What You’ll Find This Week

In July 2024, Taco Bell's parent company announced that its drive-thru voice AI was headed to hundreds of restaurants. Its chief innovation officer said two years of fine tuning and testing had made the company "confident in its effectiveness." Thirteen months later, the AI was taking orders at more than 500 Taco Bells. Taco Bell spent two years testing the AI, and it still took a rollout to hundreds of restaurants to learn how the system handled the reality of hungry humans screaming for Crunchwraps at 2am.

This week, I'm looking at why a pilot built to win approval can't tell you whether something will work at scale, what a study of Wisconsin farmers says about this year's AI rollbacks, and why every pilot needs a written answer, before it starts, to whether it's there to prove something or to persuade someone.

Don’t Miss Our Latest Podcast

This Week’s Article

Taco Bell didn't rush its drive-thru AI. Yum Brands, which owns Taco Bell, KFC and Pizza Hut, started small, with five California stores in early 2024. Taco Bell said customers liked it, and that was enough to keep going. By the end of July, the AI was taking orders in more than 100 drive-thrus across 13 states, and Yum was ready to go big.

The timing took some nerve. The month before, McDonald's had pulled the plug on its own voice AI test at about 100 drive-thrus, after TikTok videos of its mangled orders made the rounds.

By August 2025, Taco Bell's AI was in more than 500 drive-thrus, and its customers were posting their own videos of the humorous and often frustrating failures by the system. In one, the AI got stuck in a loop, asking a man who had already ordered a Mountain Dew what he'd like to drink. In another, which took the internet by storm, a customer resorted to ordering 18,000 cups of water in an effort to force the AI to pass him to a human.

Dane Mathews, Taco Bell's chief digital and technology officer, told The Wall Street Journal the company was having an "active conversation" about where the AI belongs, and that at busy restaurants with long lines, a human might be the better choice. Let that sink in…the exact moment that AI is designed to streamline is the one place where it failed miserably.

After two rounds of expansion, the problem Taco Bell's tech chief described to the Journal was a busy restaurant with a long line, which is about as ordinary as drive-thru conditions get.

By July 2026, the AI was taking orders at more than 890 restaurants in 38 states, running on technology from Omilia, a company that builds voice AI for contact centers and drive-thrus. Mathews said Omilia's platform "has proven itself at scale in select U.S. restaurants." Note the word select. Two years of testing and nearly 900 restaurants later, Taco Bell is still working out which of its stores the AI belongs in.

I've leaned on Everett Rogers here before. Last year, in Sales Is the Wrong Lens for GTM, I used his research on how ideas spread to argue that adoption depends more on how a product is built than on how hard anyone pitches it. His book, Diffusion of Innovations, also has a passage about testing that every innovation team should read.

Rogers called pilots “demonstrations,” and his go-to example came from the farming world. A county extension agent, the government farm adviser whose job was to get new farming methods into farmers' hands, would split a farmer's field, fertilize one plot and leave the other alone, then invite the neighbors over at harvest to compare yields. A corporate pilot follows the same script. A company tries the idea in a few locations, then invites the executives over to judge the results.

Rogers said pilots/demonstrations come in two flavors, and they need opposite conditions:

The experimental kind also gets a definition of success most companies would find strange. Rogers counted it a success if the innovation got an adequate evaluation, regardless of "whether this evaluation is positive or negative." A test that comes back with a clear no did its job.

Rogers also wrote, "Often…a single demonstration is expected to fulfill both functions."

That sentence describes most corporate pilots. The same pilot is supposed to tell the company whether to scale and to win the budget for scaling. The sponsor needs a yes for the budget and a yes for their own reputation, so when the two jobs conflict, the sponsor usually picks persuasion. And a pilot run to persuade rarely comes back negative. When it comes back green, all leadership has learned is that the sponsor can make a good case; whether the product holds up once it's in every location is still anyone's guess.

Most pilots run under conditions the product will never see again. The team asked for the project. Leadership is watching. There are few enough locations that someone can drive over and fix a problem by hand. None of that survives a mass rollout and contact with the real world at scale.

Once Taco Bell's AI was in hundreds of restaurants, customers complained online about being asked again and again to add a drink, and the 18,000-water order went viral. Five stores don't draw that kind of crowd, so a small pilot never meets the critical mass of customers necessary to draw ire across the internet.

I've given a version of this advice before. In May 2025, I advised innovators to keep legal, brand and compliance out of their early experimentation if those teams would slow it down, and not to seek approval "until you've got something defensible." Rogers would have recognized the sequence: test quietly first, then go win support with what the test found.

The trouble starts when a company rolls the test and the pitch into one pilot.

Taco Bell has plenty of company. Sinch, a Swedish company that sells messaging and voice infrastructure, surveyed 2,527 senior decision-makers at large companies in ten countries early this year about the AI agents they use to talk to customers. Sixty-two percent had one live. Seventy-four percent had already rolled back or shut down an agent after launch because of what Sinch calls a governance failure.

The companies with the most mature governance rolled back even more often, at 81%. "The most advanced organizations aren't failing less; they're seeing failures sooner," said Daniel Morris, Sinch's chief product officer. In Rogers' terms, these companies are running their real test after launch, with paying customers, which suggests the pilot projects that came before launch never tested much of anything. Their monitoring works, but only after customers have hit the failures.

Rogers documented the same pattern on farms in the 1960s. He cites a study of Wisconsin farmers that found about as many farmers abandoned an innovation in a given year as adopted one for the first time. The ones who quit because they were disappointed tended to be the farmers less careful about generalizing "the results of an innovation's trial to its full-scale use." Taco Bell went from five stores to nearly 900, and it's still deciding where the AI shouldn't be.

If your company has never rolled back an AI deployment, don't celebrate yet. Sinch's numbers suggest your success might be because nobody is watching closely enough to catch the failures.

Before anyone books a store or a demo, write down the answer to one question.

Is this pilot here to prove something, or to persuade someone?

You'll need both kinds. Last year I told you to ship a visible win, "a battle people can see," because a visible win makes the bigger vision feel real. I stand by that, and plenty of innovation budgets exist because someone shipped one. The damage happens when the company treats a visible win as proof the thing will work across hundreds of locations.

A pilot built to prove something has these features from day one:

Kill signals for pilots are rarer than they should be. In a Gartner survey of 2,002 CIOs and technology executives, only 31% said their companies had guidelines for when an AI pilot should die before production. The rest have nothing in writing to stop one.

A persuasion pilot can be as public and polished as you like. Pick the best stores, put a believer in charge and invite every executive who'll come. Just don't let anyone cite its result when the company decides whether to scale.

If a project needs both, run two pilots, and run the one that proves something first. Last December I wrote that you should define the path to scale before a pilot begins. That path only leads somewhere if the pilot at the start of it actually tested something.

"We're learning a lot, I'm going to be honest with you," Mathews told the Journal in 2025. He deserves credit for saying it out loud. As of July, the AI was taking orders in nearly 900 restaurants, so whatever Taco Bell is still learning, it's learning with real customers on the other end of the speaker.

Go find the plan for the last pilot your company approved. If it doesn't say whether the pilot was meant to prove something or persuade someone, assume it was built to persuade, and treat its result like a sales pitch before anyone uses it to justify a rollout.