Your Most Valuable Context Retires by 2039

Most AI projects index the documents. The knowledge that decides whether the output is right leaves with the people who never wrote it down.

The number

By 2039, about 13.4 million people in Germany’s labour force will be past the statutory retirement age of 67. According to Destatis, that is 31% of everyone who was available to the labour market in 2024.

The 55 to 64 year olds alone counted 10.0 million people in 2024, more than any younger ten-year age group. The younger cohorts will not replace them in numbers.

Most discussions treat this as a hiring problem. For AI projects, it is a context problem.

What is not in any document

The philosopher Michael Polanyi summed it up in 1966: we can know more than we can tell.

In a company, that knowledge looks like this:

why the line, the process or the system was built this way and not the obvious way
which supplier specification is formally wrong but works in practice
which workaround keeps the month-end close alive
which failure already happened, what it cost, and how it was fixed

Some of it sits in unstructured data: emails, tickets, meeting notes, handwritten logbooks. Most of it was never written down at all. It lives in the people who have been there longest.

Why AI projects miss it

Most enterprise AI projects start by indexing documents. That gives you retrieval over what was written. It does not give you what was known.

MIT’s 2025 study of enterprise AI points at the same gap. Its central explanation for failed projects is a learning gap: systems that lose context and do not get better with feedback.

A retrieval pipeline over your SharePoint is an archive. It is not an expert.

Put experience into the harness

In the harness article I argued that the value of AI sits in the system around the model. Experienced colleagues belong inside that system, in four roles:

Authors of success criteria. They define what a correct result looks like, including the cases nobody documented.
Authors of test cases. Every incident they remember becomes an evaluation case the agent has to pass.
Owners of the error budget. They decide how wrong an agent may be in their domain, and where it may not be wrong at all.
The human gate. They review agent output before it changes anything that matters. Every correction becomes a permanent change to the harness.

This is different from a farewell interview in someone’s last month. An interview captures stories. A test case captures a rule that is checked every single time.

People-first means asking, not extracting

Nobody shares thirty years of judgment because a project plan says “knowledge transfer”. People share when the work is respected and the time is real.

Protect the time. Evaluation work is work, not a side task on top of a full calendar.
Make the contribution visible. Name the people whose test cases protect a process.
Involve the works council from day one, especially when conversations are recorded or transcribed.
Talk about experience, not age. The argument holds just as well for the 35-year-old who built the original system.

How to measure it

If you cannot measure it, it is a workshop, not a transformation. Four numbers are enough to start:

Metric	What it shows
Share of test cases written by domain experts	Whether experience is actually inside the harness
Rework rate in their domain	Whether agent output improves once their rules are encoded
Time to competence for successors	Whether knowledge reaches people, not only machines
Incidents whose cause someone already knew	How often the answer existed before it failed in production

The order matters

If you transform people-first, people solve problems. If you transform process-first, people satisfy processes.

The colleagues who leave by 2039 are not a risk to manage. They are the best source of ground truth your AI projects will ever have, and the window to use it is closing.