Last fall, Carl Addison showed his Starbucks coworkers a magic trick.
As a shift supervisor at his Seattle-area café, he was responsible for the store’s twice-a-week inventory count. Starbucks had introduced a new AI tool in September to automate the process. Called Automated Counting, it used an iPad camera to identify and tally items on the storage shelves, turning an hour-long job into one that was supposed to take as little as 10 to 12 minutes.
Addison discovered that when he aimed the iPad into the shiny steel fridge that holds the oat milk, even if he was careful, the camera picked up a reflection, and the app counted the reflected cartons, turning 5 real oat milks into 10.
It would have been amusing if baristas weren’t being warned that hand counts were no longer acceptable. And Addison’s fridge wasn’t the only one haunted. Within weeks of the tool’s rollout, it was going rogue, baristas from coast to coast tell me, marking their milks as the wrong type, swapping syrups, and in at least one photo I saw, counting the trash can as food.
Megan Queen, a store manager in Graham, Texas, had the opposite problem. Instead of conjuring inventory, her Automated Counting tool kept making items vanish. The rural café, an hour and a half outside Fort Worth, had unreliable internet. When Wi-Fi dropped mid-count, the app’s progress was wiped. Her shift supervisors counted by hand instead, only to be told that the company was now treating manual counts as no count at all.
Automated Counting had been deployed rapidly, reaching all 11,300 company-operated cafés by the end of September. Nine months later, the tool—which insiders told Fast Company may have cost north of $10 million over several years to develop and deploy—was eliminated overnight.
Along the way, baristas around the country say, they were left in the dark, then sometimes blamed for the AI’s glitches. Milk and beverage items have now returned to being counted and recorded the way everything else in the store is: with the human eye and pen and paper.
Starbucks, which declined to make executives available for this story but did provide a statement, characterizes the outcome as an example of its test-and-learn culture functioning properly: “That is what innovation looks like at Starbucks: listening, learning, and adapting.”
Automated Counting, after all, was one piece of the company’s much larger AI push, which included tools like Smart Queue, an order-sequencing engine that Starbucks says helped get customer wait times to under four minutes.
Technology has been a key part of the turnaround plan, known as “Back to Starbucks,” launched by CEO Brian Niccol in 2024. And there have been recent signs pointing to the strategy’s success: Last quarter, sales at existing stores grew 6.2% globally, reversing seven quarters of decline.
Yet what happened with Automated Counting may also represent the most sweeping rollback of workplace AI yet executed in corporate America. It offers a lesson for any company racing to deploy AI at scale: However promising a new tool may look from the boardroom, it’s only as good as it works in the field. Front-line workers are often best positioned to observe where it saves time and where it adds friction. Communication can be as vital as the tech itself.
Drawn from interviews with dozens of Starbucks workers and managers around the country—and members of the small team that built the AI tool itself—this is the full account of how the company’s foray into AI-assisted inventory, Automated Counting, collapsed.
AI was inescapable at Starbucks’s 2025 Leadership Experience in Las Vegas, the three-day event where Niccol rallied 14,000 store managers around his Back to Starbucks turnaround plan. “Not just a reset,” is how the company explained it, but “a recommitment to who we are when we are at our best”—aiming to restore the personal neighborhood-coffeehouse feel while cutting $2 billion in costs. A key aspect was using AI to reduce friction for workers. Automated Counting would liberate partners (as Starbucks employees are known internally) from a labor-intensive task so they could get back to making drinks and connecting with customers.
As part of the plan, Starbucks has also invested more than $500 million to increase worker hours in stores, tripled paid parental leave, and pledged to put an assistant store manager in most corporate-run cafés by the end of this year.
Automated Counting would harness first-of-its-kind augmented reality technology from a Seattle-area startup called NomadGo to count coffee bags, milk, syrups, and other beverage items. Employees told me their bosses returned from Las Vegas stoked, relating how the tool could make inventory count up to eight times faster, potentially saving them 90 minutes per week.
At cafés, baristas took turns doing the prep work: repositioning items on storage shelves to be fixed distances apart and stacking all inventory so it was upright and aligned in nice, neat rows—a mass Marie Kondo-ing of thousands of storage rooms. The app also required a new-generation iPad Pro; locations running older models had to get a new one, which retailed for around $1,000.
NomadGo CEO David Greschler and his 30-person company had created the technology to address a real problem in retail: Inventory has always been a tedious job, typically done by a relatively senior employee as the sun’s rising or long after it has set. In NomadGo’s controlled tests, the platform reported a 99% accuracy rate.
For NomadGo, Starbucks was a “white whale” of a client, former employees tell me. Not just because Starbucks was a high-profile catch, but because the coffee giant was eyeing a systemwide launch, enabling the Seattle area-based NomadGo to, as Greschler told me, “stress-test and improve” its product in the wild. Retail has become a public laboratory for such experiments, and there have been several high-profile debacles. Last August, Taco Bell sidelined a drive-through order bot that seemed adept mainly at generating viral TikToks. McDonald’s did the same with its IBM-powered “automated order taking” test after the system added butter packets to ice cream orders and ran up nearly $300 of McNuggets on one car’s bill while the customers filmed themselves shouting “Stop! Stop!” as the chicken kept piling up.
Nobel Prize winner Daron Acemoglu and Yale economist Pascual Restrepo call this “so-so automation”: It’s not superintelligent robots taking over our workplaces. Instead, reliably mediocre ones are being installed that displace human work while annoying us all.
But sometimes errors are worse than so-so. Pizza Hut is being sued over an AI delivery system that franchisees allege caused $100 million of lost business. Target unveiled a “Help AI” chatbot to assist employees only for it to tell a store worker to confront an active shooter with a baseball bat.
At an industry conference during the rollout of Automated Counting, Greschler marveled at the scale of Starbucks’s operation, noting that people usually hear about AI projects that “are, I don’t know, 3 stores, 20 stores, 50 stores.” The Starbucks AI inventory manager beside him on the panel added that Automated Counting was proving to be a hit with workers because “when you hit complete, you get confetti. Partners love the animation associated with that.”
The party wouldn’t last long. Automated Counting was struggling to take stock of coffee for the espresso machines (which, to be fair, comes in lumpy 5-pound silver packs). It was confusing caramel syrup with cinnamon dulce. Bags of Starbucks’s most popular brewed coffee—Pike Place—got clocked as syrup.
Nothing seemed to trip it up like milk. A California worker told me that his store’s iPad treated all the whole and 2% milks as nonfat. Another said that after a milk brand changed its labels, the app quit recognizing all milk for a few weeks.
At her café in Alabama, Sara Gattis would stare at five jugs while the iPad counted four. “For almost every row, I was having to go in and change it by hand,” she says. The count was taking longer than before, and her café had enough operational chaos already: It had opened in late August, the day before Pumpkin Spice Lattes launched.
When Gattis saw baristas on Reddit describing a button inside the app that would briefly let a hand count override the AI, she grabbed her store’s iPad, found it, and showed her coworkers. “None of us went back to the AI,” she says. The crew didn’t receive warnings, though others who hacked their iPads have claimed online that they did.
Workers who tried to report their store’s tech problems up the corporate ladder tell me they got nowhere. “Pretty much radio silence until the day they announced, ‘Effective this week, we will no longer be doing Automated Counting,’” Addison says. “For that nine-month period, I didn’t receive any support, any updates.”
A Houston shift supervisor who asked me not to use their name says they reported each problem they encountered on the company’s internal messaging platform, MyDaily. “It would miss a whole section, or it wouldn’t even turn on,” this person says of Automated Counting. But their manager told staff that using AI was mandatory, compliance was being tracked, and the company was “threatening disciplinary action” against those who didn’t at least open the app.
Some took their complaints to Starbucks’s own channels. The company runs a TikTok for partners. On unrelated posts, baristas began leaving comments like, “Hey! Shift here! We hate the AI!!” and “It actually takes more time than just counting.” The latter drew a reply from the account itself: “When tested in coffeehouses, partners shared feedback that once they got familiar with the new tech, it made counting much faster!”
Days after Automated Counting rolled out last September, baristas started seeing odd things happening with the ordering process. Starbucks was known to have had issues with product shortages, a problem that has dogged the chain since the pandemic—and that the last three CEOs have all tried to fix. Suddenly, however, there were surpluses.
“I had more than my fair share of nights where I was filling up bag after bag of sandwiches and pastries, and nobody could explain why we were shipped these, because they weren’t something anybody remembered ordering,” Addison says. The uneaten food he gathered got donated or tossed.
I saw a picture from a Connecticut Starbucks of 50 croissants, two dozen bagels, and a stack of cookies that the system over-ordered, piled onto the counter. The store later discarded nearly $700 worth of food items in one day. Another photo, from a New Hampshire store, showed 123 croissants and 69 slices of banana walnut and lemon loaf being discarded. On Reddit, a worker posted an image of 15 boxes of uneaten bacon-and-egg breakfast sandwiches, begging: “Lord please make it stop.”
Back in Texas, store manager Queen says her store’s reordering process got stuck in a loop nobody could stop. “For weeks, we were getting a case of pepper packets and a case of salt packets,” she recalls.
By now, “missed counts” were coming up in her check-ins with her district manager. When inventory isn’t logged properly, the system flags the job as incomplete—a lapse Starbucks ties to product shortages. “That was one of the KPIs that was tracked so closely that if we missed it, it was a huge red flag,” Queen says.
Screenshots she shared with me of alerts known as “missed actions” from that time period—which she was careful to scrub of any sensitive or identifying details—date the problems back to mid-September. Counts that her store did by hand because the AI wasn’t working appear underneath “missed actions.” In one message to her boss that she had also sent me a screenshot of, she asked why orders tied to her store’s inventory counts kept defaulting to “autoship,” a fail-safe triggered when no count exists. The screenshots show no replies, even after she reported that she was handling the mess herself. (The district manager did not respond to an interview request.)
Queen estimates that her store racked up at least 10 of these responses. In meetings, she says, her boss would say, “That’s a lot of counts to miss, Megan,” then point at a piece of paper and say: “I have the data right here.” Queen says she was “encouraged to write up” the workers responsible, adding that eventually, “Starbucks was like, ‘You’re not writing people up, you aren’t managing your store right, you aren’t managing your employees right, they’re lying to you.’”
By April, she says, she was dealing with other issues. Her district manager was leaning on her to keep her staff off the overtime rolls by working the hours herself. Then, she claims, Starbucks took away her performance bonus. She soon announced she was quitting and dropped off her laptop and keys. Weeks later, she says, “I see this itty-bitty headline—‘Starbucks quietly retires AI’—and I lost it.”
Starbucks’s decision to end Automated Counting had blindsided NomadGo, too, though employees at the startup had been aware of issues with the technology. On April 3, CEO Greschler says, Starbucks informed NomadGo that it was pulling the plug. “It was a complete surprise to us,” he says. “But there’s nothing you can do when leadership and strategy change.” Within days, “basically the whole company had to be let go,” a former NomadGo employee told me—including the technical staff who ran the Starbucks system. (NomadGo still works with clients, including Burger King franchises.)
In cafés, Starbucks kept the AI counts going for another six weeks. On May 18, workers discovered a line buried in their weekly employee newsletter. “Starting today, Automated Counting will be retired,” it read, instructing them to rip the QR codes off shelves and “own your inventory” again.
For months, baristas had gravitated to the Starbucks Reddit channel to troubleshoot and vent about Automated Counting, with threads sometimes going hundreds of comments deep. Now it filled with celebratory posts like, “Your prayers are answered” and “Rest in pieces, Automated Counting.” In its statement to Fast Company, Starbucks said that Automated Counting was “designed to simplify a routine task and give partners more time with their customers. When it fell short, we listened to feedback and changed course.”
People involved with NomadGo say the startup has worked with 7 of America’s 20 largest quick-service food and beverage brands, but that Starbucks had been a special challenge. Starbucks has boasted that baristas can craft 170,000 drinks (though one outside estimate multiplied every size, shot, milk, syrup, foam, drizzle, topping, powder, and other inclusions together and reached 383 billion possible permutations for the latte alone). That requires a lot of product SKUs. Some, like those for the 1,500 cup-and-lid combinations stocked nationwide, were excluded from Automated Counting from the outset.
NomadGo employees were struck by the fact that an item as basic as milk varied so much state to state, sometimes even store to store. Two people involved in developing the tool argue that the issue wasn’t faulty computer vision; it was Starbucks’s data stack.
By Starbucks’s own admission, that system is dated. Employees on its technology team said in January that the computer network used for inventory and reorders is “fundamentally” still an IBM AS/400 system purchased in the 1990s. The R&D behind Augmented Counting actually spanned the tenures of Starbucks’s last three CEOs, these people said, which complicated matters because NomadGo’s point people often got reset after each leadership change.
Starbucks declined to respond on the record to questions about its inventory. Greschler said his company couldn’t discuss the partnership, but he did address the underlying technology in an email, explaining that it works best with static inventory. Even subtle changes present a “classic challenge” for computer-vision models, he said. Running a limited-time offer or adding Santa to holiday packaging can require “four to six weeks of work” to ensure the system will recognize the new products. Former NomadGo employees have told me that there were a number of times when they learned about a new Starbucks product once it was already on café storage shelves.
When Starbucks announced its AI mission in a blog post in January, it explained that AI initiatives would be used to help baristas do their jobs better. “If it does that, we scale it. If not, we move on.” In June, soon after Automated Counting was retired, parts of the post were rewritten.
Now the company seems less inclined to jettison an AI tool completely, and it seems focused on how AI could support the entire Starbucks ecosystem, not just employees (“for customers, partners, and the long-term health of our business”). “If it strengthens the experience, we scale it. If not, we iterate on it,” is how it reads now. In other words, AI isn’t going anywhere.
Bloomberg reported in late May that Starbucks has started tying tech worker bonuses to how much AI they’re using, and tracking how many of the company’s Back to Starbucks priority initiatives are powered by AI. In July, it came out that Starbucks is developing in-house tools—with AI’s help—to replace its inventory and supply management software, which currently comes from Microsoft and IBM.
The company’s current AI initiative now includes a ChatGPT order bot to help customers find a drink “based on their mood or vibe of the day.” Employees, for their part, have Green Dot Assist, a voice-activated “real-time companion” that answers work-related questions on the floor. Four workers told me that it, too, is prone to glitches. “Nine times out of ten, when I ask for help, it will say, ‘I’m not sure what you’re asking,’” Gattis tells me. For now, she has gone back to looking up answers herself.
The blog post with Starbucks’s AI manifesto lists scheduling among the tasks that algorithms now assist with. The company says nearly 95% of baristas today get their preferred shifts. Yet glitches remain. Addison says one of his colleagues was recently scheduled to work for eight straight days. Her first day off was Juneteenth, when the shift would have paid overtime.
“No human would schedule another person for eight days in a row and make their day off be the one day where they could get time-and-a-half,” he says. “It sort of feels like managers aren’t allowed to blame the automated scheduling. We keep raising concerns, and the response from management is, ‘Sorry, I don’t know how I did that.’ The answer is: Well, you didn’t do it. Let’s be honest. It was the computer that did it.”