What is the story about?
Artificial intelligence has been widely promoted as a technology that can unlock productivity, automate repetitive work and help organisations make better decisions. But Microsoft CEO Satya Nadella believes the same systems could quietly weaken businesses if companies fail to protect the knowledge they generate while using them.
In a detailed post on X, Nadella introduced what he calls the "Reverse Information Paradox", a concept that reframes one of the biggest economic questions surrounding enterprise AI. His argument is that companies are not simply paying for access to AI models—they may also be handing over the expertise, workflows and decision-making processes that make them unique.
The idea builds on Nobel Prize-winning economist Kenneth Arrow's "Information Paradox", but Nadella argues the relationship between buyers and sellers has fundamentally changed in the age of generative AI.
From buying intelligence to giving it away
Arrow's theory suggested that information is difficult to sell because a buyer can only judge its value after receiving it. Nadella believes AI turns that problem on its head.
"In the AI age, the buyer risks giving away knowledge, just to use what they bought," he wrote.
According to Nadella, organisations effectively pay twice when adopting AI. They first spend money to access advanced models, then provide the context those models need to become useful. That context includes company documents, internal processes, specialised terminology and business expertise.
"The better you want the model to perform, the more of that knowledge you have to feed it," he said.
As businesses increasingly depend on AI assistants and agents, Nadella argues that a growing share of their intellectual capital is embedded within interactions with those systems. Over time, this creates an imbalance where AI providers continue improving their products while enterprises risk losing control over the knowledge that differentiates them from competitors.
Nadella argues the issue extends well beyond conventional concerns about privacy or confidential data. He says AI systems also capture what he describes as "intelligence exhaust"—the valuable by-products created every time employees interact with AI.
This includes prompts, corrections made when models generate inaccurate answers, evaluation methods, workflow patterns, tool selections, feedback loops and decision-making processes.
Rather than viewing these interactions as routine usage data, Nadella says they represent institutional expertise accumulated over years of operating a business.
"In consuming intelligence, you are creating intelligence. And what you create should belong to you," he wrote.
He warned that this knowledge can gradually leave an organisation "trace by trace, correction by correction, eval by eval", even if no confidential document is explicitly shared.
Nadella also questioned current industry practices in which AI providers often limit customers' ability to distil or adapt foundation models while retaining broad rights over how customer interactions improve future AI systems. If learning primarily benefits model developers, he argued, economic value will increasingly concentrate with companies that own AI infrastructure rather than those generating the expertise.
To address what he sees as a structural imbalance, Nadella proposed a stronger "trust boundary" for enterprise AI—one designed to protect not only business data but also organisational learning.
He argued that companies should retain ownership of AI memory, interaction traces, evaluation frameworks, adapted model weights, feedback systems and institutional context. Enterprises should also have the freedom to use AI-generated outputs to improve or train their own models.
Referencing Palantir CEO Alex Karp, Nadella said businesses ultimately want control over "their compute, their models, their data stack, and their alpha", rather than allowing those assets to become dependent on external AI providers.
He outlined five principles that organisations should adopt: maintaining control over proprietary AI knowledge, building secure environments where AI can learn from internal workflows, keeping orchestration layers independent of individual model providers, optimising model selection for cost and performance, and creating continuous learning systems that allow AI investments to compound over time.
Nadella concluded that enterprises must rethink intellectual property for the AI era. During the cloud computing revolution, organisations primarily accumulated data. In the next phase of computing, he argued, the most valuable asset will be the learning generated through AI interactions. Businesses, he said, should be able to harness artificial intelligence without surrendering the expertise that gives them a competitive edge.
In a detailed post on X, Nadella introduced what he calls the "Reverse Information Paradox", a concept that reframes one of the biggest economic questions surrounding enterprise AI. His argument is that companies are not simply paying for access to AI models—they may also be handing over the expertise, workflows and decision-making processes that make them unique.
— Satya Nadella (@satyanadella) July 12, 2026
The idea builds on Nobel Prize-winning economist Kenneth Arrow's "Information Paradox", but Nadella argues the relationship between buyers and sellers has fundamentally changed in the age of generative AI.
From buying intelligence to giving it away
Arrow's theory suggested that information is difficult to sell because a buyer can only judge its value after receiving it. Nadella believes AI turns that problem on its head.
"In the AI age, the buyer risks giving away knowledge, just to use what they bought," he wrote.
According to Nadella, organisations effectively pay twice when adopting AI. They first spend money to access advanced models, then provide the context those models need to become useful. That context includes company documents, internal processes, specialised terminology and business expertise.
"The better you want the model to perform, the more of that knowledge you have to feed it," he said.
As businesses increasingly depend on AI assistants and agents, Nadella argues that a growing share of their intellectual capital is embedded within interactions with those systems. Over time, this creates an imbalance where AI providers continue improving their products while enterprises risk losing control over the knowledge that differentiates them from competitors.
The hidden value in 'intelligence exhaust'
Nadella argues the issue extends well beyond conventional concerns about privacy or confidential data. He says AI systems also capture what he describes as "intelligence exhaust"—the valuable by-products created every time employees interact with AI.
This includes prompts, corrections made when models generate inaccurate answers, evaluation methods, workflow patterns, tool selections, feedback loops and decision-making processes.
Rather than viewing these interactions as routine usage data, Nadella says they represent institutional expertise accumulated over years of operating a business.
"In consuming intelligence, you are creating intelligence. And what you create should belong to you," he wrote.
He warned that this knowledge can gradually leave an organisation "trace by trace, correction by correction, eval by eval", even if no confidential document is explicitly shared.
Nadella also questioned current industry practices in which AI providers often limit customers' ability to distil or adapt foundation models while retaining broad rights over how customer interactions improve future AI systems. If learning primarily benefits model developers, he argued, economic value will increasingly concentrate with companies that own AI infrastructure rather than those generating the expertise.
A new blueprint for enterprise AI
To address what he sees as a structural imbalance, Nadella proposed a stronger "trust boundary" for enterprise AI—one designed to protect not only business data but also organisational learning.
He argued that companies should retain ownership of AI memory, interaction traces, evaluation frameworks, adapted model weights, feedback systems and institutional context. Enterprises should also have the freedom to use AI-generated outputs to improve or train their own models.
Referencing Palantir CEO Alex Karp, Nadella said businesses ultimately want control over "their compute, their models, their data stack, and their alpha", rather than allowing those assets to become dependent on external AI providers.
He outlined five principles that organisations should adopt: maintaining control over proprietary AI knowledge, building secure environments where AI can learn from internal workflows, keeping orchestration layers independent of individual model providers, optimising model selection for cost and performance, and creating continuous learning systems that allow AI investments to compound over time.
Nadella concluded that enterprises must rethink intellectual property for the AI era. During the cloud computing revolution, organisations primarily accumulated data. In the next phase of computing, he argued, the most valuable asset will be the learning generated through AI interactions. Businesses, he said, should be able to harness artificial intelligence without surrendering the expertise that gives them a competitive edge.
