Field notes · essay
Manage flow, not inventory
Ask a knowledge program how it's doing and you'll usually get a count. Twelve thousand articles. Four hundred new ones last quarter. Coverage up six points. Every one of those numbers describes a warehouse: what's on the shelves, how full they are, how fast the shelving is growing.
A warehouse is the wrong picture. A knowledge base isn't a stockpile you fill; it's a current you keep moving. Every answer in it is doing one of two things: climbing toward a permanent fix, or overdue for retirement. There is no third state where an answer sits on a shelf and stays true forever. Programs that behave as if there is end up with the museum the founding essay warned about: a corpus of things that once were true.
The trade at the center of the Content Graduation Model fits in one sentence: stop managing content as inventory in channels, and start managing it as flow between them. Here's what that trade looks like once you actually run it.
Six tiers, traffic in both directions
The trust ladder has six tiers. At the bottom sits conversation: forum threads, chat, case notes. Abundant, fast, close to real problems, and unverified. Above it, community-validated answers. Then knowledge base articles. Then official documentation. Then training and rich formats. At the top, the product itself: in-product guidance, better defaults, outright fixes, where the knowledge arrives with zero reading required.
Each rung adds validation, structure, reach, and cost. Content earns every move through demonstrated demand and validated quality, never through where it happened to be authored. A forum answer that keeps getting accepted is a knowledge base article waiting to be claimed. A procedure thousands grind through every month is a guided flow waiting to happen. A workaround everyone depends on is a product requirement.
Inventory thinking only sees the up direction, because up adds to the count. Flow thinking runs the ladder both ways. Demand below threshold: archived. Superseded: retired with a redirect. Invalidated by a release: withdrawn before it misleads. Made unnecessary by a product fix: celebrated, then removed. The down direction isn't cleanup you get to when there's time. It's half the operation, and once an AI assistant grounds in the corpus, it's the half that keeps the assistant from confidently prescribing last year's workaround.
MTTE, the clock that finds product gaps
Inventory metrics can't tell a knowledge gap from a product gap; both show up as an article with a lot of views. MTTE can.
Mean time to event is the average interval before the same question, error, or issue resurfaces. It converts "I feel like we answer this a lot" into a frequency you can rank and a clock you can re-check. Then it splits the signal in two. A question with a short MTTE is a standing gap in the knowledge layer: promote the answer, or create the coverage. A workaround with a short MTTE is a standing gap in the product.
The second case is where inventory thinking does real damage. A popular workaround article wins on every inventory dashboard: high views, strong helpfulness votes, a permanent seat in the top ten. Run the arithmetic instead. An MTTE of eleven days means somebody hits the same wall about thirty-three times a year, and every one of those hits is an answer someone pays for: you in channel cost, the customer in time. Cost per answer can only fall so far on a question that keeps getting asked. The cheapest answer is the one nobody needs.
A recurring workaround is a product requirement in disguise, and your most-viewed article may be your best-documented bug.
So route it. On a standing cadence, content strategy and product management review content demand together and ask one question: which of our most-needed content should not need to exist? Candidates land in the backlog pre-evidenced, with demand data, the affected population, and the full text of the current workaround attached. Nobody has to argue the problem is real. The corpus already wrote it down.
Machines detect, humans decide, machines carry
Flow needs a pass that runs on a schedule, or it settles back into inventory. The graduation pass divides the labor along one line.
Machines detect. Harvest sweeps the source channels on a schedule; quiet candidates get re-checked when their MTTE horizon elapses instead of being forgotten; ratified thresholds sort clean passes from clean fails automatically and route the in-betweens to a person.
Humans decide. Someone accountable confirms the answer is correct, safe, and complete. That step does not automate. It's the difference between a knowledge layer and a rumor mill, and the same gate governs the way down, because retiring the wrong asset damages trust exactly as much as promoting a wrong one.
Machines carry. The rewrite to the destination channel's standard, the publish, and the loop closure: the graduated link posted back on the thread where the question first lived, crediting the person who answered it. Loop closure isn't a courtesy. It's how contributors learn the ladder exists, and it turns a community into a willing supplier instead of a mined resource.
AI speeds up every carrying step and most of the detecting ones. It never gets the middle one.
Measure the current, not the warehouse
A knowledge program run as flow gets measured on speed and direction. Time from detection to publication. The share of new assets that came from harvested demand rather than speculation. The share of the corpus still inside its re-evaluation horizon. Workarounds converted into product fixes per quarter, with the retired content to prove each one.
Notice what's missing: the article count. A program can grow its count every quarter while its flow stalls, and that is the warehouse failure in a single sentence. The reverse holds too. A shrinking workaround library is a quality metric. It means the top rung of the ladder is doing its job.
Which brings me to the most trust-building sentence a corpus can contain:
"Fixed in version 4.8: this workaround is no longer needed."
No new article can do what that sentence does. It tells the reader the problem is gone. It tells them someone was watching. And it pulls an answer out of the grounding layer before an assistant can serve it six months after it became wrong. Workaround library, minus one. That's the metric moving the right way.
Keep the shelves if you like. Just stop counting them.
The full model (the six tiers, the four signal families behind every move, the graduation pass, and the cadence that routes demand into the product) is published at machinereadyknowledge.com/graduation. Why the corpus now grounds machines as well as people is the subject of the doctrine.
KCS® is a service mark of the Consortium for Service Innovation™. This essay reflects an independent, KCS-informed methodology and is not affiliated with or endorsed by the Consortium.