Almost every company has that one process where everybody points at the same person. The complaint from a particular plant. The line that gets run differently from the two beside it. The account where you need to know who actually decides.
As long as that person is around, it works. They are reachable, they answer quickly, and their answers are good. A single point of knowledge therefore does not register day to day as a problem. It registers as reliability.
Raise the topic and four objections come back. All four are fair, and one of them is strong enough that it does not work in our favour. In order.
"We have that documented"
Usually that is even true. There is a work instruction, a wiki, a folder. The only question is what is in it.
What gets documented is normally the standard case: the order of the steps, the responsibilities, the limits. But a single point of knowledge does not sit in the standard case. It sits in the deviation. How do you tell that this batch will cause trouble before the readings look wrong. When is it worth chasing this supplier, and when are you better off waiting. Which exception was granted once three years ago and has quietly been practice ever since.
42 percent of job-related knowledge is documented nowhere (Panopto/McKinsey). That share is not spread evenly across topics. It collects exactly where experience is required.
There is also the fact that documentation preserves the moment it was written. The work instruction describes the process as it was imagined at the time. The detours that have grown since, because reality did not cooperate, are not in it. Those detours are precisely why the person gets asked.
A useful test for this is unspectacular: take a work instruction and ask the person who does the job what is missing from it. Not what is wrong, what is missing. The answer rarely takes longer than two minutes.
"But she is still here"
This is the objection that comes up most often, and it conflates two different risks.
One is departure. Retirement, resignation, illness. That one gets discussed, and succession planning exists for it. The frame is tightening: in Germany around 1.3 million people retire each year while about 800,000 enter the labour market, leaving a shortfall of roughly 4.3 million workers by 2036 (IW Köln, 2026). Comparable curves apply across much of Western Europe. How to prioritise critical roles for this is covered in the piece on the retirement wave.
The second risk operates every day, even when nobody leaves. A single point of knowledge is a bottleneck, and bottlenecks cost lead time. The decision waits until the person is back from holiday. The colleague does not trust herself to assess the case alone. The follow-up question goes into a calendar that is already full.
This second risk rarely gets addressed because it has no trigger. There is no date on which it becomes acute. It costs a little every month, and nobody receives a notification about it. It shows up most clearly when new colleagues are being brought up to speed, because there all the questions arrive at once. That sequence is covered in the piece on onboarding videos and the first 90 days.
You can observe it without measuring anything. Ask in a meeting which decision is currently waiting for one particular person to have time.
"When she leaves, we will do a handover"
The plan is plausible and fails on the arithmetic of time.
A handover starts when the resignation is on the table or the retirement date is approaching. That leaves four to twelve weeks, usually alongside normal operations and the induction of the successor. In that window you can transfer what can be written down. Thirty years of accumulated judgement cannot be retrieved in that window, because the decisive cases do not come to mind on demand. They come to mind when a concrete case asks for them.
There is a second problem that never appears in the handover protocol: a handover goes to one person. If that person moves on in two years, the same exercise starts again. The single point of knowledge has not been dissolved. It has relocated.
How you know a handover only captured the processes:
- The successor is in post, and after three months she is phoning her predecessor
- There is a protocol, and nobody has opened it since the handover
- For exceptional cases people still write to the old address
- The team describes the successor as "not quite as safe a pair of hands yet" without being able to pin it to a task
"A single point of knowledge is also retention"
Here the objection gets uncomfortable, and it is right.
Being the only one who knows how something works confers standing. That is recognition, often built over years, and in some cases it is the reason the person is still there. If you launch a capture project and give the impression that the point is to make people replaceable, you will get participation that happens formally and yields nothing. Nobody will tell you about the interesting cases.
This objection cannot be argued away. It can only be taken seriously, and that happens in how the project is designed.
What makes the difference, and none of it is a tool:
- The person stays visible as the source instead of disappearing anonymously into a database
- The benefit shows up for her first, through fewer follow-up questions on things long since settled
- It is about the cases she explains three times a year anyway, not her entire working life
- The time required is named up front and small
- Nobody is turned into an author who does not want to write
- The order follows risk, not age
That last point is the one most often violated. Prioritising knowledge capture by year of birth tells the participants, without saying it, what the exercise is really about. Prioritising by critical process says something else, and it also catches the forty-year-old on whom one process depends.
Where it realistically starts
Not with the people, with the processes. Naming people creates resistance. Naming processes creates agreement, because a team usually knows which three or four things only one person can do.
The effort per process then decides whether anything happens at all. A documentation task that costs half a day loses against everyday work. An AI-guided interview needs around 30 minutes from the experienced person and puts its follow-up questions where a statement stays incomplete. Half an hour is something a shift supervisor will give you. Half a day is not. How the procedure works is covered in the piece on AI expert interviews for knowledge capture.
At a German primary insurer, more than 70 key people were captured this way over twelve months and the material turned into documentation, learning videos and queryable chats. The case is anonymised; the order of magnitude is real. What mattered was not the volume but that every knowledge object carries an owner, a validity date and a verification status. Without that you end up with another repository in which nobody can tell what still holds. Why this is the same point for AI systems is covered in the piece on RAG architecture for enterprise.
That leaves one question which cannot be answered from here, and it is the one that matters.
If you had to name your three most critical processes: would your specialists name the same three? In our experience the risk that appears in no plan sits precisely in that difference.
