Business AI was promoted as affordable, nonstop digital assistance displacing high-paid staff. During 2026, accumulating proof indicates costs may shift in the opposite direction, with the clearest case emerging from within Microsoft.

During December 2025, Microsoft deployed Anthropic Claude Code across thousands of workers in its Experiences and Devices group, covering units handling Windows, Microsoft 365, Outlook, Teams, and Surface. Roughly six months afterward, the firm began withdrawing most of those subscriptions and shifting engineers to its own GitHub Copilot CLI ahead of the June 30, 2026 cutoff.

The unexpected element is that the tool was not withdrawn because staff rejected it. The reverse proved true. Usage proved far heavier than anticipated.

The move surfaced just prior to Anthropic revealing IPO intentions, among the most watched offerings of the period, together with the SpaceX IPO and OpenAI IPO.

Microsoft retained access to Claude models through Copilot CLI and its wider Foundry collaboration, which features commitments reaching 5 billion dollars toward Anthropic. The costly overlying service built atop the models was what concluded.

A token-driven expense challenge

The core issue stems from pricing. In contrast to conventional software subscriptions, leading AI platforms bill according to token consumption. Each query, code inspection, or troubleshooting step increases the total. Reported monthly expenses per engineer spanned 500 to 2000 dollars, while domestic AI software rates rose 20-37 percent during the prior year. An industry advisor noted one client incurred half a billion dollars in a single month after omitting limits on Claude activity.

Microsoft is not unique. Uber leadership reportedly exhausted the firm entire 2026 AI allocation within four months. Even Nvidia, supplier of the processors enabling much of the infrastructure, recognizes the pattern. Its VP Bryan Catanzaro indicated that compute expenses for his group surpass personnel compensation. Duolingo, which publicly adopted an AI-first policy by substituting contractors, is now reassessing the extent of its AI outlays.

Gartner has positioned generative AI within its trough of disillusionment and anticipates 25 percent of scheduled 2026 AI expenditures will defer into 2027, reflecting how early enthusiasm meets purchasing constraints.

The firm further projects that merely 28 percent of AI infrastructure initiatives ultimately meet their projected returns. Lower token prices alone may not resolve matters, since agentic systems demand substantially more tokens per operation, allowing volume growth to exceed per-unit savings.

The wider outlook remains measured. McKinsey calculates that only about 5 percent of roles are presently fully automatable, although roughly 60 percent feature at least 30 percent of duties open to automation. The emphasis is moving toward adaptation instead of outright substitution.

The takeaway is not that enterprise AI is ineffective. Rather, enterprise AI demands careful oversight.

Successful organizations will apply AI as a targeted instrument for efficiency, avoiding unrestricted consumption. This involves directing standard tasks toward lower-cost models, preserving advanced capabilities for critical work, and evaluating returns at the workflow level.

The promotional period is concluding. The phase of measured accountability is underway, whether stated openly or not.