When the CEO of Microsoft publishes a 1,500-word essay on competitive strategy and it draws 28 million views in 72 hours, enterprise leaders should pay attention.
Satya Nadella's recent essay, "A Frontier Without an Ecosystem Is Not Stable", reframes the entire enterprise AI strategy conversation — away from "which model do we use?" and toward "what capability are we building?"
The concept at the center of it: token capital.
What Is Token Capital?
Nadella defines token capital as the AI capability a firm builds and owns — accumulated through its own workflows, proprietary data, evaluations, and embedded institutional knowledge.
He pairs it with the more familiar concept of human capital: the judgment, relationships, ingenuity, and pattern recognition of your people.
The real competitive advantage in the AI era is not which foundation model you subscribe to. It is the learning loop you build between these two forms of capital — where human direction shapes AI execution, and AI execution generates signal that makes human judgment sharper.
"The loop between human capital and token capital will be the new IP for firms." — Satya Nadella
"You can offload a task, or even a job, but you can never offload your learning." — Satya Nadella
This is the strategic fault line. Every enterprise will automate workflows. But only some will build the architecture that makes those workflows generate proprietary intelligence — intelligence that compounds with every run.
The Warning: API Consumers vs. Loop Builders
Nadella's most pointed warning is about a scenario many enterprise leaders are quietly sleepwalking into: ceding value to a few frontier models that eat everything they see.
His analogy is globalization. Offshoring created efficiency gains, but value concentrated at the top — and the consequences are still playing out. AI could follow the same pattern. If every enterprise rents intelligence from the same handful of foundation models, using the same prompts and the same generic automations, no one builds a durable edge. The model provider captures the value. The enterprise remains a cost center.
The firms that win are frontier firms: companies that treat AI not as a subscription service but as a core production asset — one that compounds with every workflow it runs.
Model Sovereignty: The Test Nadella Sets
Nadella offers a precise test for whether you have actually built token capital:
"A company should be able to switch out a 'generalist' model without losing the 'company veteran' expertise built into their learning system."
If swapping one foundation model for another means your AI forgets your business context, your approval logic, your exception patterns, and your institutional knowledge — what you have is not token capital. It is a well-configured API client.
Real token capital is portable. The institutional knowledge lives in your architecture — in the semantic layer that understands your ERP fields, the knowledge graph connecting your business relationships, the rules embedded in your execution layer — not locked inside any single vendor's platform.
Nadella's private evals framework reinforces this: your success metrics should measure improvement against outcomes that matter to your business, not external model benchmarks. If a newer model scores higher on public leaderboards but your invoice exception rate goes up, the model is worse for you.
Why Most "AI Strategies" Aren't Building Token Capital
The typical enterprise AI deployment doesn't generate proprietary capability. It generates convenience.
A chatbot on top of SharePoint. Copilot in your inbox. A prompt wrapper around your CRM. These tools reduce friction — but they don't build a learning loop. Every query starts cold. Every workflow runs on generic, shared intelligence. Nothing compounds.
Building actual token capital requires three things most generic AI tools don't provide:
1. Contextualized business data — not just data access
An AI agent that can query your ERP can technically read a customer account number. But reading a field is not the same as understanding it. Token capital starts accumulating when the AI knows that field is a customer account, tied to a credit limit, connected to an open dispute and a 90-day payment history — and can reason across all of it simultaneously.
2. Embedded business rules and audit trails
Generic AI tools apply generic logic. Token capital requires your business rules — approval thresholds, exception policies, compliance requirements — embedded in the agent's execution layer, with every decision logged in a human-readable audit trail. Rules in prompts are brittle. Rules in architecture compound.
3. Workflows that generate proprietary signal
Every time an AI agent resolves an invoice exception, matches a purchase order against an approved price change, or escalates a credit block based on account history — that outcome becomes signal. It reinforces what good execution looks like in your specific business context. That is the loop Nadella is describing.
The Architecture That Makes Token Capital Possible
Building a proprietary learning loop inside enterprise operations is not a prompt engineering problem. It is an architecture problem.
Standard AI tools can access your systems, but they cannot contextualize them. They can read your data, but cannot understand what it means, how it relates, or what action should follow. Rollio's Contextual Data Engine operates across three layers designed to solve exactly this:
- Semantic Index: Every ERP field gets a plain-language business description. Agents understand what data means in your business context — not just what it says.
- Knowledge Graph: An invoice dispute connects to the email thread, the purchase order, the approved price change, and the account manager — not because a developer mapped it, but because the graph infers it.
- Progressive Skills: Business rules, approval logic, and compliance policies are injected precisely when needed — applied with surgical precision, not dumped into context windows.
This architecture satisfies Nadella's sovereignty test. The institutional intelligence lives in the Contextual Data Engine — not in any single model. Swap the underlying LLM and your business context, your rules, and your accumulated judgment remain intact.
What Frontier Firms Are Actually Doing
The enterprises building real token capital are not asking "how do we add AI to our existing processes?" They are asking "what does our process look like if AI executes it autonomously — and where does human judgment add the most value?"
In Order-to-Cash, AI handles invoice matching, exception routing, and payment reconciliation — while human teams focus on relationship decisions and strategic escalations.
In Finance Operations, the close cycle compresses because AI agents resolve exceptions in hours rather than days — generating a feedback loop that continuously improves accuracy.
In IT Service Management, ticket routing, fulfillment, and escalation happen autonomously — while the team's attention shifts to systemic improvements instead of queue management.
In each case the outcome is not just efficiency. It is accumulated capability — an AI that understands your business better after six months than it did at deployment. See also: why predictability is the real differentiator in enterprise AI.
Five Questions to Audit Your Current AI Strategy
- Does our AI get smarter as it runs? Or does every query start fresh with no institutional memory?
- Is our business logic embedded in execution, or just in prompts? Prompt-based rules are brittle. Embedded rules compound.
- Could we switch foundation models without losing our AI's business expertise? If not, you are renting — not building.
- Can we audit what our AI did and why? Token capital requires explainability — not just outputs.
- What is the surface area of our learning loop? The more workflows AI touches with real business context, the faster token capital accumulates.
The Strategic Imperative
Nadella warns of a world where "a small number of AI systems capture all the economic returns, while entire industries find their knowledge commoditized right out from underneath them." That is not just a societal risk — it is a business risk for every enterprise that delays building its own loop.
The companies that will look back on this decade as a turning point are not the ones that moved fastest to adopt a model. They are the ones that moved smartest to build a loop.
Token capital is the new IP. The question is whether your current AI architecture is designed to accumulate it — or just consume it.
Want to understand what a proprietary learning loop looks like in your operations? Book an architecture call with the Rollio team.
