Nadella warns of hidden AI cost: proprietary knowledge leaks with every interaction

Microsoft CEO Satya Nadella recently outlined what he calls a "reverse information paradox" in enterprise AI—a hidden cost that goes far beyond the upfront price tag. In a detailed post, Nadella argues that using AI systems to their full potential forces organizations to sacrifice something more valuable than money: proprietary knowledge and institutional expertise.

The dual price of enterprise AI#

When companies deploy AI, they face a two-part expense. The first is straightforward—the subscription or licensing fees. The second, Nadella contends, is far more costly: organizations must feed these systems their most sensitive operational secrets to unlock meaningful performance. "The better you want the model to perform, the more of that knowledge you have to feed it," he writes.

This inverts economist Kenneth Arrow's classic information paradox, which centered on sellers' difficulty in proving information value without revealing it. In enterprise AI, the burden flips. Buyers must expose proprietary processes to get useful results, creating an asymmetric information exchange where the real cost remains largely invisible.

Organizational knowledge as operational exhaust#

Each interaction with an enterprise AI system leaves a trail of data—corrections, refinements, evaluations—that accumulates into a detailed map of how an organization actually works. Nadella describes this as "exhaust": "It's the kind of knowledge a competitor could never buy, and the kind that leaks almost imperceptibly: trace by trace, correction by correction, eval by eval."

Over time, these thousands of interactions create an internal knowledge repository that may exceed the value of the original training data seeded into the system. The more employees rely on AI tools, the more an organization's competitive differentiation becomes embedded within them—accessible to whoever controls that system.

The Microsoft angle and the irony#

The argument gains additional weight given Microsoft's own position. The company sells Copilot, a product that extracts maximum value from enterprise information by accessing Microsoft Graph data—documents, emails, chats—that users can already reach. Research has found that Copilot accessed nearly three million sensitive records per organization in the first half of 2025, while audits revealed roughly 80% of Microsoft 365 tenants had significant data-sharing risks.