Column: Count the Repeats Before You Pay for Drive-Thru AI

There is one number from this year's drive-thru data I keep coming back to: 29.1%.

That is the share of mystery shoppers at voice AI drive-thrus who had to repeat part of their order, according to Intouch Insight's 2026 drive-thru study. At the same brands with a person on the headset, it was 20.4%.

To be fair to the machines, the rest of the report reads like good news for them. Order accuracy at AI locations hit 90.8%, up from 82.5% a year earlier, and beat the 86.7% at the same brands' human-staffed lanes. AI offered an upsell on 79.2% of visits, against 59.2%. Overall satisfaction ran 94.2% to 91.7%. Total visit time was a wash, 5:14 against 5:15: shoppers waited about 11 seconds less and spent about 10 seconds more at the speaker, per the full report. One caveat Intouch states plainly: its 120 AI visits at Burger King, Taco Bell and Wendy's were selected, not randomized.

So why fixate on repeats? Because the franchisee pays for them.

The rollouts are not waiting for anyone's spreadsheet. McDonald's says most U.S. restaurants will have its Archy voice ordering by the end of 2027, part of a modernization bill operators are already eyeing. Taco Bell has voice AI in more than 900 drive-thrus, and Yum wants more than 1,500 by year end. Bojangles went systemwide in October.

The pitch to owners is labor. McDonald's says Archy can free at least 50 labor hours a week per restaurant. But as Restaurant Technology News pointed out this week, the public figures leave open how often crew members step in. A fix at the speaker costs seconds. An error caught at the pickup window means a remake while the cars behind it wait. Wendy's measured its own pilot by how often the tool finished an order without crew help, which is a different thing from an order filled correctly.

Here is my position. No franchisee should accept a voice AI rollout, mandated or optional, without store-level reporting on three numbers: how often a person takes over, how often guests repeat themselves, and how many corrections reach the window. If the brand cannot produce those by store, the 50-hour figure is a projection, not a result.

The fair pushback is that repeats are the price of a system that is still learning, and the curve is bending the right way, from 34.2% last year to 29.1% now. I accept that. Guests also liked the AI visits more overall. But only 63% of shoppers called the AI interaction easy, and 21.7% called it difficult. One in five is a lot of difficult at a lunch rush.

What I don't know is whether owners of quieter stores can actually bank the labor savings when someone still has to watch the lane. Nobody has published that yet, and I'd like to see it before the invoices arrive.

Where we sit, for the record: Revscale, which publishes this site, sells a Voice Agent that answers inbound phone calls for franchise and multi-location businesses, not drive-thru orders. It captures caller details and intent, transfers to a person when needed, and records and transcribes every call. The handoff is built in on purpose, because it is the moment an operator most needs to be able to see and count. Whatever tool an owner ends up with, on the phone or in the lane, the transcript is how you check the brand's math.

Accuracy is the headline. Repeats are the invoice.

Revscale Media is published by Revscale.