The Substrate Problem
What I see today is enterprise knowledge that has accumulated over years but sits behind the worker, not at the point of use. Tomes of documents span intranets, wikis, shared drives, and Slack or Teams channels — but the more revealing pattern is that most knowledge still resides with its producers, tribally. Participation in the collective repository feels less like knowledge management and more like passing digital paper.
I've watched too many SharePoint rebuilds, Confluence migrations, and team-channel consolidations leave more behind than they carried forward. New taxonomies and ontologies don't coexist with old ones; they demand obliteration, because the old structure's assumptions are baked into the documents themselves. You can't cleanly re-sort content that was written against a different mental model.
And through all of it, knowledge management systems continue to suffer the same fatal flaw: they require a level of digital gardening that produces more friction than flow.
"Participation in the collective repository feels less like knowledge management and more like passing digital paper."
The natural response to this sprawl has been retrieval-augmented generation. Point a language model at the tomes, let it find what's relevant, let it summarize. On the surface, RAG looks like the compression layer the knowledge worker has been waiting for — finally, an interface that can surface the right passage from the right document without requiring a human to know where to look.
In practice, RAG on a legacy knowledge base fails twice.
It fails computationally because the corpus was never shaped for machine retrieval. Unstructured SharePoint sites, nested wiki pages, decade-old shared drives, and Slack threads masquerading as documentation all have to be ingested, chunked, embedded, and indexed — at cost, at scale, and with diminishing returns as the corpus grows. Every re-org, migration, and taxonomy change invalidates prior work. The compute is not the hard part; the compute is the tax you pay for the real problem underneath.
It fails semantically because those artifacts were written for human readers navigating digital paper, not for machines extracting grounded answers. Documents assume context the reader brings. Pages reference "the Smith memo" or "last quarter's decision" without links. Policies contradict themselves across versions nobody reconciled. Tribal knowledge — the kind that lives in the producer, not the document — was never written down in the first place. A retrieval system cannot retrieve what was never captured, and cannot disambiguate what was never written to be disambiguated.
The deeper shift is this: today's knowledge worker is not solving a scarcity problem. They are solving an abundance problem. They have too much information, not too little. What they need is compression — signal pulled from noise, at the moment of decision. The previous era of KM was optimized for the opposite problem: finding the scarce document in the pile. We built systems, habits, and careers around that frame. The frame is now wrong.
AI does not fix this by being added on top. AI inherits whatever substrate it retrieves from. If the substrate is tribal, fragmented, and machine-hostile, a language model will produce fluent, confident answers drawn from a corpus that was never trustworthy to begin with. The output will sound better than the input deserves.
"The output will sound better than the input deserves."
What healthy looks like is easier to name than to build. It's the minimum amount of knowledge needed to make a decision — nothing more.
Claude Shannon asked, decades ago, what the smallest signal was that could faithfully reconstruct a waveform. The question for the modern knowledge worker is structurally similar. What is the smallest payload of context that reconstructs a good decision at the moment of use? Is it 140 characters? The preview on a watch-face notification? A single sentence pulled from a forty-page policy? The answer is situational, but the framing is not. The goal is compression to sufficiency, not accumulation to completeness.
"The goal is compression to sufficiency, not accumulation to completeness."
This is the quiet inversion. The previous era of KM optimized for surfacing the scarce document in the pile. The current era needs the opposite: the smallest piece of trustworthy signal, delivered at the point the worker is already looking.
What this requires is something most organizations have not built and most vendors have not sold: a knowledge substrate co-built by workers and AI, together, in the flow of real work. Not a separate gardening activity. Not a mandatory tagging discipline. Not another migration. Instead — subtle feedback loops inside the work itself, where the act of doing the work scaffolds and fills the corpus as a byproduct. The worker produces the decision; the system captures the shape of how the decision was reached; the next worker arrives with that shape already compressed and ready.
This was not possible five years ago. The tools to do it quietly, in the background of work, without demanding attention from the people doing the work, simply did not exist. They exist now. That changes what we should expect from knowledge management — and what we should stop accepting.
The closest thing to healthy knowledge management in 2026 is not running inside enterprises. It is running on individual workers' laptops — Notion, Obsidian, Roam, Logseq, and the quietly sophisticated personal knowledge systems a generation of thinkers has been building for themselves. These tools already do a compressed version of what the cognitive enterprise needs: capture in the flow of work, light structure, AI assistance without ceremony. But they stop at the edge of the individual. The bridge to team and enterprise still requires integration stacks most organizations cannot sustain, and in practice the pairing collapses into cut-and-paste prompting.
"The substrate has to change before the intelligence layer can be trusted. Everything else is sequencing."
The substrate has to change before the intelligence layer can be trusted. Everything else is sequencing.