Your Company Doesn't Need an AI Strategy
— It Needs an AI Learning System (Rapid Research Brief)

Executive Summary

Key Takeaways

• AI strategy is a dead end. The Fable 5 export-control ban showed that vendor-anchored strategies collapse when a single policy move disables the model.
• Token capital compounds. Human capital becomes more valuable as token capital grows — human agency directs the learning loop.
• Build private evals, not generic training. Evals measured against your quality metrics, care standards, and compliance requirements.
• Microsoft is betting on the execution layer. Frontier Tuning + Copilot Cowork = own the layer where AI meets real work.
• We are at a waypoint, not the destination. Mollick: build training for continuous adaptation, not today's tools.

Key Findings

1 The Fable 5 Disruption: Why Vendor Strategy Is Brittle

The White House banned Anthropic's Fable 5 model via export controls, exposing that organizations treating AI as a vendor relationship are one policy decision away from losing their AI capability. UC Berkeley professor Andrew Reddie: "If creating models that are impossible to jailbreak becomes the de facto standard for the United States, then it will have no AI models."

Health system implication: If your AI strategy is "we use Epic's AI features" or "we've deployed Copilot," you're exposed to the same brittleness. Build model-portable IP — prompt libraries, eval suites, agent workflows — that survive any vendor change.

2 Token Capital: Satya Nadella's Framework

Nadella distinguishes between human capital (knowledge, judgment, relationships) and token capital (AI capability a firm builds and owns). Formula: Token Capital = Human Capital × Scaffolding × Feedback Loops. Companies that treat AI as a vendor relationship are outsourcing not just tasks but their learning.

Health system implication: Clinical judgment is the highest-value human capital. An AI learning system amplifies it by encoding clinical expertise into evals. Every time a clinician corrects an AI output, that correction should feed back into the system.

3 Frontier Tuning: Training AI on Your Own Workflows

Microsoft's Frontier Tuning lets organizations train models directly on their own workflows via RL environments. Mustafa Suleiman: "a training gym for AI" — addressing both AI sovereignty and AI budget.

Health system implication: Train models on your Epic workflows, care protocols, compliance rules. The model that knows your processes is exponentially more valuable than a generic model.

4 Copilot Cowork: The Execution Layer Goes GA

Copilot Cowork went GA with multi-model support. Seat license + usage pricing signals Microsoft's intent to own the execution layer as the part of the AI stack with pricing power.

Health system implication: Copilot is the execution layer, not the strategy. The durable skill is prompt engineering, agent orchestration, and eval design — portable across any execution layer.

5 The Applied AI Layer: Accenture's Warning

Accenture's stock dropped 18% to a near-decade low. Critics: real AI implementation requires deep domain expertise generalist consultancies lack. Aaron Levie: the applied AI layer is substantially more complex than early critics assumed.

Health system implication: Don't outsource your AI learning to consultants who don't understand clinical workflows. The applied AI layer for healthcare is built by clinicians who understand both the medicine and the AI — that's the upskilling imperative.

6 You Can Never Offload Your Learning

The episode's most powerful line: "You can offload a task or even a job, but you can never offload your learning." If your upskilling program is structured as "AI will handle X so you don't need to learn it," you're building fragility. The right framing: "AI accelerates X so you can focus your learning on higher-order judgment."

Health system implication: Every workflow redesign should ask: what human judgment does this preserve and amplify? Build a 4-week sprint from casual AI use to running a personal AI system with custom agents, private evals, and encoded clinical protocols.

7 Ethan Mollick's Caution: A Waypoint, Not a Stable Phase

We genuinely don't know the best approaches to rebuilding organizations around AI agents yet. The temptation will be to impose strict spend limits and bias toward known ROI — exactly the opposite of what the ecosystem approach requires. "A comfortable waypoint that feels like a stable phase but almost certainly isn't."

Health system implication: Don't build training optimized for today's tools. Build infrastructure for continuous adaptation. The durable asset is institutional learning capacity, not any particular tool or workflow.

Risks, Gaps & Uncertainty

• Vendor lock-in risk. Microsoft's ecosystem is a walled garden. Health systems training exclusively on Microsoft's stack face the same lock-in as Epic today.
• Compliance gap. HIPAA, FDA, CMS constraints aren't addressed in the Frontier Tuning framework. Data governance questions remain unresolved.
• Eval design is the hard part. Private evals require clinical expertise, data science capacity, and ongoing maintenance most health systems lack.
• Speed of change. Healthcare adapts slower than the AI landscape due to retraining and compliance validation requirements.
• Equity and access. Token capital compounding could widen gaps between well-resourced and under-resourced health systems.

Recommended Next Actions

Annotated References