Enterprise AI was sold as a story of cheap, tireless digital labor replacing expensive humans. In 2026, a growing pile of evidence suggests economics can run the other way, and the most telling example comes from inside Microsoft itself.

In December 2025, Microsoft rolled out Anthropic’s Claude Code to thousands of employees across its Experiences & Devices division, including teams working on Windows, Microsoft 365, Outlook, Teams, and Surface. Less than six months later, it started canceling most of those licenses, redirecting engineers toward its own GitHub Copilot CLI ahead of June 30, 2026 deadline.

And here's the surprising part: The tool wasn’t removed because engineers disliked it. Quite the opposite. It was removed because they used it too much. 

The decision became public just two days before Anthropic announced its IPO plans, one of the most anticipated listings of the year, alongside the SpaceX (SPCX) IPO and OpenAI IPO.

Crucially, Microsoft did not abandon Claude entirely. Anthropic's models remain available through Copilot CLI and Microsoft's broader Foundry partnership, which includes an investment of up to $5 billion in Anthropic. What ended was the expensive product layer sitting on top of the models.

A token-based cost problem

The root cause is pricing. Unlike traditional software licenses, frontier AI tools charge by token usage. Every prompt, code review and debugging request adds to the bill. Per-engineer costs reportedly ranged from $500 to $2,000 per month, while U.S. AI software prices climbed 20–37% over the past year. One AI consultant told Axios that a client spent half a billion dollars in a single month after failing to cap employee Claude usage.

Microsoft isn't alone. Uber's CTO reportedly burned through the company's entire 2026 AI budget in just four months. Even Nvidia — which sells the chips powering much of this ecosystem — acknowledges the issue. VP Bryan Catanzaro told Axios that, for his team, compute costs exceed employee salaries. Duolingo, which loudly embraced an "AI first" strategy by replacing contractors, is now reconsidering the scale of its AI spending.

Gartner has now placed generative AI in its “trough of disillusionment”, and expects 25% of planned 2026 AI spending to slip into 2027, as a sign that the proof-of-concept enthusiasm is colliding with procurement realities.

It also estimates that only 28% of AI infrastructure projects ultimately deliver in line with their business case. Cheaper tokens won't necessarily solve the problem: Agentic models consume far more tokens per task, so usage growth can outpace declining unit costs.

The broader picture is still more sobering. McKinsey estimates that only around 5% of jobs are currently fully automatable, with roughly 60% have at least 30% of their tasks exposed to automation. The story is increasingly about transformation rather than replacement.

The lesson isn’t that enterprise AI fails. It's that enterprise AI requires discipline. 

The winners will treat AI as a practical tool for optimization, not a service to be consumed indiscriminately. That means routing routine work to cheaper models, reserving frontier reasoning for high-value tasks, and measuring ROI at the level of individual workflows.

The hype phase is winding down. The accountability phase is beginning, whether anyone says it out loud or not.

This article was written in cooperation with TradingView