Restaurant Technology and AI

Drive-Thru Voice AI: What Pilots Show

What public drive-thru voice AI pilots show about accuracy, speed and guest acceptance, and how to design a pilot with clear pass marks.

AssistCall research teamPublished Reviewed by Faruque Amin, VP of Sales and Marketing6 min read

On this page
  1. What drive-thru voice AI actually does
  2. What the public record shows
  3. Why the menu is the hardest part
  4. Where operators stand
  5. What a good pilot measures
  6. How to design the pilot
  7. Questions to put to any voice AI vendor
  8. What this means for independents
  9. Example: a fictional six-week pilot
  10. Frequently asked questions
  11. References
Restaurant Technology and AI: Drive-Thru Voice AI: What Pilots ShowArticle coverRESTAURANT TECHNOLOGY AND AIDrive-Thru Voice AI:What Pilots ShowAva for RestaurantsInsightsRestaurant Technology and AI: Drive-Thru Voice AI: What Pilots ShowArticle coverRESTAURANT TECHNOLOGY AND AIDrive-Thru Voice AI:What Pilots ShowAva for Restaurants

Key takeaways

  • Off-premises traffic reached 83% at limited-service restaurants in 2024[1], so the drive-thru lane is where speed is won or lost.
  • The most public test so far, McDonald's work with IBM in more than 100 restaurants, was shut off in 2024 after reported accuracy problems[2].
  • Operators' top AI worries are complexity, cost and customer acceptance[3]; a pilot should measure all three.
  • Set pass marks for accuracy, speed, handoffs and upsell before the first car, and keep a person on the headset.

What drive-thru voice AI actually does

A drive-thru voice system greets the car, listens to the order, matches it to menu items and modifiers, reads the order back, shows it on the confirmation board and sends it to the POS and kitchen display. Most systems also suggest add-ons and hand the lane to a crew member when they get stuck. The crew still takes payment, hands out food and fixes problems at the window.

The appeal is obvious. Limited-service restaurants depend on off-premises traffic, which rose from 76% of all their traffic in 2019 to 83% in 2024[1]. Speed matters to almost everyone ordering that way: 94% of consumers say speed is critical when ordering off-premises[4]. If a system can take orders consistently while the crew focuses on food and the window, the lane moves faster.

What the public record shows

Most drive-thru AI results are reported by vendors or by brands in investor updates, which makes independent evidence thin. The clearest public data point is a cautionary one. McDonald's tested an automated order-taking system with IBM in more than 100 restaurants, and a franchisee memo said it would be shut off by July 26, 2024[2]. CNBC reported, citing two sources, that the system had trouble handling different speech styles, which hurt order accuracy; McDonald's declined to comment on accuracy[2]. McDonald's said voice ordering would still be part of its restaurants' future and that it would decide on a long-term option later[2].

Read that as a lesson in difficulty rather than a verdict on the idea. Drive-thru audio is one of the hardest settings for speech recognition: engine noise, wind, children in the back seat, accents, people changing their minds mid-sentence and menus with dozens of modifiers. Systems that work well in a quiet demo can struggle in a real lane at noon.

When a vendor quotes an accuracy figure, ask three things: accuracy of what (complete orders or individual items), measured how (human review of recordings or system logs), and in which conditions (all hours, or only quiet ones).

Why the menu is the hardest part

Speech recognition gets the headlines, but most failed orders in a drive-thru come from the gap between how the menu is built in the POS and how people talk. Guests say a number two with no pickles and a Diet Coke, large, but the POS needs a combo, a size, a modifier and a drink swap in the right order. Limited-time offers make it worse: the marketing name, the nickname guests use and the POS button can all differ, and each new promotion has to be loaded before launch day, not after.

  • Build a spoken-name list for every item: the official name plus the two or three things guests actually say.
  • Map every combo and modifier path, including the ones crew handle by habit, such as splitting a combo across two people.
  • Load limited-time offers into the voice system in the same release as the POS, and test them before launch.
  • Decide what the system should do with off-menu requests: hand off at once rather than guess.
  • Re-test after every menu or price change, because a small POS change can break a spoken path.

Where operators stand

Interest is real but measured. In 2024, 16% of operators said they planned to invest in AI integration, including voice recognition[5]. By 2026, only 3.2% were implementing AI projects companywide, while 20% were running small-scale pilots[6]. The concerns that hold operators back line up with the risks in the lane.

Operators' top concerns about AIShare of US restaurant operators citing each concern, 2026
Operators' top concerns about AIBar chart. Complexity: 46.1%; Cost: 43.5%; Customer acceptance: 41.9%.ComplexityComplexity: 46.1%46.1%CostCost: 43.5%43.5%Customer acceptanceCustomer acceptance: 41.9%41.9%Operators' top concerns about AIBar chart. Complexity: 46.1%; Cost: 43.5%; Customer acceptance: 41.9%.ComplexityComplexity: 46.1%46.1%CostCost: 43.5%43.5%Customer acceptanceCustomer acceptance: 41.9%41.9%

Source: National Restaurant Association, 2026[3]

Show the data
ItemValue
Complexity46.1%
Cost43.5%
Customer acceptance41.9%

Customer acceptance is less of a wall than it might seem among younger guests. Half of Gen Z adults (50%) and 52% of millennials say they would consider ordering from an AI-generated video assistant[7]. Expect more friction from guests who are less comfortable with automation and from long, customized orders, so make sure the handoff to a person is instant and obvious.

What a good pilot measures

Decide every pass mark before the first car and measure the crew baseline in the same weeks at a control store.
MetricWhy it mattersHow to measurePass mark
Order accuracyWrong orders cost remakes, refunds and repeat visitsHuman review of a random sample of recordings against the ticket, dailySet before the pilot; compare with crew baseline
Completion without handoffShows how much work the system really takesShare of orders finished with no crew interventionTrack by daypart; expect lower at peak
Speed of serviceLane time drives throughputTimer from greeting to order confirmed, and total lane timeNo worse than crew baseline at peak
Handoff timeA stuck car is the worst outcomeSeconds from the system giving up to a crew member speakingA few seconds at most
Upsell rateOften the vendor's headline benefitAdd-on items per order versus baselineHigher without a drop in satisfaction
Guest complaintsAcceptance in practiceComplaints mentioning the voice system, per 1,000 ordersNo rise against baseline
Crew workloadThe real labor benefitCrew survey plus headset timeCrew say it helps, not adds work

How to design the pilot

  1. Pick a control store

    Choose a similar location without the system and measure the same metrics there over the same weeks.

  2. Clean the menu data

    Every item, modifier, combo and limited-time offer must be in the system with the names guests actually say. Most early errors are menu-data errors.

  3. Start off-peak

    Run the system at quieter dayparts for the first two weeks, with a crew member listening on every order.

  4. Review recordings daily

    Sample orders each day, score them and send the vendor a list of failures with timestamps.

  5. Add the rush only when ready

    Move to peak hours only after accuracy and handoff time meet the pass marks off-peak.

  6. Decide on evidence

    At the end, compare with the control store and the pass marks. Expand, extend the pilot or stop.

Keep a person on the headset throughout. The point of the pilot is to learn where the system fails without letting guests pay for it.

Questions to put to any voice AI vendor

  1. How is accuracy measured: complete orders correct, or items correct? Who checks, and how often?
  2. What happens when the system is unsure: does it ask again, guess or hand off, and how fast?
  3. How long does it take to load a new item or limited-time offer, and who does it?
  4. Which POS and kitchen display systems are supported in production, not only in testing?
  5. Are lane recordings stored, for how long, and are they used to train models for other brands?
  6. How is the system told about local conditions, such as a sold-out item or a closed fryer?
  7. What does the contract say if accuracy falls below an agreed level?

Ask for answers in writing and compare vendors on the same list. The answers about sold-out items and new menu items are often the most revealing, because they show how much daily work the system will add for your managers.

What this means for independents

Most independent restaurants do not have a drive-thru, but the lessons transfer to phone orders. Taking a full order item by item by voice is hard, and errors are expensive. A simpler model avoids most of that risk: answer the takeout call, handle questions, then text the guest your official online ordering link so they build the order themselves. Ava works this way for takeout calls, sending the link only when the caller agrees to a text. For AI uses that never touch the order itself, see AI review reply guardrails and AI sales and labor forecasting.

Example: a fictional six-week pilot

Example (fictional): a regional burger chain tested voice ordering in one store with a nearby control store. In weeks one and two, off-peak only, human review found errors mostly on combos with drink swaps and on the new spicy chicken, which guests called by three different names. After the menu data was fixed, completion without handoff rose and accuracy matched the control store off-peak. At lunch peak in weeks four to six, handoffs doubled and lane time was slightly worse than the control. The chain decided to keep the system for breakfast and late night only, retest peak after the next vendor update, and measure again.

The example shows the usual pattern: results vary by daypart, data quality matters as much as the model, and a partial rollout can be the right answer. Before any rollout, connect the data flows covered in connecting POS, reservations and ordering and track lane metrics in a weekly restaurant dashboard.

Frequently asked questions

Does drive-thru voice AI work?

It can, but results vary widely by conditions, menu complexity and daypart. The most public test, at McDonald's with IBM, was shut off after reported accuracy problems[2]. Judge any system on a measured pilot in your own lanes.

Will voice AI replace drive-thru crew?

Not in current pilots. Crew still handle payment, food, the window and every order the system cannot finish. The realistic benefit is moving order-taking work, not removing positions.

Do guests accept talking to an AI at the drive-thru?

Many younger guests are open to it: about half of Gen Z and millennial adults would consider ordering from an AI video assistant[7]. Acceptance falls when the system makes mistakes or the handoff to a person is slow.

What should a pilot measure first?

Order accuracy from human review of recordings, then handoff time and lane speed against a control store. Upsell results only matter once accuracy is solid.

References

  1. National Restaurant Association (2025). Report: Takeout, drive-thru, delivery are more popular than ever (opens in a new tab) ↑

  2. CNBC (2024). McDonald's to end IBM AI drive-thru test (opens in a new tab) ↑

  3. National Restaurant Association (2026). Research Insight: The Future of Restaurant Technology (press release, November 5, 2026) (opens in a new tab) ↑

  4. National Restaurant Association (2025). From Trend to Transformation: Off-Premises Dining Now Essential for Restaurant Consumers, Operators (opens in a new tab) ↑

  5. National Restaurant Association (2024). Serving Up Technology: New Data Shows How Tech Integration is Transforming the Restaurant Experience (Restaurant Technology Landscape Report 2024) (opens in a new tab) ↑

  6. National Restaurant Association (2026). Research Insight: The Future of Restaurant Technology (press release, November 5, 2026) (opens in a new tab) ↑

  7. National Restaurant Association (2025). From Trend to Transformation: Off-Premises Dining Now Essential for Restaurant Consumers, Operators (opens in a new tab) ↑

Book a demo

See Ava handle your calls.

Book a short demo. We’ll walk through your calls and workflow, and agree what Ava should take care of.

Book a demo