Centralising knowledge: the one-stop shop your AI rollout is missing

Here's a pattern I see in almost every firm a few months into AI adoption. Usage is genuinely happening — pockets of it are impressive. But the knowledge about how to use the tools well is scattered across DMs, personal notes, half-remembered training sessions and individual heads. Five teams are solving the same problem in parallel, badly, because none of them know the sixth team solved it in March.

The firm builds five times and learns once — it should be the other way round. Every prompt that works, every skill that gets saved, every lesson from a failed experiment is intellectual property. Unmanaged, it evaporates. Centralised, it compounds. BCG's 10-20-70 finding — that 70% of AI transformation effort belongs in people and process^[1] — is usually read as a training point, but knowledge infrastructure is the other half of that 70%: the difference between a firm that learns as one organism and a firm that learns as forty individuals.

01One front door

The fix is boring and effective: a single internal hub — the one place everyone knows to go. Not a wiki nobody reads; a working destination with everything a user needs, from first login to advanced builds. When someone asks "where do I find…?", the answer is always the same URL.

02What goes on the shelves

Two design rules make or break it. First: contribution must be near-zero effort. "Send the prompt to the hub owner" beats "fill in this six-field form" — a curator can tidy later; friction at the point of contribution kills the pipeline at birth. Second: everything has an owner and a version. A skills library where nobody knows if the redline skill reflects the current standard position is worse than no library, because people trust it wrongly. This is governance work as much as knowledge work — the hub is also where the approved-tools list and the "what not to paste into what" rules live, one click from the assets they govern.

A firm that centralises its AI knowledge learns once and deploys forty times. A firm that doesn't learns forty times and deploys once — usually the wrong thing.

03Context is the compounding asset

There's a deeper version of this argument. The models are increasingly commodity — every firm has access to roughly the same intelligence. What a competitor cannot download is your firm's context: how deals get evaluated, what committee actually asks, which sectors you know cold, what the house style sounds like. Systematising that knowledge — capturing it, structuring it, making it available to both people and models — is the single move that unlocks the higher-value use cases, because almost all of them depend on it. A queryable knowledge base is only as good as what got captured; an IC-question skill is only as sharp as the historical decisions behind it.

The industry has quietly converged on the same conclusion. Anthropic's Agent Skills — folders of instructions and resources a model loads on demand — are, at heart, institutional knowledge made executable, and they became an open standard precisely because organisations needed a portable format for exactly this^[2][3]. MIT's enterprise research found the failed 95% of pilots shared a common flaw: tools and organisations that couldn't retain feedback or accumulate learning^[4]. The hub is the retention mechanism.

04Start smaller than you think

None of this needs a platform procurement. Version one is a well-structured space in whatever you already have — SharePoint, Notion, an intranet page — with six sections and a named owner. It can be live in a week. What it needs isn't budget; it's curation, a couple of champions who seed it with their best material, and communication channels that keep pointing people back to it. Setting up exactly this — structure, seeding, governance and the operating rhythm around it — is part of what an embedding engagement covers.

One URL. Everything in it. Start this week.