The Argument · 01
Every enterprise has three main assets. Digitization focused only on two.
Knowledge management has spent thirty years moving the enterprise from paper to systems of record. Each wave captured a layer of the business and made it queryable. One layer was always left behind, the one everyone actually depends on.
Knowledge is what the organization knows.
Facts, documents, policies and precedent. Explicit and storable, the layer wikis, search and RAG already address, however imperfectly. Wikis describe what; they rarely describe how, and almost never why.
Expertise is how the organization thinks.
Judgment, heuristics and the criteria that separate a yes from a no. It is tacit, fragmented and locked in a handful of senior people. An estimated 42% of an employee’s job-critical skill is tacit, not documentable by a colleague.
Knowledge management digitized the filing cabinet. It never digitized the expert. Expertise is the last undigitized enterprise asset, and the one that drives the outcomes.
lost per knowledge worker per week recreating undocumented information.
Panopto Workplace Knowledge Report
of job-critical skills are tacit, invisible to any document.
Panopto Workplace Knowledge Report
success rate for internal AI builds; vendor-built reaches ~67%.
MIT NANDA, State of AI in Business 2025
The arc of the enterprise system of record
Each decade promoted a new layer of the business into a system of record. Expertise is the next, and last.
The Argument · 02
AI made it even worse, because your agent “forgets every Monday”.
The first instinct was to point a large language model at the problem. But the dominant ways of building agents all share one defect: knowledge dissolves at the boundary of the session. Nothing the agent learns on Friday survives to Monday. The cost is not the model; it is the senior expert time you spend twice.
Generic copilots
ChatGPT · Copilot · Gemini
What survives a session
Nothing. Knowledge lives in the prompt and chat history. Each session restarts cold.
RAG assistants
Glean · Writer · Sana · Dust
What survives a session
Chunks, not structure. Vector fragments of documents. The reasoning and lineage are lost.
Orchestration frameworks
LangGraph · Semantic Kernel
What survives a session
Code. Knowledge is hard-coded into tool graphs. Fixes mean engineering tickets.
Retrieval-augmented generation looks like knowledge management, but a chunk is not a unit of knowledge, practically an unrelated needle in a haystack. Retrieve enough chunks and you can answer a question; you cannot encode how an expert weighs an ambiguous case. That reasoning was never written down, so there is nothing to retrieve. The result is visible in the numbers.
of GenAI pilots deliver no measurable business return.
MIT NANDA, 2025
of agentic AI projects will be cancelled by end of 2027.
Gartner, June 2025
leaders expect agents in their AI strategy within 12–18 months.
Microsoft Work Trend Index 2025
The bottleneck is not models, infrastructure or talent. It is that these systems do not retain feedback or improve over time. Knowledge management for the agent era has to make knowledge, not the agent; the thing that lasts.
Why now: three technology curves have converged
None of this was buildable five years ago. For the first time, expertise can be modeled and continuously evolved.