Inventory Is in One Person's Head

The matrix that knows everything except the work

A retirement wave shows up in the age data two years before it lands, and every serious organization responds the same way. Someone builds a matrix. Silos across the top, names down the side, cells colored by who can do what, and a column with one expert and no backup is a single point of failure circled in red. The matrix is real work and it is not wrong. It tells you who to worry about. What it does not tell you, and was never built to tell you, is the one thing the decision turns on, which is what exactly is inside the red cell. I have sat in the room where the wall of cells gets reviewed: a team agrees the master machinist in the corner of red is irreplaceable, resolves to capture his knowledge, and the capture turns out to be an exit interview or a request to write up the procedures. Six months later the man is gone, the procedures read like a parts list, and the first time the line drifts nobody can say why. The matrix finds the person. The knowledge is the asset, and it was left as a formality. That gap, between knowing whom to fear and knowing what they hold, is where most continuity programs quietly fail, and the people who study this field have been saying so since before the matrix software existed.

What the canon actually found

The founding text of this field is twenty years old and still unbeaten on the diagnosis. David DeLong's Lost Knowledge^[1] made the move that everything since rests on: stop counting heads and start from capabilities. Ask leadership which strategic capabilities are at risk, then trace those down to the people who carry them. His other contribution is the taxonomy that explains why documentation keeps failing. He splits knowledge into four forms that each die a different death: human knowledge is skill and heuristics, social knowledge is the trust network of who calls whom, cultural knowledge is how decisions really get made, and structured knowledge, the documents everyone points to as the safety net, decays the moment the people who understood its context leave, because the document records the what and never the gloss that made it readable. A wiki without its author is not a backup. It is a fossil.

If DeLong named the forms, Dorothy Leonard and Walter Swap^[2] found the move that gets the deep ones out, and their warning is the single most useful sentence in the literature: if you ask people in an exit interview to transfer their critical knowledge, they will give you the most common experiences of their careers, which is exactly the part the organization already has. The ordinary interview produces the ordinary knowledge. Their fix is smart questioning. Never ask for general lessons. Ask for specific rare events, hard calls, and near misses, the decision moments where the expert diverged from what a competent generalist would have done, because the divergence is the expertise. They also draw the line that humbles every capture effort: explicit knowledge documents handle, implicit the expert can say if asked the right way, and tacit the expert cannot say at all and must be observed. The tacit layer comes out in no interview. It needs their OPPTY sequence, observation and joint problem solving with a near-expert receiver. There is a floor under what listening recovers, and pretending otherwise is how programs oversell.

The naturalistic decision-making school turned that instinct into a protocol. Gary Klein, Beth Crandall, and Robert Hoffman^[3] built the Critical Decision Method on fireground commanders and intensive-care nurses, and the structure is the part worth stealing. You do not interview the job. You interview one real, non-routine incident where the expert's skill was genuinely challenged, sweeping it more than once: a loose telling, a fixed timeline, then a return to each decision point probing what they noticed, what they expected, what options they rejected. The closing move is the what-if query, what a less experienced person standing there would have missed. Evaluations of that work found over 90 percent of the items it elicited were genuinely cognitive, the noticing and the judgment, rather than the procedural steps a manual already holds.

The industry that priced the lesson

The nuclear sector is worth singling out because it is the one place that turned this into machinery and lived with the consequences. The Tennessee Valley Authority's nuclear division built the canonical attrition process in the early 2000s, and the IAEA codified it^[4] across its 2006 and 2017 knowledge-loss publications. For every incumbent, score a retirement factor from the departure date and a position risk factor from how critical and how unique the knowledge is, multiply them into a total attrition factor, and let the system compute it down to the department. That is the matrix, done rigorously, by the people with the most to lose if it was wrong. And here is the sentence in the IAEA write-up that should be printed over every continuity dashboard in every industry: the matrix only finds the people; the interviews are where the actual knowledge inventory happens. They built the best risk-scoring system in the world and then said in writing that the score is not the inventory. The score tells you whose door to knock on. The conversation behind the door is the deliverable.

The same sector supplies the rule that exposes how late most organizations start. The published nuclear practice, carried in INPO's knowledge-transfer guidance^[5], is to hire the successor two to three years ahead of the expected retirement, so the new person can run alongside the expert before the expert leaves. Set that against how a typical capture runs, a few weeks of write-ups at the end, and the contrast is the whole argument. The petroleum literature puts the ramp to an autonomous professional at three to eight years. These are not numbers a better interview compresses. They are the reason the interview has to happen early, while there is still time to act on what it finds.

The scale of what is unwritten

It would be easy to read all of this as a niche problem for nuclear plants and machine shops. The base rates say otherwise. Panopto's workplace-knowledge survey^[6] put a number on the intuition that the floor runs on the undocumented: about 42 percent of the knowledge an employee uses to do the job is unique to them, held by no one else and written nowhere. The matrix that reports a role as covered is reporting on the position, not on the 42 percent. The same work estimates a thousand-person firm loses roughly 2.4 million dollars a year to not finding knowledge that exists somewhere in the building, and IDC put Fortune 500 knowledge-sharing failure^[7] at 31.5 billion dollars a year. And the exit waves that trigger the loss are not niche. The GAO has flagged for years^[8] that a large share of the federal workforce is retirement-eligible within five years, and the 2025 reductions made it concrete: a heavy share of the approved deferred resignations were workers over fifty-five, with agencies naming loss of institutional knowledge as a direct operational risk. The legacy-systems corner is sharpest of all. The COBOL workforce that runs the core of banking and benefits^[9] averages past fifty-five, most mainframe teams run understaffed, and the people who know which 1998 patch must never be touched are the ones near the door. The vocabulary on the exit form changes. The mechanism does not.

Why it stays unasked

Pull the threads together and they converge on one uncomfortable finding. The knowledge that decides whether an operation survives a departure is rarely missing in the sense of being unknowable. It is unasked, because the channels built to ask it are built to miss it. The exit interview, by Leonard's account, retrieves the common and not the rare. The matrix, by the IAEA's own admission, finds the person and not the inventory. The wiki page, by DeLong's taxonomy, captures the layer that rots without its author. And the best witness to whether a capability is covered, the apprentice who quietly hands the hard jobs back, keeps that to himself, because admitting the gap looks like sabotaging his own standing. The truth is distributed across the floor in pieces, each held for a different reason it will not volunteer.

A scene makes the recovery move concrete. Ask a senior operator, late in his tenure, what the new control system must not design out, and he shrugs and says it is fine. Ask him instead about the last time he reached for something on the new screens that was not there, one specific recent shift, and he describes an alarm grouping his anticipation depended on, a thing he would have lost in the changeover and never been able to name in the abstract. The general question got the safe answer. The episodic one got the inventory. That is the difference the canon has described for thirty years, from Klein's incident sweeps to Leonard's rare-case rule to the IAEA's elicitation behind the score.

Which points, finally, at the kind of instrument this calls for. Not another matrix, which finds the people and stops. Not another exit form, which collects the knowledge the organization already has. Something closer to what the literature has prescribed and few have run at scale: a neutral, confidential conversation anchored to a real recent moment rather than a request to generalize, with the right person, early enough that the answer can still change a decision. That is the instrument we are building at Latent Variables. The field told us where the inventory sits and how to ask for it twenty years ago. The open problem was only ever doing it before the date on the calendar arrives.