Starbucks told NomadGo on 3 April that it was scrapping the AI inventory tool the Redmond startup had built for it. Within days the 30-person company cut a large part of its workforce, according to GeekWire.
The cuts included the technical team that ran the Starbucks account. Starbucks did not tell its own baristas for another six weeks.
“There’s nothing you can do when leadership and strategy change,” David Greschler, NomadGo’s chief executive, said. He called the decision a complete surprise.
The account comes from a Fast Company investigation published on Monday. It draws on dozens of Starbucks employees and on the startup itself.
We covered the cancellation in May. The vendor’s side of it has been missing until now. Baristas heard on 18 May. The memo told them to pull the QR tracking codes off the backroom shelves and count by hand again.
What actually broke #
The failures were mundane and physical. A shift supervisor near Seattle aimed the iPad at a steel fridge. The camera caught the reflection, and five cartons of oat milk registered as ten.
Elsewhere the app marked milks as the wrong type, swapped syrups, and in one photograph counted a bin as food.
A store manager in Graham, Texas, had the opposite problem. Patchy Wi-Fi wiped her count partway through, and her managers had already ruled hand counts unacceptable. That left the store with no usable number at all.
The model was not the problem #
NomadGo’s computer vision hit 99% accuracy in controlled tests. Its launch announcement promised counts up to eight times faster than manual methods.
Greschler’s explanation is more interesting than a simple accuracy failure. Computer vision struggles when the inventory itself keeps changing, he said.
Seasonal cups and limited-time packaging could need up to six weeks of retraining. His developers sometimes learned about new items only once they reached the shelves.
The other constraint was older still. People who worked on the tool pointed to the Starbucks backend. It runs on a legacy IBM AS/400 system from the 1990s, which makes real-time store data hard to move.
Insiders estimated the programme may have cost north of $10m over several years.
None of this was a pilot #
The Starbucks AI inventory tool, called Automated Counting, reached all 11,300 company-operated cafés in North America by the end of September 2025. It was gone by 18 May.
Enterprise AI failures are usually described as pilots that never scaled. This one scaled first, which is what made being wrong so expensive.
MIT’s NANDA initiative found that 95% of enterprise generative AI pilots delivered no measurable profit impact. Most enterprise AI spending still has not left the lab.
The UK’s Office for National Statistics found the same shape in adoption data, with AI use widening rather than deepening. Starbucks went the other way and went deep everywhere at once.
Starbucks also defended the tool in public shortly before killing it. Reuters reported the miscounts in February, and the company said then that adoption had improved product availability, CNBC reported. Roughly eight weeks later it told NomadGo to stop.
The leadership change Greschler did not name #
Greschler blamed a change of leadership and strategy. He did not say which one. The sequence is on the record.
Deb Hall Lefevre, then Starbucks chief technology officer, praised the deployment in NomadGo’s launch release on 3 September 2025.
She resigned on 29 September, the same month the rollout finished. Starbucks hired a permanent replacement from Amazon in December.
The tool outlived its executive sponsor by about six months.
What Starbucks says now #
“We use technology to support human connection, not to replace it,” a Starbucks spokesperson told GeekWire. The company pointed to $500m spent on putting more staff in its coffeehouses.
“When it fell short, we listened to feedback and changed course.”
That deserves a fair hearing. Killing a deployed tool across 11,300 sites is harder than letting a pilot quietly expire, and plenty of firms keep paying for software their staff have stopped trusting.
Starbucks has kept its other AI work running. That includes Green Dot Assist for baristas and a ChatGPT integration for customers.
The part the statement does not cover is the supplier. A tool that generates work rather than saving it is a familiar failure, and researchers have given it a name.
Less discussed is who absorbs the cost when a large customer changes its mind. The Starbucks AI inventory tool failed at 11,300 sites, and the bill for that landed on a company of 30 people.
Get the TNW newsletter #
Get the most important tech news in your inbox each week.