<cite index="52-1,52-2">Microsoft CEO Satya Nadella warned that enterprises adopting AI face a hidden cost beyond subscription fees: the proprietary knowledge they are forced to hand over just to make the technology useful. In an X post, he called this the 'Reverse Information Paradox,' arguing firms should protect their 'intelligence exhaust' through better control, capability, choice, cost efficiency and continuous learning.</cite>
<cite index="51-2">To obtain meaningful responses from an AI model, one needs to supply the system with the context of the business, its processes, errors, and corrections. This 'intelligence exhaust'—referring to the byproducts of prompts, corrections, and evaluations that workers make when employing an AI system—is not garbage information but institutional knowledge that cannot easily be purchased from a competitor and is leaking out gradually with each correction in the model.</cite> <cite index="52-5">Nadella highlighted that these corrections represent accumulated prompts employees write, the tools AI agents use, and particularly the corrections made when a model gets something wrong, which are quietly distilled into institutional know-how that leaks out gradually, 'trace by trace, correction by correction, eval by eval.'</cite>
<cite index="51-3">Nadella has come up with a solution he calls the 'five Cs': Control (organizations should have control of their evaluations and institutional knowledge rather than leaving this in the hands of a third-party vendor); Capability (creating private, tenant-aware environments in which models are trained using actual organizational data); Choice (not getting locked down into a particular model or vendor); Cost (making efficient use of different models and workflows); and creating a continuous learning loop that allows AI investments to compound the value of the firm.</cite> <cite index="52-4">Nadella argued that enterprises need a new kind of trust boundary, one that keeps not just data but 'intelligence exhaust' inside the organisation without explicit consent to share it, noting that while it is reasonable for AI companies to have fair use rights to train on public data, it is inconsistent for those same companies to then impose restrictive terms on distillation while reserving the right to learn from customer usage.</cite>