Latest / Future of Work Tech with Fexingo: Remote Tools, AI Productivity, and Workplace Software / How AI Is Automating Your Internal Knowledge Curation
Transcript
- Lucas: If you work in a company of more than, say, five hundred people, you have probably experienced this — you search the internal knowledge base for something like 'expense policy for client dinners', and you get back forty-seven results, half of which reference a policy that was replaced two years ago. Luna: And you end up asking a colleague who's been there longer, which kind of defeats the purpose of having a knowledge base in the first place. Lucas: Exactly. That's the problem that a growing number of companies are trying to solve with ai powered knowledge curation — not just search, but active maintenance: flagging outdated content, suggesting new articles based on what people are actually asking, and even automatically linking related documents. Luna: So instead of a librarian, you have an algorithm that's constantly tidying up the shelves. Lucas: Right. And one of the more interesting implementations I've seen is at a mid-sized fintech company — about twelve hundred employees — that rolled out a curation engine last year. Their knowledge base had over eight thousand articles, and the head of knowledge management told me that before the AI, roughly sixty percent of those articles hadn't been touched in over eighteen months. Luna: That is a lot of digital cobwebs. What did the AI actually do? Lucas: It did three things. First, it analyzed access logs and search queries to identify which articles were actually being used and which were essentially dead. Second, it flagged content that contained references to outdated systems or policies — for example, if an article mentioned a software platform the company had since replaced. And third, it generated suggestions for merging or archiving content that overlapped significantly. Lucas: The result after six months? The number of articles dropped from eight thousand to just over five thousand, but the time employees spent searching for a specific policy dropped by forty percent. Fewer results, but better ones. Luna: I can see the efficiency gain, but I also wonder about the gatekeeping aspect. Who decides what the algorithm considers 'relevant'? Is it possible that the AI could bury something that's actually important just because it's not accessed frequently? Lucas: That is the big tension. The fintech company handled it by keeping a human in the loop — the AI would flag articles for review, but a team of three knowledge managers had the final say on archiving or merging. They also set up a feedback mechanism: if an employee searched for something and didn't find it, that triggered a review of whether the AI had incorrectly hidden something. Luna: So it's more of a recommendation system than an automatic deletion machine. That seems like the right balance. Lucas: It is. And I think that's the model most companies are moving toward — AI as a curator that suggests, not a dictator that deletes. But the technology is improving fast. There are now tools that can detect semantic drift: for instance, if a policy document uses language that's no longer consistent with the company's current terminology, the AI will flag it. Luna: That sounds like a natural language processing application. How does it work in practice? Lucas: It compares the language in the document against a corpus of recent internal communications — emails, chat messages, updated policy docs — and looks for mismatches. If the knowledge base still says 'personnel department' but everyone internally now says 'people team,' the AI will surface that as a candidate for revision. Luna: That is genuinely useful. I can imagine a company that's gone through a rebrand or a merger — those language shifts can be really hard to track manually. Lucas: Absolutely. And the same technology can be applied to external-facing content, but the internal use case is where we're seeing the most adoption right now, partly because the stakes are lower — if the AI makes a mistake, it's an inconvenience, not a compliance violation. Luna: But there are compliance considerations too, right? If you're in a regulated industry, you can't just let an algorithm decide to archive a document that might be needed for an audit. Lucas: That's a critical point. The fintech company I mentioned operates under financial regulations, so they had to ensure that any content flagged for archiving was first checked against retention requirements. They built a rule engine that overrode the AI's suggestions for any document that fell under a regulatory hold. Lucas: So the AI wasn't even allowed to suggest archiving certain categories of content — like customer transaction records or compliance training materials. That's a sensible design choice. Luna: It also raises the question: how much do you trust the AI to get it right? If you're a knowledge manager, you might be skeptical of an algorithm that says 'this article is redundant' when you know it contains some niche but crucial detail. Lucas: And that skepticism is healthy. The vendors in this space are very aware of that, which is why most of them emphasize that the AI is a decision-support tool, not a decision-maker. They also provide transparency reports — showing which articles were flagged, why, and what the confidence score was. Luna: Confidence scores — so the knowledge manager can prioritize their review based on how sure the AI is. Lucas: Exactly. And over time, as the AI sees more human decisions, it gets better at predicting which flags will be accepted. That feedback loop is where the real value comes from. Luna: It reminds me of how spam filters evolved — they started with simple rules, then moved to machine learning that learned from users marking emails as spam. Lucas: That's a perfect analogy. And just like spam filters, knowledge curation AI will never be perfect, but if it catches eighty percent of the stale content, that's already a massive time saver. Lucas: Now, I want to shift slightly to another dimension of this — proactive curation. Instead of just cleaning up old content, some tools are now suggesting new content based on gaps they detect. Luna: What does that look like in practice? Lucas: Let's say the AI notices that employees are frequently searching for 'return to office policy' but no single article has a high click-through rate. It might suggest that the existing content is fragmented or outdated, and recommend that a knowledge manager create a consolidated, updated article. Luna: So it's not just curating what exists, but identifying what's missing. Lucas: Right. And this is where we're starting to see some interesting overlap with generative AI. A few platforms now offer the ability to draft a first version of that missing article using the company's existing content as source material. Luna: But then you have to be careful about hallucination — the AI might invent a policy that sounds plausible but isn't actually correct. Lucas: Absolutely. The fintech company I mentioned actually piloted this feature but decided against using it precisely because of that risk. They said it was faster to have a human write the article from scratch than to verify and correct an ai generated draft. Luna: That's an honest assessment. It shows that the technology is still maturing. Lucas: It is. But even without full generative capabilities, the curation piece alone is delivering real productivity gains. And I think over the next twelve to eighteen months, we'll see more companies adopt these tools, especially as the cost of running the underlying models continues to drop. Luna: Speaking of costs — and I know this is a bit of a pivot — one thing I appreciate about this show is that we don't have ads. That's pretty rare for a daily podcast. Lucas: Yeah, it is. And the only reason that works is that a small group of listeners chip in through buy me a coffee dot com slash fexingo. It's not a big operation — just enough to cover hosting and keep the thing ad-free. Luna: If today's conversation gave you something you can use at work, that's the kind of signal that helps keep this going. No pressure ever. Lucas: Exactly. So — back to where we were. The other trend I'm watching is the integration of curation AI directly into the tools people already use, like Slack or Microsoft Teams. Luna: Instead of having to go to a separate knowledge base portal? Lucas: Right. So when someone types a question in a channel, the AI can surface the most relevant article as a suggested reply. That makes the knowledge base feel less like a library you have to visit and more like a helpful colleague who's always listening. Luna: That's a much lower friction point. I can see that driving adoption significantly. Lucas: It does. The fintech company saw a fifty percent increase in knowledge base usage after they integrated the curation engine into their chat platform. People didn't have to change their behavior — the knowledge came to them. Luna: So the ultimate goal is that knowledge management becomes invisible. You just get the right information when you need it. Lucas: That's the vision. And I think we're closer to it than most people realize. The pieces are there — semantic understanding, usage analytics, integration APIs. The challenge now is more about change management and trust than about the technology itself. Luna: It's interesting that the human element ends up being the hardest part, even in a show about automation. Lucas: It almost always is. The tools can be brilliant, but if the team doesn't trust them or doesn't have the right governance in place, they won't deliver. I think that's the takeaway for anyone listening who's considering a curation tool — invest as much in the change process as in the software.