In the last few weeks, two discussions have kept running in parallel with the clients. One is about business continuity and demographic change: who replaces the plant manager who’s been running that line since the 90s, the process engineer who just knows which valve sticks in cold weather, the account manager who remembers why that one customer gets special terms. The other is about agentic business capability: pilots, use cases, first production deployments.
Nobody seems to notice these two conversations are about the same problem.
Germany’s Federal Statistical Office expects around 13 million people currently in the workforce to pass retirement age by 2040 — roughly 30 percent of today’s labor force. Bavaria is named among the states hit hardest.
What’s leaving isn’t headcount. It’s the informal decision network.
The org chart never captured most of what makes an experienced team fast. It’s the shortcuts, the exceptions nobody wrote down, the reason a request goes to a specific person and not the ticket queue, because that person remembers the incident three years ago that explains why.
Consultants who study succession in large organizations have a name for this: the informal decision-making network — the part of the company that actually gets things done, running underneath the official structure. When the people retire, that network doesn’t get inherited. It just goes quiet.
Most companies still treat this as an HR problem: hire faster, document more, run a few knowledge-transfer workshops before the farewell party. That helps at the margin. It doesn’t solve it, because most of what’s leaving was never explicit enough to write down in the first place.
At the same time, something genuinely new is arriving: agentic business capability.
Call it what it is: AI systems that don’t just answer a question or draft a document, but carry out a multi-step piece of business work on their own — process a claim, reconcile an order discrepancy, adjust a production schedule, escalate the right exception to the right person.
The step from „AI that advises“ to „AI that acts“ is the real shift happening in enterprise IT right now, and it’s why boardrooms that were skeptical eighteen months ago are running budget lines for it today.
But an agent that can act is not the same as an agent that understands the business.
Here’s where the two conversations collide. Most agentic pilots I see are narrow by design — one agent, one process, one system. That’s the right way to start. But it also means each of these agents is building its own private, local idea of what „urgent,“ „the customer,“ or „critical defect“ means, based only on whatever it can see in its own corner of the stack. Two agents in two departments can be working the same case and disagree on the basics, because nobody ever wrote down, in a form a machine can use, what these terms mean company-wide — the same thing that used to live only in one experienced person’s head.
Scale that pilot into ten, twenty, fifty agents across a company, and you don’t get an intelligent organization. You get fifty confident, isolated opinions.
This is exactly the gap a semantic layer is built to close.
A semantic layer is a shared definition of what a business’s terms, entities, and rules actually mean — independent of which system happens to store the data. Not the data itself, but the meaning behind it: that „customer“ means the same thing in sales, service, and finance; that a „critical defect“ triggers the same escalation everywhere; that the exception one veteran employee always made is written down as a rule, not just a habit.
Built properly, it’s the closest thing a company has to capturing what its retiring experts knew — before they’re gone, and in a form the next generation of agents, not just the next generation of employees, can actually use.
My view
Most of the investment right now is going into the agents themselves — the visible, demoable part. Very little is going into the layer underneath that would let those agents actually understand the business they’re working in. That’s the wrong order, and I don’t think most organizations have fully registered it yet.
The uncomfortable part is that a semantic layer can’t be bought as a product and switched on. It has to be built out of the same institutional knowledge that’s currently walking out the door — which means the window to capture it is closing at roughly the pace people are retiring.
Companies that start now will end up with agents that actually know their business. Companies that wait will end up with very capable agents that know nothing at all.
ich bin eingeladen worden auf der Embedded meets Agile am 26.04 in München über das Thema „Design Thinking vs Business Canvas Modell“ zu sprechen. Ziel ist es hier aufzuzeigen, wo die Unterschiede liegen, worauf gerade in Umfeld Embedded geachtet werden muss und wo in einer Kombination aus Fuzzy und strukturiertem Ansatz Mehrwerte liegen können.
Consultant's Digest 