Coding is the undeniable PMF use case for AI, but most knowledge work still has a ways to go

Coding works because all the context lives in a git repo that is versioned, structured, in one place, and usually with a test suite that tells you what is correct
Knowledge work on the other hand is based on information that resides in a bunch of different places - Slack, email, various systems, and often in people's heads
So if you want knowledge work to be automated like code, you need a "context repo" or as often called a "company brain"
But a company brain that just takes files, puts them somewhere and slaps RAG on top is not the right answer
What you need is an ontology-based system - setting that up is hard, and maintaining it is even harder
There's a reason why we're seeing a lot of enterprise AI pilots fail or not generate the ROI, it's because the "context repo" is not getting done properly
Here's the kicker though - if done properly, the context layer can go from being a cost companies pay for a new tool, to an ASSET they build that accrues value
It becomes an asset when agentic workflows are built on the context layer and the telemetry data from those agents feeds RL environments that enable the company to post-train models that encode how the company does work, and those models are owned by the company
Owning a model with your data on how the company actually works is the final boss move