The two weeks' notice
August 22, 2026
Nothing reveals the shape of an organization's dependencies like a resignation letter. All at once everyone can see which knowledge was never written down, and exactly whose head it was in.
The handover that could not happen
The notice arrives on a Tuesday and the panic is oddly specific. Not about the role — a replacement can be hired. About the eleven things only this person knows: why the March reconciliation has that exception, which client tolerates a late invoice and which absolutely does not, what the third step of the renewal process is actually for, and the reason a rule everyone follows was written in the first place.
So a handover is scheduled. Two weeks, some documents produced, a lot of goodwill on both sides. It does not work, and it was never going to. The knowledge in question is not a list of facts. It is a set of judgements, each one shaped by an incident nobody recorded. You cannot transfer in ten days what accumulated over four years of being the person who was there.
This is the elephant we call Guru: tribal knowledge and hero culture, where one person is the bottleneck. It is the only elephant that says thank you on its way out the door.
The test
You will not find this number. You can measure your own in an afternoon.
There is no trustworthy public figure for how much of a role’s knowledge sits with exactly one person. We went looking for one and refused everything we found: the figures in circulation trace back to vendor surveys with undisclosed methods, and the strongest study we could verify at its primary source measures software repositories, not roles. The absence is worth stating plainly rather than filling with someone else’s marketing.
It also matters less than it appears, because the figure that governs your risk is not the industry’s. Take one role and list the decisions it made last quarter. For each, ask who else could have made it the same way, with the same context, without asking. The decisions where the answer is nobody are your exposure — named, countable, and yours.
We cited a number here. We took it out.
This section used to carry a 2018 figure: the share of role-critical knowledge held by exactly one person. It came from a vendor survey. When we went back to bind it to a verbatim quote at its primary source, we could not. The method was never disclosed, and the number had been repeated far more often than it had been measured.
So we withdrew it. Not softened, not hedged behind studies suggest. Removed, with the reason written down where you can read it. The source stays in the list below, and its note records exactly what happened and when.
From a number that supports the argument. To an argument that survives without one. A figure that withstands scrutiny earns its place. A figure that does not has to go, even when it was doing useful work, and even when nobody would ever have checked.
Nobody is hoarding. So how does a Guru form?
By being useful, repeatedly, in an organization that rewards rescue more than it rewards redundancy. Consider what actually happens to the two behaviours. The person who unblocks the crisis at seven in the evening is thanked by name in a meeting. The person who spends a slow afternoon writing down how the process works so that nobody has to be unblocked is doing invisible work, and produces a document that will be read by someone who never learns whom to thank.
Worse, that second person is reducing their own indispensability, and everyone knows it. Nobody says this out loud and nobody has to. The incentive is legible without being stated. Under those conditions, knowledge concentrating in a few heads is not a failure of the culture. It is the culture working exactly as configured.
The organizational cost follows from the same shape. Everything important queues behind the same handful of people, so throughput is capped no matter how many others you hire. And the process cannot be improved, because improving it requires understanding it, and understanding lives in one head that is fully occupied being the bottleneck.
What does it cost the Guru?
This is the part most treatments of the subject skip, and it changes what you do about it. The Guru is usually not the beneficiary of this arrangement. They are its most constrained participant.
They cannot take a holiday without the queue forming behind them. They frequently cannot be promoted, because nobody can backfill what they do. The reward for becoming indispensable is being made permanent in the role you have outgrown. They carry a low-grade obligation to be reachable. And they are the person who gets interrupted all day, which is why they never have the uninterrupted afternoon that writing it down would require.
Which means the intervention is not exhortation. Telling a Guru to document more is asking the busiest constrained person in the building to fix a structural problem in their spare time, and it will fail for the same reason it has already failed. The transfer has to become somebody's actual job, with actual time, or it does not happen.
How do you know the Guru is in your room?
The tell is never a complaint. It is a workaround everyone treats as normal.
- A name appears in planning conversations where a role should: not “finance approves”, but “we need Miriam for that”.
- Holidays are scheduled around the business rather than the person, and everyone considers this reasonable.
- A process was last improved before the current owner took it over.
- New hires reach competence quickly on the documented parts and stall for months on the rest.
- Someone senior can describe what the person does but not how they decide.
Won't an AI assistant just absorb what they know?
It is the obvious hope, and it is backwards in an instructive way. An assistant is extremely good at making written knowledge findable, summarised and reusable. The Guru's knowledge is, by definition, the part that was never written. There is nothing in the corpus to absorb.
So pointing a model at your documentation will faithfully amplify whatever was already written down, and leave everything else exactly where it was. It will also produce a convincing impression of coverage, which is worse than no impression at all. The one genuinely useful role for the technology here is upstream: lowering the cost of capture, so that a decision can be explained once, in speech, at the moment it is made, and turned into something durable without an uninterrupted afternoon.
That is the honest version of AI and the Guru. It does not extract judgement. It can make recording judgement cheap enough that a busy person will actually do it, which is a real contribution and a much smaller claim.
How do you get the Guru out of the room?
The same three moves we bring to any elephant, pointed at this one.
Map
Find the single points of dependency before one of them resigns. Ask, for each critical process, who could run it on Monday if the usual person were unreachable, and treat a hesitation as the answer. The result is a short ranked list of exposures. That is a different and more useful artifact than a list of undocumented processes.
Prove
Take the highest exposure and transfer it for real. A second person runs it, with the Guru available but not intervening, until they hit the judgement calls. Those calls are the actual content, so capture them as they surface. One process, one cycle. The test is not a document produced. It is a holiday taken.
Scale
Change what gets recognised, or it grows back. As long as rescue is celebrated and redundancy is invisible, the next Guru is already forming. Fund the capture as work, name the people who make themselves replaceable, and let the second person be the one who answers next time.
Find out who your organization cannot afford to lose
The Elephant Safari names your herd in ten questions and ranks them by what they cost you. Or start from the business problem instead: having grown, and watching the processes break. Get the Guru out of the room. Keep what they know.
Every figure in this article traces to a primary source. See it in Knowledge