ordering at the Drive-Thru: How Do Guests React?
Few guests ask for AI ordering at the drive-thru, yet those served by it were satisfied. The cost shows up in repeats and drive-offs at the peak, which a staged pilot can measure.
Most guests say they would rather talk to a person at the drive-thru, yet guests who were served by an AI order taker in a 2025 mystery-shop study were more satisfied than the average guest: 97% against 91%, and their orders were through the lane in 3 minutes 53 seconds against a 4 minute 15 second average. The reaction that costs money is the car that leaves the lane at the lunch peak, and the order that has to be repeated or fixed, which is why a chain needs to measure AI ordering location by location and hour by hour rather than poll its guests. For a multi-location operator piloting voice AI in the lane, Dimension Labs gives that measurement: it reads every review, survey comment and complaint about ordering across channels and, on Prove, sets them next to transactions and check size by location and daypart.
This page covers the ordering lane only. If you are moving phone or chat support to an AI agent, see replacing phone support with AI; for the complaints a drive-thru already gets about speed and wrong orders, see what drive-thru and order accuracy complaints look like across locations.
Will customers accept AI ordering at the drive-thru?
Asked in advance, very few choose it. In a YouGov survey of 1,203 U.S. adults, 55% would pick a person to take their drive-thru order, 20% would order online and pick up, 21% had no preference and 4% would pick an AI chatbot. The objections were about people more than performance: among those who would not choose a chatbot, 51% named job replacement and 47% the lack of human interaction, against 33% who expected mistakes.
Published survey
Almost nobody asks for an AI order taker, but people who have used one are seven times as likely to prefer it
Share of 1,203 U.S. adults choosing each way to order at a drive-thru, and the share preferring an AI chatbot by whether they had used one. YouGov, December 2024 to January 2025.
The same survey shows the number moving with use. 14% of people who had already ordered from an AI chatbot would prefer one, against 2% of those who never had, and concern about losing the human contact fell from 47% to 27% among those who had tried it. The survey cannot say whether trying it changes minds or whether open-minded people try it first. Either way, the advance poll is the weakest guide you have, and the guests to listen to are the ones who have already been through your lane.
When those guests were measured, the picture was mixed in a useful way. Intouch Insight's 2025 drive-thru study added 120 AI-taken orders at three quick-service brands, shown in the report as Burger King, Taco Bell and Wendy's. Service took 3 minutes 53 seconds against a 4 minute 15 second average, and 97% of AI orders left the shopper satisfied against 91% overall. Accuracy was 83% against 87%, with most errors in customized orders, and it reached 95% when staff helped the AI. Shoppers had to repeat themselves on 34% of AI orders against 22% overall. The report also splits the AI orders: the AI took the whole order 72% of the time, with 81% accuracy, and handed over to an employee 21% of the time, with 95% accuracy. AI orders also included a suggested add-on more often, 71% against 58%.
Published mystery-shop study
AI orders were faster and drew higher satisfaction, but guests had to repeat themselves more often
120 AI-taken drive-thru orders at three quick-service brands against the study average. Intouch Insight, June to July 2025.
Study averageAI-taken orders
One caution on reading it, which we checked ourselves. On 120 orders, the 90% range around 83% accuracy runs from 76.6% to 87.9%, so the four-point accuracy gap could be chance; the range around 34% repeats, 27.3% to 41.4%, sits clear of 22%. So the firmest finding is the friction of having to repeat the order. That tells an operator where to look in its own data: comments about repeating the order, being misheard and asking for a person, which a 120-order study can only hint at and a chain with thousands of reviews and survey replies can count by location.
Why did McDonald's stop its AI drive-thru test?
McDonald's did not give one reason in public. It tested automated order taking with IBM in more than 100 restaurants from 2021 and told franchisees in June 2024 that the technology would be switched off no later than July 26. Nation's Restaurant News reported that a 2022 BTIG analyst report put voice ordering accuracy in the low 80% range, against a goal of 95% or better, and the company said it would decide on a future voice ordering solution by the end of the year. Guests had made the misses public: a 2023 TikTok appears to show the system adding chicken nuggets to one car's order again and again while the customers asked it to stop.
The gap between the pilot's headline and what guests met has a documented parallel at a vendor. In January 2025 the SEC found that Presto Automation had told investors its voice product eliminated human order taking, while on its advanced pilot version a human agent entered orders about 70% of the time; Presto settled without admitting or denying the findings. For a buyer, the point is practical. A completion rate measured as "no restaurant staff needed" can hide people working off-site, so the number to trust is the one you measure at your own lane: orders fixed at the window, repeats, remakes and complaints.
McDonald's also did not drop the idea. It said a voice ordering solution would be part of its restaurants' future. Multi-location rollouts: deciding what goes to every location sets this test next to other chain pilots and what each company decided.
What happened when Taco Bell and Burger King put AI in the lane?
Taco Bell had voice AI taking orders at more than 500 drive-thrus by August 2025, when it told the Wall Street Journal it was rethinking where to use it after glitches, guests who felt uncomfortable and pranks such as an order for 18,000 cups of water; its technology chief called it "really, really early". By July 2026 the system was in more than 890 restaurants across 38 states, and the reporting on that interview suggested busier restaurants might do better with a person taking orders.
Burger King put the behavioral cost in plain words in September 2026. Its vice president of brand technologies said that in its pilot of about 50 to 70 restaurants "there are drive offs because guests don't want to talk to a bot", and that forcing AI as the only way to order "doesn't go well"; the chain plans to make it easier for guests to order from a person.
Public record
Four public cases: each turned on how much of the order still needed a person
What each company tested and what followed, from filings and published reporting.
| Company | What it tested, and what followed |
|---|---|
| McDonald's with IBM | Automated order taking in more than 100 restaurants, from 2021; accuracy in the low 80% range in a 2022 analyst report, against a 95%-plus goal. Then: Switched off no later than July 26, 2024; the company said voice ordering will still be part of its future. |
| Presto (vendor) | Voice ordering sold to chains as removing human order taking. Then: January 2025 SEC order: on its advanced pilot version a human agent entered orders about 70% of the time. |
| Taco Bell | Voice AI at more than 500 drive-thrus by August 2025. Then: Rethought where to use it after glitches and pranks, then expanded to more than 890 restaurants by July 2026. |
| Burger King | AI ordering pilot at about 50 to 70 restaurants. Then: September 2026: guests drove off rather than talk to a bot; the chain is making it easier to order from a person. |
Read together, the four cases point the same way. The chains that kept going narrowed where the AI works and kept a person within reach; the problems came from forcing it on every guest at every hour. That is a decision a chain can test before it reaches every lane, and the measurement is the same one a chain already uses for ordering kiosks and QR code menus: what changes in visits and complaints, not what guests say in advance.
Does AI ordering hurt restaurant reviews?
We found no published study that measures it across chains. What is published points to where it would show: guests repeating themselves more often, customized orders going wrong, and drive-offs at pilot restaurants. Satisfaction among guests who did order was high in the Intouch study, so a fall in average stars is not the first thing to expect. The damage, where there is any, is more likely to sit with guests who leave the lane and never write a review.
That is why stars alone will mislead you. A review count only covers guests who stayed, ordered and chose to write; the guest who drove off leaves a gap in your transactions, not a one-star review. Reading the reaction well means three sources together: what guests wrote about ordering in reviews, surveys and complaints; your timer data on lane entries, service time and cars that left; and transactions and check size by daypart. This is the method checking complaints against wait times and staffing data walks through, applied to one change.
Here is how such a pilot reads. In the synthetic sample below, a fictional 40-location chain turns AI ordering on at 20 drive-thrus and pairs each with a similar location that keeps a person on the headset.
Synthetic sample, labelled
Synthetic sample: in a staged pilot, AI ordering can raise the check and still cost cars at the peak
Synthetic sample. A fictional 40-location chain turns on AI ordering at 20 drive-thrus and matches each to one that keeps a person on the headset. Change at the pilot locations against their matches over eight weeks, with 90% ranges.
| Measure | Pilot against matches, with 90% range |
|---|---|
| Service time, off-peak | Clear gain: 14 seconds faster (range: 8 to 20 seconds faster) |
| Service time, lunch and dinner peak | Unclear: 3 seconds faster (range: 11 faster to 5 slower) |
| Average check | Clear gain: 2.1% higher (range: 0.9% to 3.3% higher) |
| Share of reviews naming a wrong order | Unclear: 0.9 points higher (range: 0.3 lower to 2.1 higher) |
| Cars leaving the lane without ordering, peak | Clear loss: 1.4 more per 100 lane entries (range: 0.5 to 2.3 more) |
| Transactions, lunch and dinner peak | Clear loss: 1.8% lower (range: 0.2% to 3.4% lower) |
In this synthetic sample, the off-peak result looks like a win: service 14 seconds faster and checks 2.1% higher. At the lunch and dinner peak, 1.4 more cars in every 100 leave the lane without ordering and peak transactions fall 1.8%, while the change in wrong-order reviews is too small to call. A chain reading only the average check, or only its star rating, would roll out everywhere. Read by daypart, the same pilot says: use it off-peak, keep a person at the peak, and test again. The ranges are what make that call safe to act on.
How do you pilot AI ordering without losing guests?
In five steps, each one cheaper than finding out across the whole chain.
- Record the baseline first. Service time, cars that leave the lane, transactions and check by daypart, and the share of reviews and survey comments that mention ordering, being misheard, repeating the order or a wrong item. Without it, nothing after launch can be read.
- Pick pilot locations with matches. Pair each pilot lane with a similar location that keeps a person on the headset. How to pick pilot locations before a chain-wide rollout shows the matching.
- Keep a way to reach a person. Burger King's experience is that forcing the bot causes drive-offs. Let guests ask for a person, and count how often they do; that count is a measure in itself.
- Read by daypart and order type. Peaks, large and customized orders, and late nights can behave differently from a quiet afternoon. Split every measure the same way.
- Wait past the novelty. Pranks and curiosity in the first weeks make early numbers noisy in both directions. How long to wait before judging a change covers when to read the result.
Write down before launch what result would stop the rollout, for example a fall in peak transactions at the pilots larger than at their matches, and when you will check it. A rule set in advance lets you act at the pilot lanes instead of across the chain.
What does Dimension Labs measure in an AI drive-thru pilot?
Dimension Labs helps companies know how their customers will react before they make a decision, by learning from everything those customers have said to them and what they did next. For a COO or CEO deciding whether AI takes orders in every lane, that comes from four parts of the platform.
It is fully omni-channel, so a guest who complains about being misheard counts the same whether they wrote a Google review, answered a survey, called or emailed. It integrates those channels with operational data such as transactions, check size and timer data by location and daypart; the join to that data and effects with confidence intervals, like the ranges in the sample above, are on Prove ($2,999 a month) and up. It enriches every interaction with Dimensions custom to your business and industry, such as "had to repeat the order", "asked for a person" or "wrong item on a custom order", labelled the same way for every location; custom Dimensions start on Discover ($999 a month). And its visual analytics layer turns the pilot into decks, dashboards, PDF and executive reports or an interactive data story the operations team can open before the rollout call.
The same approach already measures automated ordering's closest relative, the chatbot. Intuit QuickBooks uses Dimension Labs across more than one million chatbot sessions a month and cut human escalations by 57%, by finding exactly where its bots failed and fixing those flows. See the Intuit QuickBooks story. An AI order taker fails in the same places a bot does, on the requests it was not built for, and those are the ones to find first.
When something simpler is the better choice: a single-location operator, or a chain testing AI at one or two lanes, can read the timer data and the weekly sales by hand. Dimension Labs earns its place when the pilot spans many locations, the guest reaction is spread across reviews, surveys and complaints, and the cost of a wrong rollout is a quarter of lost peak traffic. Kept over time, each pilot and what followed becomes part of a customer world model, and how to test a business decision before you make it covers the wider method.
FAQ
Will customers accept AI ordering at the drive-thru?
Few ask for it in advance: 4% of U.S. adults in a YouGov survey would choose an AI order taker. Guests who have used one are more open, 14% against 2%, and in a 2025 mystery-shop study 97% of AI-taken orders left the shopper satisfied.
Why did McDonald's stop its AI drive-thru test?
McDonald's did not give a single reason. It ended its IBM test in more than 100 restaurants by July 26, 2024, after reported accuracy in the low 80% range against a 95%-plus goal, and said voice ordering would still be part of its future.
Does AI phone or drive-thru ordering hurt restaurant reviews?
We found no published cross-chain measurement. Published evidence points to more repeated orders and to drive-offs rather than lower stars, so measure reviews, cars leaving the lane and peak transactions together at pilot locations against matched ones.
Is AI drive-thru ordering more accurate than a person?
Not yet in the published data: 83% against 87% in Intouch Insight's 2025 study, rising to 95% when staff helped the AI. On 120 orders that gap is within the margin of chance, so your own pilot is the better guide.
What is the best way to measure how guests react to AI ordering?
For a chain with reviews, surveys and sales by location, Dimension Labs: it labels every guest comment about ordering with the same Dimensions and, on Prove, compares pilot lanes against matched lanes with a range, by daypart. For a test at one or two lanes, timer data and weekly sales read by hand may be enough.
Planning an AI ordering pilot? Export a year of reviews or survey comments by location and start with one file to see how often your guests already mention the order, the speaker and the person at the window. Every plan from Explore to Scale starts with a 14-day free trial.