This is a post about knowledge management, steeped in a boozy metaphor…
Knowledge management is quietly becoming context management. Generative AI didn’t cause the shift, it just made it impossible to ignore.
I’ve spent enough years around distilling to know where the romance goes: straight to the copper. Visitors want their picture taken next to the still. Nobody asks to see the fermenters. But any distiller will tell you the spirit is made long before it touches metal. It happens in the grain, the mash, the fermentation, and above all in the cuts, the decidedly unglamorous discipline of adjudicating what makes it into the bottle and what goes down the drain. The still only concentrates what you feed it; it doesn’t correct it. Give it a thin, sloppy wash and it will hand you back the same flaws at higher proof: faster and cleaner-looking than what went in, and utterly sure of itself. Starting to sound familiar?
Generative AI is the shiniest still ever built. And right now, organizations everywhere are pouring everything they’ve ever published into it: every stale article, every contradictory answer, every workaround that stopped being true two releases ago… and wondering why the spirit burns going down.
That, in one image, is what’s happening to knowledge management. The tools can very capably write. They can summarize, translate, and “answer.” But the story was never really about words, words just got cheap. Meaning didn’t. When content can be generated on demand, the competitive edge moves upstream, into the unglamorous work of capturing, governing, and delivering context.
Every article now has two readers
Traditional KM had a clear job: capture answers, organize content, publish articles, make things findable. That work is still necessary. It’s just no longer sufficient; the audience changed underneath us.
Every article you publish today has two readers. The first is the one you designed for: a human, skimming, squinting at a screenshot, filling gaps with judgment. The second arrived over the last few years with no announcement: the retrieval system feeding your chatbot, your agent-assist panel, your customers’ own AI tools. That reader doesn’t skim and doesn’t infer. It pulls a section out of context and builds an answer on top of it. Machines don’t fill gaps, they fill them with confidence.
The machine reader also has a bigger job than retrieval: it composes. To compose well it needs to know who’s asking, what they’re trying to accomplish, where they are in the workflow, which constraints apply (product version, entitlement, region, policy), and what’s already happened. None of that is “content.” All of it is context.
So the future of knowledge management is not a bigger pile of articles; most organizations have too many already. It’s better context signals, real governance, and feedback loops tight enough that AI can produce useful output without making things up and without taking your organization down with it.
What the new still does, beautifully
I don’t want to be the guy scowling at the shiny equipment, because the upside is exceptionally real… when the wash to charge the still is clean.
When provided structured case notes and strong article standards, AI is a superb first-draft machine. The reframe I push on every knowledge program: capture is labor, so price it. Cap what you ask of the person in the queue at thirty seconds (the messy notes, the error string, the two-line fix that actually worked), and let AI draft the article from that raw material. A human curator with a three-minute budget and real authority either approves or kills it. Twenty years we’ve spent asking our scarcest resource (expert attention, mid-queue) to handle the one step in the pipeline that’s easiest to automate: writing prose. That era is, thankfully, gone.
AI is also very good at resurfacing known solutions earlier (most support volume is a rerun) and at compressing complexity: case timelines, escalation summaries, the leadership read-in nobody has time to write. Pattern-finding too, across thousands of cases and articles, at a pace no analyst can match. It hands you the pattern so you can spend your time on the story behind it.
Where it burns
The failure modes aren’t mysterious, and surprisingly none of them are the model’s fault. A system that is confident, fast, and accountable to no one will happily publish something that sounds right but is wrong… which is why the rule at the heart of everything I’ve written on this is four words long: AI drafts, humans ship. Nothing AI-generated goes in front of a customer without a human gate, and no gate is worth much without traceable sources behind the draft.
Point a model at a messy corpus and it won’t fix the mess. It will serve the mess: too many options, conflicting guidance, retired truths delivered as current ones. This is why content lifecycle management just got promoted from clean-up work to strategic discipline: it’s the cuts. Retire the obsolete, merge the duplicates, and be honest about what deserves the bottle. And the one that keeps me up at night: an ungoverned corpus wired to an eager model is how PII, customer confidential information, and internal IP walk out the front door. And what helps me sleep like a baby: the core KCS® instinct here was right all along… capture it, reuse it, improve it, and move as much of it as you safely can toward public, external content… because the more of your knowledge that can live in the open, the less you have to fear on the day you finally point AI at it.
Keep the soul, replace the machinery
I came up in my career by learning, adopting, and becoming a champion for KCS, and I’ll defend its ideals to anyone: capture in the workflow, reuse is review, learn through iteration, sufficient to solve. Those principles are as true as ever, but the assumptions underneath the methodology (articles written by humans, for humans, justified by deflection numbers… and deflection is a messy conversation worthy of a post or series of its own) belong to a world that no longer exists.
The work I’ve been doing, and writing down as I go, is keeping the soul while rebuilding the machinery. A few pieces of that rebuild:
- A governed knowledge loop – capture, curation, and delivery orbiting a versioned system of record, with a human curation gate that is non-negotiable. That gate is the difference between a knowledge layer and chatter in a rumor mill.
- Content managed as a flow of assets, not static inventory. Picture your channels as a ladder of increasing trust: forum threads at the bottom, validated articles and official docs above, and at the very top, the product itself. Content earns its way up through demonstrated demand and validation, and the highest graduation of all is for content to stop being content. A workaround that hundreds of customers depend on is a product defect wearing a costume. Route it into the backlog, ship the fix, retire the article, and celebrate. A workaround library that keeps shrinking means the product keeps getting better.
- … And honest economics. Deflection measures what didn’t happen, and your CFO knows it. Cost per answer measures what did: every answer your organization delivers travels through some channel, and each channel has a real price. An assisted answer costs somewhere between a nice lunch and a decent bar tab; a self-served one costs pennies. Once you start seeing answers that way (as things with a unit cost, whoever or whatever delivers them) knowledge work stops being a faith-based line item in somebody’s budget.
What the machines still can’t do
Nobody’s model can pull thirty years of lived experience out of a support engineer’s head. Making tacit knowledge explicit, translating messy human reality into usable and reviewable guidance, remains a human craft, and it’s the craft this whole knowledge industry is built on. Trust doesn’t come out of the model either; it gets made at that human gate, one validated answer at a time. I’ll borrow a line from a respected colleague that deserves to be tattooed on every AI roadmap: “Should is not a use case.”
If your AI rollout forces your people to contort their workflow to feed the model, you’re distilling for that copper still instead of the drinker.
The winners of the next few years won’t be the teams with the most content. They’ll be the ones with clean inputs, governance that actually governs, and the discipline to retire things on purpose; teams treating context as an asset rather than exhaust. That’s the shift, as plainly as I can put it: knowledge keeping got us here. Context keeping is what comes next.
So here’s the question worth sitting with: which reader is your knowledge base actually written for, and who is making the cuts?
I’ve also written this thinking down, properly, rather than leaving it scattered across blog posts. The full doctrine and operating model are free to read at contextkeeping.com, including a one-page maturity model you can grade yourself against, with no email wall. The machine-ready article standard lives at machinereadyknowledge.com. And for individuals or teams that want the working parts instead of just the argument (this includes article templates, curation gates, governance policies, and the capture prompt library), the toolkit modules are self-serve on the site, available for purchase individually or as a bundle.
Transparency note: AI was used to initially draft this article. But I dogfood my own tools and practices. As such I have put this through the rigorous paces of the KCS and ContextKeeping doctrines to ensure accuracy and polish are all executed by me, personally, before publishing… Always a human in the loop to validate the content. The end result is an article that remains more me than machine.
Featured image art by Howard O’Donnell, ©2014 HowardODonnell.com
KCS® is a service mark of the Consortium for Service Innovation™.The above content has been remixed/derived from the KCS v6 Practices Guide by Consortium for Service Innovation which is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. Permissions beyond the scope of this license may be available at info@serviceinnovation.org.
You must be logged in to post a comment.