Your Company Doesn’t Need an AI Strategy

What It Covers

Satya Nadella's viral essay argues companies must stop treating AI as a vendor selection problem and instead build compounding "learning loops" that combine human judgment with AI capability — creating proprietary institutional intelligence that persists regardless of which model powers it underneath.

Key Questions Answered

• •Token Capital Framework: Nadella defines two assets every company must now build: human capital (judgment, relationships, pattern recognition) and token capital (owned AI capability). The critical insight is that human capital grows *more* valuable as token capital increases — human direction prevents AI from running in circles without purpose or organizational context.
• •The Zero Multiplier Test: Analyst Mark Egenstad distills Nadella's framework into one formula: token capital equals human capital multiplied by scaffolding multiplied by feedback loops. If any factor is zero, the result is zero regardless of model quality. Most enterprises currently have model access but zero scaffolding and zero feedback loop measurement in place.
• •Model Sovereignty as Strategic Requirement: The Fable Five export control crisis exposed a critical vulnerability — companies over-reliant on one or two frontier models. Nadella's test for AI sovereignty: a company should be able to swap out its generalist model entirely without losing the institutional expertise encoded in its surrounding learning system and private evaluations.
• •Private Reinforcement Learning as New IP: Microsoft's Frontier Tuning product lets enterprises run reinforcement learning environments trained on their own internal workflow traces, corrections, and accepted outputs. This converts tacit organizational knowledge into machine-operable, model-portable intelligence — shifting AI from rented capability to owned, compounding competitive advantage that rivals cannot replicate.
• •Applied AI Layer Complexity: Box CEO Aaron Levie identifies that enterprise agentic workflows require bespoke context-capture interfaces, model routing between frontier and cheaper models, and dedicated change management — far beyond raw prompting. This complexity creates durable competitive moats for platforms that solve it, contradicting early predictions that the application layer would remain thin.