---
title: "How to make Satya Nadella's vision a reality"
url: "https://atlan.com/context-and-chaos/issue/how-to-make-satya-nadellas-vision-a-reality/"
description: "The vision of AI as a compounding learning machine is right - but the context layer it needs has three specific architectural components most enterprises are missing."
keywords: "Context Engineering, Enterprise AI, AI Agents, Context Layer"
---

> Atlan is hosting Context Conference, bringing together the leaders and builders at the frontier of giving AI the context it needs to understand their business. It runs online on October 28, 2026, from 11:00 AM to 2:00 PM ET. Atlan co-founder Prukalpa Sankar opens and closes the day. Leaders from AstraZeneca, BNY and Verizon share why they invest in context and what they get from it. Registrants get early access to The AI Context Gap, a new study from MIT Technology Review Insights. Register: https://atlan.com/context-conference/

A Context and Chaos PK Editorial by Prukalpa Sankar, Co-founder of Atlan, published June 23, 2026 (9 min read). It responds to [Satya Nadella's article](https://x.com/satyanadella/article/2066182223213293753) on X about a continuous learning loop where "human capital" and "token capital" reinforce each other, a compounding "hill-climbing machine" that separates organizations that thrive from those that commoditize. The vision is right, but in most organizations the machine is stalled: the loop is hard to build, human capital is hard to capture, and the system leaks. This issue covers the architecture of context engineering needed to make it real.

## Key takeaways

- "Human capital" is not a documentation problem. It is three distinct layers: an AI-ready knowledge graph of business entities and data relationships, semantics and ontology that give canonical answers (for example "which definition of ARR?"), and skills that encode how work gets done. Most enterprise AI collapses because it treats all three as the same problem.
- The hardest part of building a context layer is capturing institutional knowledge that was never written down. Approaches: skill-building during agent deployment, mining systems like Slack and query logs, failure-driven capture, desktop shadowing, and structured AI interviews with domain experts.
- The real test of enterprise AI is not swapping a frontier model. It is whether agents can agree on business definitions, trace where they disagreed, get smarter from the correction, and whether that institutional intelligence stays portable when infrastructure changes.

## What "human capital" means: the three elements of context

Examples of human capital: private evals that know what "correct" means for your business, reinforcement from real organizational traces, a knowledge base that makes institutional expertise queryable. Most initiatives treat context (human capital) as documentation; it is made of three things:

- **Knowledge**, the map of the business: semantic entities, definitions, metrics and relationships. For example what "ARR" means in the finance motion versus the sales motion, or what a qualified lead looks like in your ICP.
- **Expertise**, how work gets done: procedures and playbooks encoding tacit knowledge and institutional intuition that never reach core systems.
- **Norms**, the programmatic rules of acceptable action: which data an agent can access, who approves a workflow, which customer can get which discount, what data can leave a jurisdiction.

## The context layer: a unified, actionable, governed place for human capital

Context cannot live in a prompt or a database; it needs a place where agents and humans can build on it, modify it and govern it over time. Prukalpa began calling this the [context layer](https://atlan.com/know/enterprise-context-layer/) a year earlier. It has three layers:

- **AI-ready data and knowledge graph**: makes structured data and unstructured knowledge machine-readable. Tables enriched with descriptions, join paths, SQL patterns and usage history; documents identified, classified and indexed into a governed knowledge graph. Without it, agents cannot tell trusted sources from stale ones.
- **Semantics and ontology**: shared definitions, metric logic and entity relationships. Sales and Finance often define ARR differently; today an agent uses whichever definition it retrieves first. The semantic layer gives one canonical answer and reconciles conflicting versions.
- **Skills**: the new primitive that does for procedural knowledge what code did for logic. Each skill is a reusable, versionable, testable unit of how-to, managed like code (version control, testing, lifecycle management, clear ownership). "The tenth agent inherits what the first nine learned."

The human-capital-to-token-capital linkage needs all three. Data without semantics is uninterpretable. Semantics without skills can describe the business but not operate it. Either one without governance becomes contradictory instructions no enterprise can trust.

## What it takes to build human capital

Documentation (write down meanings, index documents, connect the knowledge base) is neither durable nor scalable. Human capital is dynamic; the most valuable knowledge lives in query logs, Slack threads where decisions were made, and veterans' judgment calls. Ways to populate the context layer:

- **Skill-building during agent-building**: do the work manually with the agent once, then have AI examine the session, extract the repeatable pattern and write the skill with triggers and edge cases. This bootstraps an agent at 50% autonomy and improves it toward 90%.
- **Context mining from systems**: read query history, detect where Sales and Finance use two definitions of the same metric, and surface the conflict for a human. Slack, email and event logs already hold most of what is needed.
- **Failure-driven capture**: run the agent, watch where it fails, and capture the missing context at the point of failure, so the library grows where agents actually work.
- **Desktop shadowing**: capture what top operators do (clicks, keystrokes, app switches) and reverse-engineer workflows that systems mining misses.
- **Structured AI interviews**: AI-run conversations that surface exceptions, workarounds and decision rules, such as "Walk me through the approval process. How do you actually decide?"

None produces finished context alone. They produce candidates that a context development lifecycle (testing, review, approval, deployment) turns into governed, durable artifacts.

## Learning leaks: the failure point in the hill-climbing machine

An agent system has four types of memory. Working memory and episodic memory sit close to the agent (immediate execution state). Semantic memory and procedural memory sit inside the context layer: the knowledge the organization wants to preserve, govern and make portable across tools.

Compounding happens when episodic experience is promoted into shared enterprise knowledge: a repeated exception becomes explicit policy, a resolution path that keeps working becomes a reusable skill. The system should detect the pattern, propose the update, and ask the right owner to certify it.

The key gap is [context drift](https://atlan.com/know/context-drift-ai-agents/). When a CMO updates positioning or a finance lead changes a metric definition, the change has a blast radius; if it does not propagate, the social media skill, the SDR pitch skill and the analyst call skill run on different versions of reality. The context layer must understand the blast radius of every change and route updates to the right owners, which requires versioning, change tracking and context propagation. Without them, the learning loop becomes a learning leak.

## The ultimate test: ownership and portability

Satya's test: can a company swap its underlying generalist model without losing the veteran expertise it has built? Prukalpa adds two:

- Can your AI agents agree on what "ARR" means, trace why they gave different answers last week, and get smarter from the correction, without anyone touching a model?
- Can you move all of your human capital and context from one agent platform or data ecosystem to another without losing a day of institutional learning?

Building expertise into proprietary software, a custom GPT stack or a walled-garden orchestration layer trades model lock-in for platform lock-in, trapping context when the next breakthrough happens elsewhere. Satya's warning about a frontier without an ecosystem is right, but the answer is absolute portability, not local applications on top of big platforms. Context is core IP, the digital distillation of human capital; organizations that treat it as a durable asset rather than a consumed resource or documentation project will capture the compounding loop.

## Related reads

- [OpenAI's Frontier Proves Context Matters. But It Won't Solve It.](https://atlan.com/context-and-chaos/issue/openais-frontier-proves-context-matters-but-it-wont-solve-it/)
- [Your AI Context Layer Is Being Built on Stale Metadata](https://atlan.com/context-and-chaos/issue/your-ai-context-layer-is-being-built-on-stale-metadata/)
- [BI-Ready Is Not AI-Ready](https://atlan.com/context-and-chaos/issue/bi-ready-is-not-ai-ready/)
- [Data Governance vs AI Governance: Why It's the Wrong Battle](https://atlan.com/context-and-chaos/issue/data-governance-vs-ai-governance-why-its-the-wrong-battle/)

Context & Chaos is a community newsletter on context engineering, governance, architecture and discovery. All issues: https://atlan.com/context-and-chaos/ / Contribute: https://atlan.com/context-and-chaos/contribute/