Starbucks deployed AI inventory management across 11,000 stores.
Nine months later they told employees to go back to counting by hand.
The system couldn’t reliably count bottles of milk.
The pattern nobody is saying out loud
This would be a funny one-off story if it were just Starbucks.
It’s not just Starbucks.
Right now across the entire tech industry the CEOs who bet everything on AI are quietly reversing course. Cancelling data centers. Rehiring the humans they laid off. Getting sued by their own franchises. Walking back everything they said two years ago.
Quietly. Behind closed doors. Hoping nobody notices the contradiction.
The rehiring problem
February 2024. Klarna CEO announced their AI chatbot was doing the work of 700 full-time customer service agents. Handled 2.3 million conversations in its first month. Company cut its workforce from 7,000 down to 3,000.
Fifteen months later he told Bloomberg the AI-first push had destroyed their customer service quality. The company had been too focused on cutting costs. Investing in human support was actually the way forward.
Klarna launched a rehiring drive.
Duolingo CEO sent an internal memo declaring the company AI-first in April 2025. Phased out contractors. Said he didn’t expect the backlash. By April 2026 he had completely walked back requiring AI usage in performance reviews.
Salesforce CEO reduced customer support from 9,000 people to 5,000 using AI agents. By April 2026 he announced they were hiring a thousand new graduates. Apparently forgetting he had spent two years explaining why those jobs no longer needed to exist.
The money problem
Uber revealed 95% of their engineers use AI tools monthly. 70% of committed code is AI generated.
They burned through their entire 2026 AI coding budget in four months.
Called it a head-exploding moment.
The bigger problem — none of that usage has translated into anything Uber can actually measure. No clear connection between AI adoption stats and shipping more useful features to customers.
The cost keeps climbing. The measurable output doesn’t.
Microsoft cancelled most of its internal Claude Code licenses and forced engineers to switch to GitHub Copilot by June 30th. Not because Claude Code performed badly. Engineers liked it too much and the token pricing was too expensive at enterprise scale.
An unnamed enterprise client ran up a $500 million bill on AI tools in a single month. They gave thousands of employees unlimited access with no spending caps.
What the data actually shows
MIT found that 95% of enterprise AI projects deliver zero measurable return on investment.
A survey of over a thousand senior leaders found 39% of companies made layoffs specifically because of AI. More than half of those now admit they made the wrong decision.
The lawsuits starting to arrive
A Pizza Hut franchisee operating 111 locations filed a $100 million lawsuit against the parent company over a mandated AI dispatch system.
Before AI — over 90% of deliveries arrived within 30 minutes.
After AI — about 50% of deliveries took 45 minutes or longer. Same-store sales in New York City swung from double-digit growth to nearly negative 10% year-over-year.
The lawsuit says the AI system did the exact opposite of improving efficiency. It stripped managers of operational control and introduced algorithmic behaviors that slowed everything down.
Air Canada lost a tribunal case after their chatbot invented a bereavement fare policy that didn’t exist. The airline tried to argue the chatbot was a separate legal entity responsible for its own actions.
The tribunal called that remarkable and ruled against them.
Builder.ai — a startup valued at over a billion dollars and backed by Microsoft — claimed their AI built software automatically. Turned out human engineers were doing most of the work behind the scenes. Then $180 million in accounting fraud surfaced and the company collapsed.
The CEOs who predicted this walking it back
Sam Altman spent two years saying AI would replace nearly half of all work tasks in the near future. At a conference in Sydney he said he was delighted to be wrong. Admitted AI hadn’t displaced nearly as many jobs as he expected and his predictions about economic impact were just off.
Dario Amodei said in 2025 that AI could eliminate half of all entry-level white collar jobs within five years and push unemployment to 10 to 20%. Now he’s reframing AI as something that expands what people do rather than replacing them.
Mark Zuckerberg — who committed tens of billions to AI infrastructure — admits there’s a real possibility this follows the same pattern as past infrastructure bubbles.
Jamie Dimon has warned AI companies are massively overvalued and a crash is a real possibility.
Ray Dalio went further. Called it clearly the early stages of an AI bubble. Compared it to the years right before the dot-com crash.
What this actually means
The internet was real in 2000. Investors still lost trillions betting on companies that never figured out how to make money from it.
Goldman Sachs found the market has already added $19 trillion to AI-exposed company valuations since 2022.
The technology is real. The question — which has never been seriously asked publicly until now — is whether the business models work. Whether the ROI materializes. Whether replacing humans with AI actually saves money or just shifts costs around.
The evidence coming in from Starbucks, Klarna, Duolingo, Salesforce, Uber, Pizza Hut and dozens of others suggests the answer is more complicated than the pitch decks implied.
The pullbacks are happening behind closed doors right now.
That’s not going to last.
