Fast Food’s AI Drive-Thru Is an Upsell Engine, Not a Layoff Plan

WIRED’s reporting on AI drive-thrus reads like a labor story: robots arrive, workers leave. Look at the metric the chains are actually tracking and it is a merchandising story instead. That distinction decides whether putting a machine in front of your own customers is worth the money.

“What chains are doing is using robots to convince people to pay more for fast food. Intouch Insight’s survey found that AI drive-thru agents use ‘suggestive selling’—upselling—71 percent of the time, compared to an industry average of 58 percent.”

— Kate Taylor, WIRED

Our take

The number that matters here is not headcount. It is the thirteen-point gap in suggestive-sell rate, applied to every single car in the lane. A store running 400 cars a day is getting roughly 52 additional upsell attempts it was not getting before — the ones a tired human skips at 9pm on a Tuesday. Nobody bought a cashier replacement. They bought a salesperson who never gets bored of asking the question.

The employment data in the article supports that reading. White Castle’s headcount held steady through its rollout, and Taco Bell says locations running voice AI have higher employee retention. That makes sense to anyone who has worked a window. Drive-thru labor is six tasks braided together, and automating the one that is easiest to script does not remove the person doing the other five.

Where we would push back on the framing is that 97 percent satisfaction figure. It measures a transaction, not a relationship. Ask someone immediately after a correctly filled order and almost any process scores well; ask whether they would drive past a competitor for that brand in a year and you get a different survey. The number also comes off a base of chains that spent three-plus years and serious capital getting there — Taco Bell working with Nvidia and Omilia, White Castle four years into SoundHound. The base rate for everyone else is the wreckage the article documents: Del Taco rolling back its test, McDonald’s selling McD Tech Labs to IBM and then pulling the plug entirely. Adoption is still around 6 percent of US locations. That is not a solved technology. That is a handful of well-capitalized chains finishing a long, expensive project.

The second-order effect the article does not chase: fast food is absorbing the awkwardness cost of conversational commerce on behalf of every local business that comes after it. Once “would you like to add a drink” from a machine feels ordinary at Taco Bell, it stops feeling invasive on a dentist’s booking line or an HVAC company’s after-hours phone. Six percent adoption today is the beachhead. Small businesses get to walk in behind it without paying for the consumer education, and that window is open now rather than later.

For a Southern Nevada service business the real analog is not the ordering window, it is the phone. Nevada’s minimum wage stayed at $12.00 an hour going into 2026 while 19 other states raised theirs, so the wage-arbitrage argument is weaker here than the vendor pitch decks assume. Fully loaded, a part-time front desk hire is still a four-figure monthly line item, and that person does not work Sunday evening — which is exactly when a homeowner with a dead air conditioner starts calling down the Google results. The math that made drive-thru AI work was upsell rate. The math that makes phone AI work for a contractor is answer rate on calls currently going to voicemail. Different justification, different measurement.

The risk nobody prices into these projects is disclosure. California has required it since 2019: Business and Professions Code section 17941 makes it unlawful to use a bot to mislead someone about its artificial identity in order to incentivize a purchase, with a clean safe harbor if you simply disclose that it is a bot. A voice agent whose primary job is upselling sits squarely in that fact pattern. Utah’s amended AI Policy Act requires disclosure when a consumer clearly asks. The EU switched its version on this month. Write the disclosure line into the script on day one; retrofitting it later costs more and reads like you were hiding something.

And the actual failure mode is not the bot getting an order wrong. It is the recovery path. The viral clips in the article — 260 McNuggets, bacon on an ice cream order — went viral because nobody intervened. Set one hard rule: any repeat, any confusion, any raised voice, and a human is on the line inside ten seconds. That rule is the entire difference between a funny TikTok and a one-star review that outranks your homepage.

What this means for your business

• Justify the project on revenue lift, not payroll. Model it the way the chains actually did: incremental attach rate times average ticket times monthly volume. If that number does not clear the subscription cost, labor savings will not rescue it.
• Automate the follow-up before you automate the front door. Most local businesses lose more money to unreturned voicemails and cold form fills than to order accuracy. Fix the leak you already have first; it is cheaper and it converts faster.
• Write the disclosure line before you write the greeting. One sentence, up front, plus an obvious path to a human. It satisfies California today and the states drafting copycat bills next session.
• Define the human handoff trigger and measure it weekly. Track escalation rate and time-to-human as first-class KPIs, not diagnostics. A rising escalation rate is your early warning months before it shows up in reviews.
• Pilot on your slowest shift, one location, 30 days. Compare answer rate, booked appointments, and review sentiment against the same window last year. Kill it fast if the sentiment moves the wrong way.

The drive-thru is the loud version of a quiet shift arriving at every business with a phone number. Read Kate Taylor’s original reporting at WIRED, then go look at your own call logs.

Originally published on Digital Media Marketing Technology.