Please don't implement a "company brain" and expect a learning loop

Satya Nadella's viral post linking human capital and token capital through a learning loop has sparked discussion on long-horizon AI strategy. The author warns against implementing a "company brain" and instead advocates for a structured learning loop starting from minimum viable context that compounds through use. The approach emphasizes curated shared knowledge, introspectable multiplayer work, and executable systems to prevent AI decay.

Back to The Wire /blog Please don't implement a "company brain" and expect a learning loop The long-horizon AI strategy is not a company brain. It is a structured learning loop that starts from minimum viable context and compounds through use. Satya's piece https://x.com/satyanadella/status/2066182223213293753?s=20 is rightly going viral. He's connecting two things, human capital and token capital, through a learning loop: how humans steer AI, how AI makes humans smarter, and how the whole thing compounds. We've been researching and building exactly this for nine months. The learning loop is your long-horizon AI strategy. Without it, AI is incredible in the moment but decays silently over the long run. Long-horizon AI work spans far beyond a Codex or Claude session. It's the work where every employee wakes up each morning iterating on the overall goal of their team and the organization. Your long-horizon learning loop has four properties. Minimum Viable Context. Seed the system with the judgment only your company has, not everything it has ever said. Structured, introspectable, multiplayer. Everyone and every agent builds on the same shared foundation, and intermediate decisions can be traced and judged. Executable. Once the work is visible, agents can operate on it directly. Compounding. It measurably gets better the more it's used. The opposite of AI decay. Start from Minimum Viable Context Context is your most precious asset. It's also your most dangerous one. For that reason, do not pull the whole company into a shared knowledge graph. Don't do the full "company brain" thing. You start from a curated set of shared knowledge, and you spend real time curating it down as much as possible. Start with the decisions where your company's expertise actually matters: strategy, product tradeoffs, customer understanding, risk assessment, market interpretation, sales judgment, engineering architecture, operational playbooks. The opportunity is not to automate generic work. It is to find the places where human capital creates differentiated judgment. Humans can filter context against their own understanding of the world, built through meetings, offsites, and deep work. AI can't come to your offsite. You need to build clean context the same way you onboard a new employee. You wouldn't hand a new hire access to everything on day one and tell them to just start working. This is the start of your shared system. Make the work structured and multiplayer Once everyone and every agent works from the same Minimum Viable Context, all future work gets built and evaluated against it. That means turning your Minimum Viable Context into assumptions, source context, intermediate reasoning, examples, constraints, rubrics, decisions, and outputs. Do not bury the company's learning inside prompts, chats, or individual memories. Make the reasoning visible, editable, reusable, and shared across people, agents, and teams. Make the system executable Once the work is visible, agents can operate on it. Structure the artifacts into repeatable circuits with dependencies, inputs, intermediate outputs, and final deliverables. Now AI is not just answering questions. It is participating in a governed system of work that can be replayed, revised, inspected, and improved. So much of organizational work is interdependent: Product, Engineering, Marketing, and Support all lean on each other's output. An important artifact usually gets used downstream across many areas, well beyond the team that produced it. If AI is only used to produce final artifacts, a marketing brief, a piece of code, a support article, then all of the intermediate work behind them gets redone by each department. These intermediate artifacts are also introspectable. Humans and AI can see how a decision or document was actually produced, which is foundational to the learning loop. Conceptually, this is very similar to data engineering, where tools like dbt are used to model organizational business logic. There are a lot of patterns to share here. Everything inherently compounds We call this the stewardship pattern, and it's how the system compounds. Everyone can keep their own harness, Claude, Codex, whatever they like. The contract is that every session leaves the shared structure in better shape than it started, by updating structure, adding it where it's missing, and introspecting the artifacts. That is how personal harnesses compound across the org. Without that contract, they just fragment it. This is a human problem and an AI problem at the same time. When a foundational principle is added or changed, the explicit structure knows exactly what to update and what no longer aligns. That is the opposite of AI decay. Then change what you reward You must reward the people who make this happen. Most organizations have focused on procuring AI tools, which produces a small set of elite early-adopter employees. If every employee has their own chat, their own memories, and their own files, a few of them get sharper in private. The organization gets nothing. You may even get acceleration in conflicting directions. This is "AI hero ball," and nobody wins championships that way. Ignore AI output that doesn't use or improve the organizational system. Reward the people who build the system, and the ones who use it. More from The Wire /blog