The knowledge base nobody argues with
Someone walked me through their knowledge setup recently, and I'll admit it was elegant. Connect the email, the calendar, the drives, the meeting transcripts, the CRM. Agents ingest the lot, link every person and project and decision into a living graph, and keep it fresh while you sleep. Ask it anything about the business and it answers, quickly, with sources.
It's a pattern I'm seeing more of, and the ambition behind it is real. Some versions go further and promise synthesis: the system doesn't just fetch, it extracts what matters, flags contradictions, tidies as it goes.
But underneath every version sits the same assumption. Pull enough of the organisation in, and understanding will follow.
That assumption has a name. Yuval Noah Harari opens Nexus by calling it the naive view of information: the belief that accumulating more of it carries us naturally toward truth and wisdom. It doesn't, he argues, and it never has. Piles of information have justified as much folly as insight.
I'd put the organisational version more plainly. Information accumulates. Knowledge is chosen.
A meeting transcript holds forty minutes of talk. It holds none of the thinking that happened before anyone spoke, and nothing of what everyone in the room already knew and so never said. And what mattered in the talk itself depends on your strategy, your history, the decision you took last quarter, the client you nearly lost in March. That context lives in the organisation and almost nowhere else. So when a system extracts "the key points" on its own, a choice still got made. The machine made it, with a borrowed sense of what matters, a salience learned from everyone's meetings and no one's business.
And a meeting transcript is the tidy case. The same pipeline takes in Slack and email, which are not records of anything. They are people thinking out loud, being wrong in passing, answering confidently from memory, settling something in one channel that gets reversed in another an hour later. Almost none of it was written to be relied on. Pull it in wholesale and it comes back later as an answer with a source attached, and the source lends it a weight it never earned.
Now, none of this is an argument against the machine doing the work. I use AI for exactly this extraction, week in and week out, and I'd be pretending otherwise if I said different; done well, it is the only way the work scales. The difference sits in two small, unglamorous places.
First, the machine is told what matters, in advance, in writing: what this kind of session produces, what must always be kept, what counts as a decision here rather than a musing. That instruction is the organisation's judgement, encoded.
Second, a person reads what comes back before it becomes the record, because sometimes the machine gets it wrong, and the review is where you catch it. The choosing of what, in all that information, counts as knowledge stays yours. And once it has been through a person, it carries a stamp everyone else can rely on.
There is a difference, in other words, between delegating the work and delegating the judgement. The connect-everything pipeline delegates both.
And a judgement that is never exercised weakens the way anything unused does. Quietly, over months, the organisation loses the ability to say what it knows and why it knows it.
There's a second choice these systems remove, further downstream, and it's the one that decides whether any of this survives contact with time. Capture is a moment; correctness is a condition, and conditions have to be maintained.
Knowledge stays correct for one reason only: someone noticed it was wrong. The pricing that changed last month. The escalation path that moved when a person left. And noticing happens in use, when the person who actually holds the knowledge, the ops lead, the consultant who created the exception, reads the thing itself and thinks: that's no longer right.
If the only door into the knowledge is an AI answer, that person never sees the thing itself. They see a fluent response, and fluency is very good at hiding staleness (this is roughly its job). The frown never lands on the page. The correction never happens.
And underneath all of this sits the trust problem, which is important to flag. You hear it everywhere at the moment: "I am not sure if I can trust what AI gives me". The instinct is sound. An answer you cannot trace back to something you can open, read, and question is an answer you have to take on faith, and organisations don't run on faith.
People trust what they can check. When an answer links back to a page they can open, current, owned, visibly maintained, trust has somewhere to come from. When it stands alone, fluent and unsourced, trust has nowhere to come from. So people quietly go back to what they can verify: the colleague who was in the room, the folder they keep themselves. The system built to end fragmentation ends up re-creating it.
To be fair to the engineers: their version of this, documentation kept as code, works. It works because the people who hold the knowledge and the people who comfortably touch the files are the same people, so both choices, what to keep and what to fix, stay with the ones who can make them.
An organisation is a different animal. Its knowledge mostly lives with people who will never open a repository, and shouldn't have to. Which means the environment itself has to be built for them: a place where the ops lead can open the page behind the answer, read it in plain language, question it in the margin, and put the correction in front of the person who owns it, right there on the page. Not an archive behind glass.
If people can only meet their organisation's knowledge through what the AI says about it, that knowledge has stopped being theirs.
So the question I'd put to anyone building their AI knowledge layer right now has two parts.
When something enters your knowledge base, who decided it mattered?
And when it stops being true, who will notice?
If the answer to both is "the system," you haven't built organisational knowledge. You've built an archive with a confident voice.
The healthiest knowledge environment I've seen wasn't the tidiest, or the fullest. It had an argument running in the comments, somebody working out in the open whether the page was still right, and whether it belonged there at all.
And when the argument settled, the page changed, the owner made it stick, and the next person to arrive inherits the answer rather than the argument. So does the AI answering from it.
Side note: And the elephant? I shared two last week on LinkedIn https://lnkd.in/p/dUCPVw7K. People seemed to like them, so here's another, same watering hole (Shompole if interested). Although, sitting above that headline, this one has ended up looking rather more relevant than intended.
Comments