Want to Build a Knowledge Factory? First Solve These Four Problems
Harrison Chase proposes the knowledge factory, and the direction is agreed with — but if a company hasn't even got its data and processes in order, building the factory only automates the chaos.
Harrison Chase recently wrote a long piece on the "knowledge factory." The core argument: to get AI doing work inside an enterprise, what's missing isn't the model or the context — it's the knowledge of how the enterprise operates. That knowledge is scattered across documents, systems, and the heads of veteran employees. Companies need a system to collect, organize, and maintain this knowledge, then hand it to agents to use.
LangChain is already doing this internally: the GTM agent, the content agent, and the LangSmith Engine all run similar learning loops.
Maksymilian agrees with the direction but thinks it's not that simple. He says: walk into a company and build a knowledge factory straight from this playbook, and you'll most likely fail. Not because the concept is wrong, but because almost no company is ready. AI transformation has only just begun, and reality is far messier than the blueprint.
He lists the four most common problems:
1. Data is scattered
A typical company has ERP, TMS, WMS, plus email, Excel, and Slack. All of these tools hold data, but the same piece of data looks different in different systems. The agent has to learn on its own to find things across systems — just like a human, this grunt work can't be avoided.
2. Exceptions are everywhere
No matter how complete the SOP is, there are always cases it doesn't cover. An invoice missing a PO, a customer suddenly raising a non-standard request. In these moments, domain experts rely on the experience in their heads, not on any document. That experience never gets written down, yet it's exactly what keeps the process running.
3. People become the glue
"Copying data from one Excel to another, or entering it into ERP from email" — at some companies, this might be a role's entire job. Accumulated, these trivial actions eat up huge amounts of time. People who should be handling complex problems get used as data transmission pipes.
4. The processes themselves are terrible
Take the finance department's supplier settlement process, for example: it might contain dozens of steps, going back and forth between departments and waiting for replies. If the process itself is garbage, automation just lets the garbage flow faster.
Maksymilian's advice: don't rush to build a knowledge factory. First talk to the people who actually do the work, dig into the usage traces of their systems, and map out the existing processes completely. Then redesign the processes based on the new capabilities that AI and automation bring. Only then should you consider what systems to build.
A commenter asked: instead of building a complex knowledge factory, why not just continuously maintain a company brief document as the single source of truth? Maksymilian replied in the comments: in an enterprise, a "single source of truth" is very hard to establish. Processes are vertical: a complete workflow involves finance, sales, operations, and HR, each with its own tools, and the same information gets written different ways in different departments. As for continuously updating the brief — who maintains it? How big is it? Which information deserves to go in? When does it go stale? Constant appending also creates duplication and conflict, and when AI sees contradictory content, it only performs worse. As for how to store and update memory, the industry has no standard yet.
Harrison's article includes an example: while running a revenue report, an agent found that a column name in BigQuery had changed. It corrected it on the fly and flagged the change; that night, a curator agent reviewed the flags and interaction logs and updated the corresponding skill.
The idea is good: the work process naturally exposes knowledge gaps, which then get filled automatically. But the premise is that the enterprise must have a decent process for the agent to run. Otherwise, once the knowledge factory is built, all you feed it is experience from broken processes.
发布时间: 2026-11-23 00:56