Remember when AI was supposed to save companies money? Internal data from Microsoft suggests the opposite is happening at enterprise scale, and it's causing quiet panic in C-suites that went all-in on AI transformation.

According to financial analysis of Microsoft's own AI deployments, the cost per AI-generated output is now exceeding the equivalent cost of human labor for many common enterprise tasks. Not by a little. By enough that some divisions are quietly scaling back AI usage.

Let me break down the math that has finance teams sweating. A typical enterprise AI implementation costs:

• $25-50 per employee per month in licensing (Copilot, ChatGPT Enterprise, etc.) • $0.002-0.10 per API call for custom implementations • Compute costs that scale with usage • Integration and maintenance overhead

For simple tasks like summarizing emails or drafting boilerplate text, the ROI is there. But for complex knowledge work - the stuff companies actually care about - the costs are adding up faster than the value.

One particularly telling data point: Microsoft's own internal teams found that using AI to generate certain types of code was costing more per line than just paying engineers to write it. The AI wrote more code faster, but it required so much review, debugging, and correction that the total cost exceeded traditional development.

The issue is twofold. First, enterprise AI pricing is still based on compute, not value. You pay for tokens processed, not problems solved. That's fine when tokens are cheap, but as models get larger and usage scales, the bills are getting eye-watering.

Second, AI output quality is inconsistent enough that it requires human oversight for anything important. So companies aren't replacing human labor - they're adding AI costs on top of human labor.

Satya Nadella has bet Microsoft's future on AI. The company has poured tens of billions into infrastructure. They've reorganized entire divisions around Copilot. And now their own data suggests the unit economics don't work yet.

This doesn't mean AI is useless. It means the current generation of AI, at current prices, for current use cases, isn't delivering the cost savings executives were promised.