How to pilot AI at one drive-thru lane

Pilot AI ordering on one lane, for a fixed period, against numbers you recorded yourself before it starts. Write down the pass/fail rules with the vendor first. Then test the crew takeover and your hardest menu items with staff playing guests before a real driver hears the system.

A pilot without a baseline ends in an argument about whether the lane felt faster. The worksheet below gives you the measures, the scripts and the decisions to make in advance, so the review at the end is a comparison of numbers.

Which lane and dayparts should you pick?

Pick one lane you can watch closely. On a double lane, run the AI on one side and keep the other as your comparison for the same hours. On a single lane, compare against your own baseline weeks.

Choose at least three dayparts to review: your busiest rush, a slow period, and late night if you run one. Rush tests noise and speed. Slow periods show whether guests will talk to the system when a person is clearly free. Late night may leave fewer crew to take over.

Avoid starting the pilot in the same week as a menu change, a new promotion or a price rise. Those change order times on their own, and you will not be able to separate the effects.

What should you measure before the AI starts?

Record two weeks of the same dayparts you plan to test. Use your timer and POS reports where you can, and a paper tally where you cannot.

Measure	How to define it	Where to get it
Order time	From the greeting to the guest confirming the order	Drive-thru timer, or a stopwatch on a sample of cars
Service time	From placing the order to leaving with food	Drive-thru timer report
Cars per hour	Cars that reached the window, by hour	Timer report or POS drive-thru order count
Drive-offs	Cars that left the lane without ordering	Timer report if it tracks them; otherwise a tally
Remakes and comps	Orders remade, refunded or comped because of an order-taking error	POS void and comp reports, with a reason code
Upsell take rate	Offers accepted divided by offers made, for the items you choose	Tally sheet at the speaker; POS item mix for the offered items
Crew interventions	Orders where a person had to step in (AI weeks only)	Vendor report, checked against a crew tally

Timers already log much of this. HME says its ZOOM Nitro timer offers POS integration and a bottleneck ticker, and PAR says its drive-thru timer tracks service times, bottlenecks and drive-offs. Check that your timer settings stay the same through baseline and pilot.

For outside context, Intouch Insight's 2025 study with QSR Magazine sent shoppers through drive-thrus at 13 national chains in June and July 2025. It defined service time as placing the order until leaving with food. The average was 4 minutes 15 seconds, and 120 AI-enabled orders at three chains averaged 3 minutes 53 seconds. Your own baseline matters more than either figure.

Count upsells the same way in both periods. If the AI offers an add-on to every guest and your crew offered it to some, the take rate and the number of offers both change. For more on setting up an offer test, see the upselling pilot guide.

How do you test fallback and staff override?

Run these with staff in a car before the pilot opens, then repeat two of them in the first live week.

1. Start an order and have the crew take over halfway. Check who hears the guest and whether the half-built order survives.
2. Have the guest ask for a person. Time how long until a crew member answers.
3. Give the AI an order it cannot understand. Count how many tries before it hands off.
4. Turn the AI off for the whole lane, then back on. Check the lane works normally in between.
5. Ask the vendor how to simulate an internet or POS outage, and what the guest hears when one happens.
6. Change the crew member on duty and repeat the takeover. The second person must know the control too.

Ask the vendor to name the takeover control and show it working. HME says its NEXEO Pro headset platform lets a crew member take over the order at any point. Maple's drive-thru page describes a one-press kill switch that puts the lane back to normal, and says guests can ask for a person at any time. Test whichever control you get on your own lane.

Which menu edge cases should you script?

Write each script down and run it the same way every time. Check the POS ticket, the price and the kitchen screen after each one.

Case	Sample script	Pass if
Modifiers	"Cheeseburger, no onion, extra pickle, add bacon"	All three changes and the bacon charge are on the ticket
Combos	"Number two, large, swap the drink for a shake"	Size, swap and any upcharge match your POS rules
Mid-order change	"Actually, make that two, and drop the fries"	Quantity and removal are correct at read-back
Out of stock	Order an item your manager marked out	The AI says it is out and offers an alternative you sell
Off-menu item	Ask for a dish from another chain or last season	The AI declines it clearly and keeps the order going
Language	Place a full order in each language your guests use	Ticket matches, and the guest can switch back if needed
Several voices	Driver and passenger both order	Every item lands once, with no duplicates
Large order	Eight or more items with changes	Read-back is complete and the total matches

Customization is where AI orders missed most in the Intouch study, so weight your scripts toward modifiers and combos. To build the item list, the menu preparation worksheet covers required choices and modifier rules for phone ordering; the same item list works for the lane.

Should you tell guests they are talking to AI?

Decide before launch, with your lawyer and the vendor. Settle what the greeting says, whether a sign at the post explains it, and what happens if a guest asks.

Ask the vendor whether it records audio, how long it keeps recordings, and whether it uses them to train its system. Privacy law varies by state. In June 2021 an Illinois customer sued McDonald's, alleging its drive-thru voice AI collected his voiceprint without consent under the state's biometric privacy law. That was an allegation in a complaint, and it shows the question your counsel should answer before the pilot.

What pass/fail rules should you agree first?

Fill in the blanks with the vendor and sign it before the first live car. Use your baseline as the comparison, and set the numbers yourself.

1. Pilot length, lane, and dayparts reviewed: ________.
2. Order time and service time must not be worse than baseline by more than ________ seconds.
3. Cars per hour at peak must be at least ________.
4. Remakes and comps from order errors must not exceed ________ per 100 orders.
5. Crew interventions must fall to ________ per 100 orders by the final week.
6. Upsell take rate, counted the same way as baseline: target ________.
7. Drive-offs must not rise above baseline by more than ________ per day.
8. Any failed takeover test stops the pilot until the vendor fixes it.
9. Who reviews the numbers each week, and who can end the pilot: ________.
10. Cost of the pilot, removal of equipment, and what happens to recordings if you stop: ________.

If you have not yet decided whether the lane or the phone is the better first test, start with the phone vs drive-thru guide. Maple's drive-thru unit is in early access and priced per unit; ask about a lane trial on the drive-thru page.

Published by Maple. This AI-assisted worksheet combines an original pilot method with a named industry study, maker timer documentation, trade press and Maple's published drive-thru scope. It sets no pass mark for you and does not report results from any Maple or other vendor pilot.