Microsoft CEO Satya Nadella says companies must own the AI learning loops shaping their future

Microsoft’s CEO argues that enterprise advantage will depend on combining employee expertise with company-controlled AI systems, rather than relying on model capability alone.

Microsoft CEO Satya Nadella says companies must build AI learning loops that combine employee expertise with systems they control

Microsoft Chairman and CEO Satya Nadella has called on companies to build AI systems that preserve and compound their own institutional knowledge, arguing that competitive advantage in an AI-driven economy will depend less on selecting a single model and more on controlling the learning loop between employees and technology.

In a blog shared through LinkedIn, Nadella introduced what he described as "human capital" and "token capital", setting out a model for how businesses could organize work, intellectual property and employee learning as AI agents take on more complex tasks.

Under Nadella’s framing, human capital includes employees’ knowledge, judgment, relationships, ingenuity and ability to recognize important patterns. Token capital refers to the AI capability a business builds and owns using its workflows, data, evaluations and accumulated expertise.

Nadella argued that businesses should develop agentic systems that improve through repeated use while retaining control of the knowledge created inside the organization. He said companies should be able to replace an underlying general-purpose model without losing the specialist expertise incorporated into their wider AI systems.

The Microsoft CEO set out a strategic case for private model evaluations, reinforcement learning environments and queryable internal knowledge bases that allow companies to measure and improve AI against their own business outcomes.

Human expertise becomes part of the system

Nadella rejected the idea that expanding AI capability automatically reduces the value of employees. He argued that human agency would remain responsible for identifying objectives, connecting information across disciplines, forming relationships and determining which outputs matter.

"Without human direction, you have compute running in circles," he wrote.

The proposed model is based on a recurring exchange between employees and AI systems. Staff provide judgment, corrections, context and examples from real work, while the AI system uses those signals to improve future workflows.

Nadella said companies should not treat the transfer of work to AI as equivalent to transferring organizational learning: "You can offload a task, or even a job, but you can never offload your learning. The future of the firm is the ability to compound that learning across people and AI."

That distinction places workforce learning and knowledge management at the center of enterprise AI deployment. An organization may use an external model to power a task, but Nadella argued that the experience generated through that task should continue strengthening a system controlled by the business.

The approach would require companies to capture the knowledge currently spread across employees, documents, processes and decisions. Rather than using AI only to retrieve existing information, businesses would use real work to create new training signals and improve recurring processes.

Companies urged to measure AI against internal outcomes

Nadella said private evaluations should assess whether an AI model is improving against the results that matter to a particular organization, rather than relying only on public benchmarks created for broader model comparisons.

He also proposed private reinforcement learning environments that allow AI systems to develop using traces from real organizational workflows. Internal knowledge bases would make institutional memory easier to search while helping models use tokens more efficiently.

These components would form what Nadella described as a company-owned learning loop: "This loop becomes the new IP of the firm. I think of it as a hill climbing machine. And unlike most assets, it compounds."

Each improvement to a workflow could generate additional information for training, evaluation and refinement. Nadella argued that this would allow companies to accumulate tacit knowledge that competitors could not easily reproduce by purchasing access to the same underlying model.

The ability to change models without losing that accumulated expertise would serve as a test of enterprise control and technological sovereignty, according to the blog. A business might replace the "generalist" model supporting its system, while retaining the "company veteran" knowledge developed through its own operations.

That model also places new demands on workforce development. Employees would need the skills to evaluate outputs, identify patterns, define acceptable results and turn successful AI use into processes that other teams can repeat.

Nadella warns against AI value concentrating in a few models

Nadella also used the blog to challenge an AI economy in which a small group of model providers captures the value created from knowledge generated across other companies and industries.

"The last thing any of us want is a world where every company across every sector is ceding value to a few models that eat everything they see," he wrote.

He compared that risk with the impact of outsourcing during an earlier phase of globalization, arguing that economic indicators did not fully capture the industrial displacement experienced by affected workers and communities.

Nadella said a similar pattern could emerge if companies allow their expertise to be absorbed and commoditized without retaining the systems, learning or financial value created from it.

In his view, frontier AI models need to operate within a broader ecosystem that allows businesses, sectors and countries to develop AI assets of their own. Employees’ knowledge would be amplified through those systems, while the resulting economic gains would remain distributed across the organizations and communities contributing the expertise.

The argument also aligns enterprise AI adoption with ongoing learning rather than one-time technology deployment. Under Nadella’s model, organizations would need to continually update evaluations, convert employee feedback into system improvements and decide which institutional knowledge should remain under their control.

Nadella did not specify how Microsoft would measure token capital or how companies should account for it financially. The immediate proposal is architectural: businesses should build private evaluations, reinforcement learning environments and knowledge systems around their own workflows before more of their expertise becomes dependent on external AI models.