AI had been enthusiastically adopted by all manner of companies around the world, but some are now becoming more circumspect of its cost.
The promise was simple: deploy AI, cut costs, move faster. For a while, the numbers seemed to bear it out. But a string of high-profile reversals — spanning budget blowouts, service failures, and internal pushback — suggests the reckoning is here. A growing list of major companies is pulling back, not because AI doesn’t work, but because the bill has arrived and the ROI hasn’t always followed.
Uber
Uber’s AI experiment became a cautionary tale almost immediately. The company rolled out Anthropic’s Claude Code to roughly 5,000 engineers in December 2025. By April 2026, its entire annual AI tools budget — meant to last twelve months — was gone. Per-engineer monthly API costs ran between $500 and $2,000.
The usage numbers looked impressive: 95% of engineers used AI tools monthly, 70% of code commits were AI-driven, and agentic AI feature usage surged from 32% to 84% between February and March 2026. But Uber’s COO Andrew Macdonald admitted the link between those figures and actual consumer value remains elusive. “It’s very hard to draw a line between one of those stats and ‘Okay, now we’re actually producing 25% more useful consumer features,'” he said.
The core problem Uber exposed is structural: variable token pricing creates budgeting nightmares at scale. The more useful the tools, the higher the bill — and the harder it becomes to justify the spend.
Microsoft
Microsoft — which has staked its identity on AI leadership, being the biggest investor in OpenAI, and built GitHub Copilot — quietly began revoking internal Claude Code licenses. Engineers were redirected to GitHub Copilot CLI instead. The deadline for engineers in its Experiences and Devices division (covering Windows, Microsoft 365, Outlook, Teams, and Surface) is June 30 — the last day of Microsoft’s fiscal year.
The irony is sharp. Claude Code had reportedly become “perhaps a little too popular” among Microsoft’s own engineers, who preferred it over the company’s in-house tools. The tool worked. It just cost too much when used at enterprise scale.
Klarna
Klarna was perhaps the loudest evangelist for AI-driven workforce replacement. Between 2022 and 2024, the Swedish fintech eliminated approximately 700 positions, replaced them with an OpenAI-powered chatbot, and CEO Sebastian Siemiatkowski boasted that AI could already do every human job. At its peak, the AI system managed two-thirds to three-quarters of all customer interactions.
The reversal was equally loud. Customer satisfaction dropped by 22%. Responses were generic and couldn’t handle complex queries. By mid-2025, Klarna began rehiring human agents. Siemiatkowski later admitted: “We focused too much on efficiency and cost. The result was lower quality, and that’s not sustainable.”
Commonwealth Bank
Australia’s largest bank followed a similar trajectory. In July 2025, the Commonwealth Bank replaced 45 call-centre agents with an AI voice bot, claiming it would reduce call volumes. It did the opposite. Calls spiked, queues exploded, and managers were pulled back onto the phones. The Finance Sector Union exposed the gap between the bank’s projected and actual productivity gains.
By August 2025, CBA had reversed the redundancies, offered back pay, and apologized — admitting it “didn’t properly consider that an increase in calls would continue over a number of months.” The bank called the decision an error.
Duolingo
Duolingo’s pullback is subtler but telling. In April 2025, CEO Luis von Ahn declared the company “AI-first” and announced that employees would be assessed on how effectively they used AI tools in their work. A year later, he walked it back.
After employees began questioning whether they were being pushed to “use AI for AI’s sake,” von Ahn removed the metric from performance reviews. “The most important thing in your performance is that you are doing whatever your job is as well as possible,” he said. “A lot of times AI can help you with that. But if it can’t, I’m not going to force you to do that.”
The Bigger Picture
These cases share a pattern. AI tools were deployed fast, sometimes with internal leaderboards and usage mandates designed to drive adoption. Costs scaled faster than value. And the things AI still struggles with — empathy, nuance, complex problem resolution, measurable consumer outcomes — turned out to matter more than the headline metrics suggested.
The shift underway isn’t a rejection of AI. It is a repricing of the hype. The first wave of enterprise AI adoption was driven by enthusiasm. The second wave is being shaped by spreadsheets.
Big Tech is projected to spend $655 billion on AI infrastructure in 2026. The companies retreating now aren’t pulling out of AI entirely — they’re recalibrating, slowing down, and asking the question that should probably have come first: what, exactly, are we getting for this?