The Knowledge Exchange pairs a small team of trusted human curators with AI agents that handle the routine work of keeping knowledge connected, current and useful. The humans hold the judgement. The agents make the tending affordable.
A named lead gardener is accountable for everything the agents do. Anything sent to a person is sent by a person.
Humans hold the judgementThey repair links, spot stale content, draft welcomes and flag what people keep searching for and not finding.
Drafts only — never sendsThe exchange meets practitioners where they already work. In shared channels, agents only speak when spoken to.
No new platform to adoptEach connection the exchange draws explains itself, and one click reverses any agent's work.
Tested before launch, watched afterEvery applied piece of knowledge becomes evidence of change moving through the sector. The data belongs to you.
Collected with consentHow it works, in one picture. Practitioners meet the exchange where they already work. Agents do the continuous background work. One to five gardeners make every decision that reaches a person. The dashed line matters most: when the best answer is another practitioner, the agents make the introduction.
The evidence is clear about why. These platforms don't fail on technology. They fail because nobody is paid to keep them alive: to welcome contributors, connect new material to the people who need it, and notice what's missing. That work is called curation and facilitation, and it's almost never funded past the launch grant.
What's left behind is familiar to every practitioner: a well-built site full of well-written reports that nobody reads. A bookshelf with analytics.
Before designing anything, we asked what's true about knowledge in practitioner communities, regardless of technology. Four things held up.
A report that changes what a practitioner does tomorrow is worth more than a hundred that don't. Application is the only honest measure of quality.
A practice transfers when it fits the new community's situation, constraints and capacity. Today that matching happens by luck, or not at all.
Curation and facilitation are what keep an exchange alive, and both take time from people. This is exactly what the sector doesn't fund.
Practitioners trust named peers and known intermediaries. A design that asks a community to trust software directly won't get that trust.
A platform dies when any one of these jobs goes undone. The empty bookshelf is what remains when all four stop.
Linking knowledge to the people and places that need it. Left undone, knowledge piles up in an archive.
Spotting what's going stale, what's missing and what people keep asking for. Left undone, the exchange goes silent.
Thanking contributors, filling gaps, improving what's there. Nobody writes for a place that doesn't write back.
Built slowly through reliability and human presence. Lost instantly with one wrong answer in the wrong place.
At the centre are one to five trusted humans we call the gardeners. The agents work around them: scanning for broken links and stale content, suggesting connections, drafting welcome messages and requests for contributions, and flagging what practitioners keep searching for and not finding.
The gardeners review, approve and overrule through a dashboard built for the job. A named lead gardener is accountable for everything the agents do, and anything sent to a person is sent by a person. When investment grows, it buys more human gardening before it buys more automation, because five part-time gardeners spread across the field hear more, and are trusted by more people, than one full-time curator.
The weekend scan lands as a to-do list: three broken links repaired and logged, two stale pages with refreshed drafts attached, and one gap where fourteen practitioners searched and found nothing. A gardener clears the list in under an hour.
An agent drafts a welcome for a first-time contributor and a request asking a named practitioner to write about the gap. A gardener edits two sentences and sends both.
A practitioner's question touches on a small community that could be identified from the detail. The agent hands the conversation to the gardener who knows that community, and tells the practitioner it has done so.
The gardeners spend half an hour reviewing a sample of the week's agent conversations, looking for anything off. This is routine, not a response to an incident.
A practitioner searching on their own gets a conversation, not a list of documents. The exchange knows their context from previous visits and helps them work out how to apply what they find.
In the Teams channels, Slack workspaces and email threads where the community already gathers, the rule is simple and enforced in software: agents answer questions and fetch material when asked. They never start conversations, never correct anyone in public, and never pretend to be members of the community.
We don't claim the agents will never make a mistake. We've designed the system so that no mistake is permanent, and most of the system isn't AI at all. Six commitments hold this together.
Routine work like link checks and content scans runs as ordinary scheduled code. AI is used only where a task needs judgement.
We use the smallest model that passes our tests, and we test again every time anything changes. This keeps running costs sustainable for the long term.
Every action is attributed and can be undone with one click. Nothing is ever deleted, except when a contributor asks for their material to be removed. That request is always honoured.
When the exchange links a story to a practitioner, it says why. There are no unexplained recommendations.
The agents pass a test suite before anything goes live, and the gardeners review their work as a weekly habit.
Community-valued material is a protected class the agents cannot alter or retire. The defaults are conservative on purpose.
We also ran the design through a pre-mortem: eight plausible ways a platform like this fails, each answered by a decision already in the architecture. And the whole exchange runs inside your existing IT environment and identity, so there is no new system for your technology partner to host, secure or say no to.
Change in this sector shows up first in relationships and practice, long before it shows up in population data. No current tool can see that layer. A working exchange can, because seeing it is a side effect of doing its job: when a practitioner finds a practice, adapts it and reports how it went, that's a record of practice spreading through the sector.
This data is collected with practitioners' knowledge, aggregated before it travels, and belongs to you alone. And it compounds: every month of operation makes the record more valuable, which is a reason to start sooner rather than later.
A 2019 story that fits a practitioner's situation is worth more than this quarter's report that doesn't. So the agents spend their effort on matching stories to contexts, not on tidying. Old content isn't pruned by default; its connections are tested and renewed.
Where practitioners keep searching and finding nothing, the exchange says so, and asks a named person to fill the gap. Quality checks appear as friendly prompts ("this story would travel further with an outcome note"), never as gates. And the agents never rewrite a contributor's words, because a community stops recognising itself in sanitised prose.
Nobody should buy this on faith. Each stage produces the evidence that justifies the next, and each has an agreed point where you decide whether to continue.
A short first engagement produces working prototypes on realistic material: conversational search, agent-assisted curation and the gardeners' dashboard. It ends with a demonstration on a set date, judged against go or no-go criteria we agree at the start.
Your staff become the first community of practice, on your existing systems. The approach gets tested with a forgiving audience, and the evidence builds before anything faces the sector.
It goes public carrying test results, trained gardeners and a real track record instead of promises. This is also when the measurement record switches on.
You own the specification and the data throughout. The design can be rebuilt on any platform, everything exports in open formats, and we name an archival home for the stories from day one. A design built on the observation that platforms die owes you an answer about where the knowledge goes if this one ever does.
Practitioners will never see any of this architecture. They'll only notice whether the exchange feels looked after, current and worth returning to. We'd like to show you how it stays that way.
Get in touch