May 1, 2026

The Knowledge Drain

Every year, $31.5 billion in organizational knowledge evaporates. Not through security breaches or system failures. Through normal employee turnover — people leaving for other jobs, retiring, moving on. They take with them an irreplaceable understanding of how the company works: the exceptions to the documented policies, the reasoning behind the decisions that were never fully written up, the context that makes a decade of data meaningful rather than just large.

This number is not the cost of recruiting and training their replacements. It is the cost of the knowledge itself — the repeated mistakes, the recreated context, the tribal knowledge reconstructed from scratch by the people who remain. The knowledge was never stored anywhere a company could keep it.

It still isn't. That is the problem.

The scale of what is not captured

Seventy percent of organizational knowledge is undocumented. It exists only in the experience of the people who hold it. When those people leave — and across a typical mid-sized company, 15 to 20 percent of the workforce turns over each year — that knowledge leaves with them.

What remains is 97 percent unstructured: emails, meeting recordings, chat histories, documents scattered across drives. The knowledge is technically present in the organization, encoded in these artifacts. But it is not accessible in any form that an analyst, a new hire, or an AI agent can use without spending enormous time reconstructing it.

The cost of that reconstruction is 2.5 hours per employee per day, on average, spent searching for information that already exists somewhere in the organization. That is a third of a working day, for every knowledge worker, spent on retrieval — not creation, not decision-making, not the work that actually generates value.

Why AI makes this worse before it makes it better

The arrival of enterprise AI has added a new dimension to the problem. AI systems cannot retain organizational knowledge between sessions. Every conversation starts from baseline. An agent that helped a team resolve a critical incident on Monday has no memory of it on Tuesday. The context has to be re-established, the relevant history re-provided, the institutional knowledge re-injected through prompts and documentation and the patience of whoever is running the session.

Research now quantifies this overhead: employees using AI assistants spend an additional 35 minutes per day re-establishing lost context between sessions. Across a year, that is 146 hours per employee — nearly four full working weeks — spent compensating for the absence of organizational memory in the tools that were supposed to make them more productive.

This is not a model capability problem. It is a memory infrastructure problem. The model is capable. It has no durable access to what the company knows.

The organizations that deployed AI copilots in 2024 and 2025 discovered this quickly. The tool performed well on generic tasks. It failed on company-specific ones. The gap was not the model. The gap was the missing layer between the model and the company.

The structural cause

Knowledge drains for structural reasons, not behavioral ones.

Companies do not fail to document their knowledge because they are careless. They fail to document it because documentation is expensive, slow, and disconnected from the actual work. Writing up the reasoning behind a decision takes time. Updating a policy document every time an exception is approved takes discipline that rarely survives contact with a busy quarter. Knowledge management systems are built for retrieval, not for capture — they require someone to decide to put knowledge in before it can be retrieved.

The result is a system that captures what was intended, not what actually happened. Policies as written, not policies as practiced. Decisions as announced, not decisions as made.

The actual knowledge of a company lives in the interactions that generate the data the company already collects: the support tickets that reveal how edge cases are actually handled, the decision records that show who approved what and why, the workflow logs that map how processes deviate from documentation. The knowledge is implicit in this data. It is not structured. It is not queryable. It does not survive the departure of the people who know how to interpret it.

What infrastructure looks like

The knowledge drain is not solved by better documentation practices. It is solved by a system that captures knowledge as a byproduct of the work — one that structures what the company knows as it operates, rather than requiring a separate effort to record it after the fact.

This is what YC identified in their Summer 2026 Requests for Startups as the "company brain" — a system that pulls knowledge out of fragmented sources, structures it, keeps it current, and makes it queryable. Not a search tool. Not a knowledge base. A structured representation of how the company actually works, derived from the data the company already generates.

The infrastructure requirements follow directly from the problem. Knowledge must persist across personnel changes — it cannot live in the heads of the people who hold it. It must be queryable by agents — it cannot exist only as unstructured text. It must maintain history — knowledge that was true then and is superseded now is still knowledge; it cannot be deleted. And it must be access-controlled — not all of it belongs to everyone.

These are not product decisions. They are requirements imposed by the nature of organizational knowledge. A system that stores documents and calls it a knowledge base has not solved the knowledge drain. It has digitized it.

Dominir is built to satisfy exactly these requirements. Decisions are captured as structured records — not text — with provenance, clearance levels, and a history that is never deleted. The Kernel runs on your hardware. The knowledge does not leave. What accrues in the system is a company memory that outlasts any individual who contributed to it.

The compounding effect

The knowledge drain compounds. Every departure takes not only the knowledge that person holds, but the connections between their knowledge and the knowledge of the people who remain. Reconstructing those connections is harder than reconstructing any individual piece of knowledge.

The companies that establish a structured knowledge layer early accumulate a compounding advantage. Every decision recorded, every exception documented, every workflow captured in structured form becomes part of an asset that survives personnel changes, supports new hires, and grounds AI agents in how the company actually operates. The moat is not the software. The moat is the data that accumulates in it.

McKinsey's analysis of enterprise AI deployments found that nearly 80 percent of companies deploying gen AI report no significant bottom-line impact. The missing enabler, consistently, is productized data — structured, governed, and accessible to the systems that need to operate on it. This is the knowledge drain made visible in quarterly results.

The companies that close this gap first will not just have better AI. They will have a knowledge asset that compounds for as long as they operate. The companies that do not will rebuild from baseline with every cohort of departures.

Dominir captures and structures company knowledge as it is produced — decisions, context, and reasoning stored in a local ontology that persists across personnel changes and grounds agents in how the company actually works. Read the docs or request access.