Drive-thru voice AI: noise, upsell rules, order accuracy and crew hand-off

What it is

A drive-thru voice agent greets the car at the speaker post, takes the order with modifiers and quantities, offers the upsell your rules allow, reads the order back, writes it to the POS so it appears on the kitchen and window screens, and tells the customer to pull forward. When it cannot proceed, because the audio is unusable, the request is outside the menu or the customer asks for a person, it hands off to the crew headset in one move with the partial order intact.

Everything about this channel is harder than the phone. The microphone is outdoors. The customer has an engine running, a window half down, wind, a radio and passengers. The agent's own voice comes back through the speaker post into the microphone. Orders are fast, slangy and full of mid-order changes. And the timer is running: a lane that slows by ten seconds per car loses cars.

Public press reports large deployments in the United States. Restaurant Dive reported in July 2026 that one chain had voice AI in more than 890 restaurants across 38 states, after its parent announced in a March 2025 filing a rollout to 500 restaurants across four brands, and after the same parent slowed the rollout in August 2025 following customer complaints and prank orders. The same publication reported in May 2025 that another chain had the technology in over 160 restaurants with a goal of more than 500 by the end of the year, and in January 2026 that a third chain's trial with its technology partner had ended. Read those as reported, not as proof that any product will work on your lane.

Who buys it

• Quick-service chains and large franchisees with drive-thru as the majority of sales, who need lane-by-lane accuracy and timer data, central menu and upsell rules, and a procurement process that includes the headset and timer vendors.
• Regional chains and multi-unit franchisees piloting in a handful of stores before a system-wide decision, usually driven by labour availability at the order point.
• Drive-thru hardware and POS vendors' customers choosing between a bundled agent and a standalone one on the same lane test.

Budget owner: the chief operating officer or head of restaurant technology at chain level; the franchisee principal at store level, with the brand's technology standards team approving the integration.

KPIs

Measure the crew first. One month of order accuracy from remakes and refunds, timer order-point duration, and attachment rate on the items you care about, by lane and daypart. The agent is judged against that, not against a vendor slide. Then track accuracy at the window, intervention rate, time at the speaker, upsell within rules, hand-off time and voice-to-voice latency in the lane.

Two traps. Accuracy at the speaker is not accuracy at the window; count remakes. And an intervention rate that falls because the agent stopped handing off is an accuracy problem in disguise. Read the two together, by hour, because the dinner rush and the late-night lane behave differently.

Demo script

Insist on a lane test at one of your stores, on your microphone, speaker, headset and POS, with your menu and your upsell rules loaded. Use a real car.

1. Greeting under engine noise. Pull up with the engine running and the window half down. Pass: a short greeting, an AI disclosure if your brand requires one, and an open question. Fail: a long scripted welcome or a greeting that triggers again on the engine.
2. Fast order with modifiers. "Two number threes, one no onions, one with extra pickles, a large drink and a kids' meal with apple slices." Pass: each modifier on the right item, read back in order.
3. Quantity change within ten seconds. "Actually make that three number threes." Pass: one line updated, no duplicate.
4. Background noise trap. Turn the radio up and have a passenger talk for twenty seconds without ordering. Pass: the agent waits and adds nothing. Fail: it adds an item, or says "I didn't catch that" repeatedly.
5. Interruption over the read-back. Talk over the agent's read-back with "no, no pickles on the second one". Pass: it stops within a word and fixes the modifier. Fail: it finishes the read-back or treats its own echo as speech.
6. Upsell by rule. Decline the first suggestion. Pass: one offer, remembered as declined, no second attempt. Fail: a repeated or off-rule offer.
7. Ambiguous item. "A chicken sandwich" where you sell three. Pass: it asks which, with the names. Fail: it picks one.
8. Eight seconds of silence after the agent asks "anything else?" Pass: one short prompt, then a read-back and total. Fail: a hang-up or a repeated greeting.
9. Out-of-scope request. Ask for an item from another brand, then complain about the last visit. Pass: a plain "we don't have that" and a short hand-off offer for the complaint. Fail: an invented item or an apology loop.
10. Hand-off. Mumble a long custom order the agent cannot parse. Pass: a crew voice on the speaker within a few seconds with the partial order on the crew screen, and the customer is not asked to start over. Then watch the order on the POS and the window screen.

Score each trap pass or fail. Added items during the noise trap, a repeated upsell, or a hand-off that loses the partial order are hard stops. A vendor who wants to demo from a recording in their lab has not passed the demo.

Compliance notes

A customer at the speaker post initiates the interaction, so the TCPA consent rule in the United States is not in play unless the deployment adds outbound contacts, in which case prior express consent and the 8 a.m. to 9 p.m. window apply. Recording is the live question: many drive-thru audio systems record for quality, roughly a dozen states require all-party consent, and a sign at the order point is the usual notice; your counsel decides whether the agent should also say it. Several states have bot-disclosure laws, so disclose by default. In the United Kingdom, recording the lane is personal-data processing under UK GDPR and needs notice and a documented lawful basis; the 14-allergen information duty applies to drive-thru food as to any other, so the agent answers allergen questions from data and does not guarantee. In the European Union, the AI Act's Article 50 duty to tell people they are talking to an AI applies from 2 August 2026, which at a speaker post means a disclosure in the greeting. In Australia, state surveillance-devices laws differ, so post and announce recording. Informational, not legal advice; the compliance rows on this page carry the sources.

Build or buy

Buy. This is the one restaurant use case where building is rarely sensible: the acoustic front end, the echo path through the speaker post, the headset integration and the lane-timer plumbing are specialist work, and the public record shows that even well-funded chains have paused or ended rollouts. Your leverage is the pilot design: your lanes, your menu, your crew's measured accuracy as the baseline, a noise trap in every demo, and an exit clause tied to accuracy at the window and intervention rate by month three.