Skip to main content
Essay #005 · · Reading time 5 min · Field: IT/Digitalization · Market: Mittelstand
Inhalt
  1. Tool or architecture: two ways to decide about IT
  2. The four layers of agentic IT
  3. Why these decisions belong to you
  4. Three paths that fail
  5. What works instead
  6. The next step

The IT question after the agent turn

Essay #005 · 2026 · Reading time 5 min · Field: IT / Digitalization · Market: Mittelstand

Thirty years ago, the IT question in a Mittelstand company was clear: which system do we buy, how do we roll it out, who operates it? The answer was a piece of software: an ERP, an accounting package, a CRM. IT was a tool. The business processes stayed as they were; the software just made them faster.

With agentic IT, meaning systems that carry out tasks on their own instead of just processing input, that changes. These systems change how your company works: which processes exist, which roles remain, who decides what. That is what this essay is about. And about why you cannot hand these decisions to your IT department.

Tool or architecture: two ways to decide about IT

The tool question is: what problem do we have, and which tool solves it? The answer is a purchase or an in-house build. Responsibility sits with the IT department or a service provider. The managing director approves the budget and otherwise stays out. For classical IT, that was fine.

The architecture question is: what changes in our company when we build this technology in for good? The answer is not a purchase but a structural decision. It determines which ways of working are possible, which roles remain meaningful, where data flows and which decisions will be made automatically from now on.

Agentic IT belongs in the second category. If you buy it like a tool, you make structural decisions without noticing.

The four layers of agentic IT

Solid agentic IT has four layers. You do not need to master any of them technically. But you need to know they exist, and who in your company owns each one.

First layer: the data foundation. What data does your company have, and how do agents get to it? That sounds simple. In the Mittelstand companies I know, it is not: the data sits scattered across the ERP, the CRM, email inboxes and Excel files. Before agents can work with it, someone has to create order. That work deserves care, because everything else builds on it.

Second layer: model access. Which AI models do you use, on what terms, with what commitments from the vendor? This layer changes fastest. A long-term commitment to one specific model therefore does not hold. Keep this layer swappable: switching models must not disrupt your operations.

Third layer: orchestration. How are multiple agents and models connected into end-to-end workflows? A single agent achieves little. Several agents that work together properly can take over a substantial part of office work. This layer is the most demanding, and the one most likely to require your own development work.

Fourth layer: governance, meaning the rules of use. Who may use which agents for what? What may an agent decide on its own, and what needs human sign-off? How can you trace afterwards what happened? In my observation, this layer is the one most often forgotten, and its absence becomes the most expensive later. Data protection, liability and compliance all hang on it.

Why these decisions belong to you

Each of the four layers contains decisions that shape your company for the long term.

The data foundation determines which knowledge becomes usable for agents and which stays in silos. Who decides which silos get dissolved? The IT lead? The department heads? Or you, because it is a strategic question?

Model access determines which vendors your company becomes dependent on. Supplier dependency is a classic strategy question. Who decides it?

Orchestration determines which processes get automated and thereby locked in. That is the question of how agile your company will still be in five years.

Governance determines how your company reacts when something goes wrong. Such rules rarely get written before the first incident. If you do not set them, you leave them to chance.

If all of this runs as an IT decision, your IT department holds power it should not have. That is not ill will: it is simply the wrong body to decide strategy questions with long-term effects.

Three paths that fail

“Our service provider handles that.” The digital agency you have worked with for ten years is asked to build the agentic infrastructure. That is convenient, because the trust is there. But the provider has its own interests, and they do not fully match yours. It is likely to build an architecture that fits its own business and not necessarily the one that serves you best. Help with building individual layers is fine. It gets critical when the same partner makes the architecture decision it will then implement itself.

“We buy a suite.” A large vendor covers all four layers with one complete solution. That reduces the decision to a single signature. Tempting — but it is the deepest form of vendor lock-in possible here. If you buy all four layers from one vendor, you give up your architectural independence.

“We build everything ourselves.” Maximum control, a big internal push. That can work in exceptional cases. In my assessment, though, in most companies it ties up more people and money than the core business justifies. That capacity is then missing elsewhere.

What works instead

What holds up is a mix of clearly defined in-house parts and bought-in parts. You decide which layers your company must control itself and which it can buy. As a rule of thumb: the data foundation belongs in-house. Model access stays swappable. Orchestration is partly bought, partly your own. Governance is entirely your job.

You cannot read this split off an IT concept paper. You need to understand what each layer does and where the strategic levers sit. That does not take a computer science degree, but it does take your time and your questions.

The next step

The IT question after the agent turn is no longer an IT question. It is a leadership question. If you delegate it, you are delegating strategy, not technology.

My expectation: the companies in good shape five years from now will be those whose leadership has understood these four layers. They will not know every technical detail, but they will know the strategic consequences.

In concrete terms: put the four layers on the agenda of your next strategy meeting. For each layer, settle two questions: who owns it today? And is that the right body? Wherever the answer is “our service provider” or “nobody”, you have found your first construction site.

— Axel Roth