Latest / Future of Work Tech with Fexingo: Remote Tools, AI Productivity, and Workplace Software / How AI Is Automating Your Internal Knowledge Base
Transcript
- Lucas: Luna, when was the last time you actually searched your company's internal knowledge base and found exactly what you needed on the first try? Luna: Honestly? I think it was... maybe two jobs ago. At my last company, the wiki was a graveyard of outdated processes and half-finished templates. Lucas: Right, that's the norm. And it's a massive productivity drain. A 2025 study from the International Association of Business Communicators found that employees spend an average of two point five hours per week searching for internal information. That's over six full workdays a year. Luna: And that's just search time, not counting the cost of acting on wrong information. Lucas: Exactly. So a new wave of AI tools is trying to fix this. Not just better search, but actively maintaining and even generating knowledge base content. I've been following a company called TechFlow — midsize SaaS firm, about eight hundred employees. They rolled out an AI layer across their Notion and Confluence environments back in Q4 of last year. Luna: What does that actually look like? Is it a chatbot that answers questions from the docs? Lucas: That's part of it, but the more interesting piece is the content maintenance. The AI can flag outdated articles, suggest merges for duplicates, and even auto-generate draft updates based on recent Slack conversations or ticket resolutions. TechFlow reported a thirty percent reduction in new-hire ramp time — from eight weeks down to about five and a half. Luna: Thirty percent — that's huge. But I wonder, does the AI actually understand context, or is it just doing keyword matching on steroids? Lucas: Great question. What makes these tools different is they're building what amounts to a knowledge graph on the fly. They don't just index documents — they map relationships between concepts, projects, people, and decisions. So when a new engineer asks 'How do we handle database migrations?' the system can pull not just the migration doc, but also the relevant Slack thread from last month, the Jira ticket that triggered it, and the post-mortem from when it broke. Lucas: That's the trade-off. The TechFlow team addressed it by keeping the AI summaries at a 'sufficiency' level — enough to get you started, with explicit links to sources. They measured that users clicked through to the original documents about forty percent of the time, which they considered healthy. The AI wasn't replacing the source, it was serving as an intelligent index. Luna: Forty percent click-through feels high to me. I'd expect people to take the summary at face value more often. Lucas: Possibly. But they also trained employees to treat the AI output as a starting point. They ran lunch and learn sessions showing specific examples where the summary missed a critical exception. That cultural buy-in is probably the hardest part. Luna: It reminds me of the early days of Wikipedia. People learned pretty quickly that an article is not the final word — you check the footnotes. So the AI is like a dynamic, internal Wikipedia with a chatbot interface. Lucas: That's a good analogy. And like Wikipedia, the biggest challenge is content decay. TechFlow found that about eighteen percent of their knowledge base was outdated by more than six months. Their AI now flags articles that haven't been reviewed in that window and either auto-updates them from recent activity or tags a human owner. Luna: Is the human in the loop necessary? Could the AI just update everything automatically? Lucas: They tried that in a pilot and it created more problems than it solved. The AI occasionally made plausible-sounding but wrong updates — like changing a process based on a one-off exception. So now the system generates a draft, the content owner gets a notification, and they approve or tweak it. The approval rate is about seventy percent. Luna: That seems like a reasonable balance. But I'm curious about the cost. Running an AI that continuously indexes, summarizes, and maintains a knowledge base can't be cheap. Lucas: TechFlow's CTO shared some numbers at a conference last month. They're spending about twelve thousand dollars per month on the AI layer — that's for API calls, model hosting, and a small vector database. For eight hundred employees, that's fifteen bucks per person per month. They estimate it saves about two hours per employee per week in search time alone. At an average loaded cost of, say, fifty dollars an hour, the math works out to a return of about six to one. Luna: That's a compelling ROI. But does it require a certain baseline of documentation quality to begin with? If your existing knowledge base is a mess, can AI really make it useful? Lucas: That's one of the most common pitfalls. The AI is only as good as the raw material. TechFlow had a relatively well-structured Notion environment — they'd already invested in templates and naming conventions. Companies that dump a bunch of random PDFs and expect the AI to sort them out are going to be disappointed. The tool works best as an amplifier, not a foundation. Luna: So it's like renovating a house — you need the drywall up before you can paint. What about adoption? Did employees actually use it, or did it sit on a dashboard? Lucas: Adoption was surprisingly high. Within three months, about sixty-five percent of employees had used the AI assistant at least once a week. The key was embedding it directly into their existing workflows — it integrated with Slack, so you could just type /ask followed by your question and get an answer without leaving the chat. That frictionless access is critical. Luna: That's smart. If you have to open a separate portal, people forget it exists. Slack integration makes it feel like a colleague you can ping. Lucas: Exactly. And the system learns from those interactions. If people ask the same question multiple times, it surfaces that as a candidate for a new FAQ entry. It also tracks which answers people click 'thumbs down' on, so content owners know what's confusing. Luna: I can see that becoming a virtuous cycle. But I want to push back on something — isn't there a risk that by making knowledge too easy to retrieve, we discourage people from actually learning and retaining information? If you can always just ask the bot, why bother remembering anything? Lucas: That's a valid concern. It's the 'Google effect' extended to internal knowledge. But I think the counterargument is that what you want people to retain is the conceptual framework, not the specific procedures. Knowing that the database migration process exists and where to find the latest version is more important than memorizing each step. The AI handles the memory load, freeing up cognitive capacity for higher-level thinking. Luna: I suppose that's true if the AI is reliable. But what about the social side? When I used to ask a colleague a question, I'd often get additional context — like 'oh, by the way, Legal is reviewing that policy, so check back next week.' Does the AI capture that kind of informal knowledge? Lucas: That's the frontier. Some tools now integrate with email and chat to pull out implicit knowledge — like a manager mentioning in a team stand-up that a certain approval process is changing. But it's tricky. TechFlow's approach was to encourage employees to tag their Slack messages with a knowledge base flag if they think the information is shareable. That way, the system picks up the informal update without invading privacy. About twelve percent of their knowledge base now originates from Slack. Luna: So it's a hybrid — some explicit curation, some passive capture. I think that's the right model. But I'm still wondering about one thing: how do you measure success beyond ramp time and search hours? Are there qualitative benefits? Lucas: TechFlow did an internal survey and found that employee confidence in decision-making improved significantly. People felt more empowered to act because they trusted they could find the right information quickly. Also, the number of repeated questions to senior engineers dropped, which freed up the experts to work on more complex problems. One team lead said it was like having a junior team member who never forgets anything. Luna: That's a great line. But I wonder if there's a ceiling. Once you've indexed all your explicit knowledge, the next frontier is tacit knowledge — the stuff that's never written down. Do you think AI will ever get there? Lucas: That's the holy grail, and we're still early. Some companies are experimenting with tools that record and transcribe internal meetings, then extract decisions and action items into the knowledge base. But it raises a lot of questions about privacy, consent, and information overload. I think we'll see more of that in the next twelve months. Luna: It's a fascinating space. And it's the kind of thing that directly impacts how we work every day. Speaking of which — this show itself is ad-free and independent, and that's possible because of a small group of listeners who chip in through buy me a coffee dot com slash fexingo. It's not a big ask; just mentioning that if you get value from these deep dives, that's the engine that keeps them coming. Lucas: Yeah, I appreciate that. And we don't interrupt the flow with sponsors, so this really is a listener-funded project. It's cool that it works. Luna: Alright, back to knowledge bases. Lucas, what's one thing you'd recommend to a company that wants to start this journey tomorrow? Lucas: Start small. Pick one team — ideally one that's already organized and motivated — and pilot the AI layer with them for a month. Measure their before and after search time, ramp time for new members, and user satisfaction. Let the data speak, then scale. Don't try to boil the ocean. Luna: Sound advice. And maybe start with cleaning up that existing wiki first. Lucas: Absolutely. A clean foundation makes all the difference. That's the bit that no AI can shortcut — yet.