The Use of Knowledge in the Company

Every company now rents the same models. When competitors and customers get the same APIs, yesterday's unusually polished proposal becomes the norm. What's left is know-how: how you price a job, why you passed on a deal, what a customer was promised, what your best people know and never wrote down.

Friedrich Hayek described this problem across a whole economy in The Use of Knowledge in Society. The knowledge an economy runs on is "the knowledge of the particular circumstances of time and place." Which machine is sitting idle, which ship has room on this week's run. It is spread across millions of people, and no single authority can hold it. His answer was to leave decisions with the people on the spot.

A company is a small economy with the same problem. A smarter AI model has the general knowledge in abundance and none of the particulars. Those sit with whoever was there. That is what the labs are actively buying from data vendors, and it's the one part of the stack your competitors can't rent.

Know-how sits in layers, and each one depends on the ones below it:

1. What people say. Slack, Teams, email and calls. This is where work gets coordinated, and where most of what a company knows first appears.
2. What gets written down. Docs, decks and procedures. They're cleaner than conversation, and usually out of date.
3. What actually happened. The money and the records, in Stripe, Ramp, the books and the CRM. This is the ground truth the other two describe.
4. Why. The reasons behind decisions, which only exist once the first three are connected.
5. A company that learns from its own results. At the top, the company gets better at its work without anyone having to teach it.

The fourth level is where the value is, and the hardest to reach: the reasons behind a decision are never written down in one place.

Picture a real estate firm whose partners have asked an agent to watch the deal pipeline. A broker brings back a property the firm passed on last year, at a lower price. The old model is in Excel, the reason they passed is in a call recording, and the final costs of a similar refurbishment are in Procore. Nobody has put them together.

To answer the question that matters, whether the reason they passed still applies, the agent has to connect three layers and ask the people who know. Maya, Jon and Leah each changed the same work directly, so whatever comes after the meeting starts from what they've already decided. That reason, the why, is the most valuable thing the firm knows.

AI can climb these layers for you. But most of the agents that could do it are built around one person, or keep what they learn where you can't see it. When that person leaves the company, the agent and everything it learned leave too.

The alternative is one agent the whole team shares, living where the team already works, with a memory the company owns.

Ezra lives in your Slack. Anyone on the team can hand it work, correct it, or approve what it sends, in the same thread. It builds the company's memory out of that work, connecting what people said, what they wrote and what happened. The more it works, the better it gets. The memory stays in your company's own deployment, and you can inspect it, correct it and export it whenever you want.

Coding agents showed programmers what an agent can do with their own work, one terminal at a time. An agent in the team's channel can do the same for everyone else. Most teams already have a few people doing serious work with agents, in private. In a shared channel, the rest of the team can watch how they use it and copy what works.

One person can already keep a linked vault of everything they know and point any AI at it. A company should have the same, built from the work as it happens.