Think about what happens when you hire a talented new employee. They arrive smart, capable, and well trained, but they don’t know your firm yet. They don’t know which system has the real numbers and which one is out of date, or which client needs a light touch. They don’t know why an old project from 2022 still shapes decisions today. Over time, they learn. They pick up the unwritten rules, build a mental map of who owns what, and eventually work without much hand-holding.
AI works the same way, except the “new hire” shows up again every time you upgrade a model, switch vendors, or start a new session. Claude, Codex, and other AI tools arrive with huge general skill, but they know nothing about your organization. Ask a capable AI for help and it will try its best. Still, it won’t know which internal system is the real source of truth, what your compliance team requires, or which shortcuts are actually landmines. The result is generic output that needs heavy correction — or worse, output that sounds confident but is wrong.
Firms that get real value from AI have learned something important: the quality of AI output depends on the quality of the context you give it. A well-briefed AI, pointed at accurate and curated knowledge, produces far better results than the same model working from a generic prompt. And this isn’t a one-time task. Models change every month, vendors change too, and your organization changes constantly. Without a real way to capture, curate, and refresh what your AI knows about your business, you keep re-onboarding that new hire, over and over, forever.
This is why Knowledge Management deserves a seat at the table in every AI strategy discussion. It shouldn’t be an afterthought or a one-time documentation project. It should be a living capability that your organization actively maintains.
Your Organization Needs a Knowledge Tree
Knowledge Management might sound like a return to the corporate wiki graveyards of the past. But something is different this time: AI doesn’t just search knowledge, it reasons over it, and it rewards structure. A useful mental model here is a tree.
- The Roots are your organization’s core standards: shared definitions, naming conventions, and a small set of rules. They keep every piece of knowledge consistent and easy for a machine to read. This is where Data Governance plays a key role. It doesn’t dictate every detail. Instead, it sets light conventions that let separately maintained knowledge still work together as one whole.
- The Trunk is the central index. It tells an AI agent, or a curious employee, what branches exist and how to find them.
- The Branches are where the real knowledge lives. Each branch covers one domain, such as Portfolio Management, Investment Operations, Distribution, or IT. Each branch is self-contained and owned by the team closest to that subject.
Be careful about what actually belongs on a branch, because the urge to dump everything into the Knowledge Tree is exactly how it turns into another swamp. A database is not knowledge. A CRM export is not knowledge. Raw transaction records are not knowledge. What belongs on a branch is information about those things. What is the database for? Which fields actually matter, and who owns it? What does a “good” client relationship look like for this account? Why does a policy exist, not just what it says? Generally, if the answer already lives in a governed data platform, the branch should point to it or describe it — it shouldn’t copy it.
This structure matters because it solves the problem that sinks most knowledge efforts. A single, giant knowledge base becomes unmanageable once more than a handful of people touch it. And nobody wants to wait on a central team to update a client note or fix an old policy reference. A Knowledge Tree lets ownership stay local while it still adds up to something enterprise-wide.
Distribute the Work, Not the Standards
This is where Data Governance translates directly into AI enablement. Most firms already understand the value of a Data Owner, who is accountable for a data domain, and Data Stewards, who are responsible for its day-to-day quality. The same model applies cleanly to knowledge.
Every Knowledge Branch needs an Owner — usually a subject matter expert or department leader — who is accountable for that domain’s accuracy and relevance. Supporting them are Stewards, the people who actually update content as things change: a policy revision, a new client detail, a lesson learned from a recent project.
This federated approach is deliberate. A command-and-control model, where one central team tries to curate everything, cannot keep pace with a real organization. Give ownership to the people who already have the expertise, and let governance define just enough shared structure to keep everything working together. That’s what allows Knowledge Management to scale.
Beyond the SharePoint Swamp
Every firm already has a version of this problem. You’ve heard it described the same way: “it’s all in Confluence” or “it’s on the shared drive somewhere.” Those tools aren’t wrong. They’re just built for a different reader.
Why “It’s Already in Confluence” Doesn’t Solve This
Confluence, SharePoint, and similar platforms are built for humans who browse and read prose. An AI agent can query them, but it’s working against the grain. Content isn’t broken into one clear concept per page. Relationships between related pages are rarely explicit. And there’s no consistent metadata an agent can use to know what a page is about before reading the whole thing. That’s the “SharePoint swamp” — technically searchable, but hard for a machine to actually use. The fix isn’t ripping these platforms out. It’s building, or exporting into, a lean, structured layer made for how an AI reads, with real ownership behind it.
A Format Worth Watching
Google Cloud published a first draft of its Open Knowledge Format in June 2026. It formalizes a pattern AI-native teams were already using: markdown files with a small block of structured metadata, organized so relationships between concepts are explicit. It’s deliberately simple — no proprietary database, no required software — and that’s the point. Treat it as a set of good habits worth adopting. For curated, stable knowledge, the structure can remove the need for large-scale retrieval search altogether, since an agent can just follow the links. For the large, messy, constantly changing content nobody’s had time to hand-curate, a retrieval system still fills the gap.
The Cost and Consistency Case
Two outcomes make the strongest case for this discipline: cost and consistency.
- Cost: Research comparing structured retrieval to naive, unstructured retrieval shows real token savings, commonly ranging from 25% to well over 90% for similar tasks. But structure alone doesn’t guarantee this — a poorly curated knowledge base can actually cost more tokens than a simple search. The savings come from the curation habit, not the technology label.
- Consistency: Research from Google DeepMind and University College London found that AI models can be surprisingly unstable across a long conversation. They sometimes abandon a correct answer when challenged, even by a wrong counterargument. A steady, curated source of truth the AI can return to is a direct defense against that drift.
Knowledge Is Never “Done”
Here’s the most important shift in mindset: Knowledge Management for AI is not a project with an end date. It’s an ongoing habit, much like data quality itself. Policies change, projects finish and create new lessons, and client relationships evolve. A Knowledge Tree that reflects reality in January but goes stale by June isn’t just unhelpful — it actively hurts the AI’s usefulness, and worse, it erodes trust in it.
This is where AI becomes part of the solution, not just the one benefiting from it. AI-assisted workflows can help Stewards draft updates and flag conflicts between related branches. They can even spot gaps by noticing when the AI keeps failing to find a good answer to the same question. Instead of waiting for a yearly audit, the system can nudge its own maintainers toward what needs attention. The goal is a knowledge base with a pulse, one that grows and fixes itself about as fast as the organization it describes.
A Knowledge Management Checklist for AI Enablement
Before investing in a Knowledge Tree, take an honest look at where the gaps actually are. We’ve identified seven properties that tend to separate a knowledge base that truly enables AI from one that just piles up documents:
- Centralized standards — A small, shared set of rules for file structure, required metadata, and naming, so an AI can read knowledge from different teams without translation.
- Distributed ownership — A named Owner, who is accountable, and Stewards, who are responsible, for every branch, so maintenance doesn’t bottleneck on a central team that can’t keep up with the business.
- AI-friendly formatting — Plain text, ideally markdown, structured so an AI agent can read it directly, without an extra translation or extraction step.
- Version control and change management — A record of who changed what, when, and why, so compliance teams and AI agents can both trust, and trace, what they’re reading.
- Accessibility to AI agents — A real connection point, like an MCP server, an API, or a supported integration, that lets your AI tools actually reach the knowledge, not just a folder humans happen to browse.
- Transparency — Human-readable and searchable by design, so the same knowledge base serves people and AI equally well, and nothing is a black box only a machine can read.
- Aggregation — The ability to combine knowledge across branches for one query, so the value compounds — an answer that draws on IT, compliance, and client context at once is worth more than any single branch.
Firms running Microsoft 365 Copilot often think this is already solved, because Copilot’s semantic index can already search across SharePoint and Confluence content today. That’s real, but it answers a different question. It makes existing pages easier to find. It doesn’t make them curated, current, or clearly owned. A well-governed Confluence space can score well on ownership and version history. But the pages themselves are still long-form prose, not the one-concept-per-file structure an AI reads best. Better search is not the same as better content. Most firms will find they have two or three of these seven properties checked off — and AI-friendly formatting usually isn’t one of them.
Conclusion
AI enablement is too often framed as a model choice or a tooling problem. It’s really a knowledge problem. The firms that pull ahead over the next few years will treat institutional knowledge as a curated, governed, always-evolving asset. It’s not a static pile someone builds once and forgets.
A Knowledge Tree gives your AI the same thing a great new hire eventually earns: a real understanding of how your organization works. It’s rooted in light governance standards, structured into owned and stewarded branches, and kept alive through steady curation. With the right structure in place, you don’t have to wait months for it to ramp up, and you don’t have to start over every time the technology changes.
Olmstead brings together three things most firms don’t find in one place: deep experience in the investment management industry, a thoughtful Data Governance practice, and the ability to develop AI solutions that work. We customize the knowledge structures, ownership models, and curation practices that let AI perform like it truly understands your business. Contact us today to learn more.