Agentic AI for manufacturers: operating model first, then agents.

An AI agent carries out multi-step tasks itself: it plans steps, calls systems and acts, where a chatbot only answers and an automation repeats a fixed sequence. That is why it needs what a chatbot does not: an accountable person, a level of autonomy per task, and human sign-off wherever an action can no longer be undone.

You find out by comparing it with the process you actually run today, not with a perfect one. In the audit we measure the current process, agree beforehand which number counts as evidence and when to stop, and set build, buy and platform against each other. If it does not pay, the board paper says so, with the reasons.

Three questions decide it: how close the process is to your core, which data and systems the agent has to reach, and what you will still be able to change in three years. I set build, buy and platform side by side, with cost and dependencies.

With a specification written before the selection, criteria fixed in advance, and a sign-off by someone who earns nothing from the contract. That is the role I take.

Yes, for every agent that acts rather than only assists. Who owns it, what counts as success and where a human decides have to be settled before it takes over a process.

By settling how an agent will evolve before the first one goes live. Models get replaced, vendors change their interfaces, rules get added. So an agent needs what a connected product needs: versions, a re-test against fixed cases after every change, a rollback path and a planned end of support. It also needs access to data and systems that does not depend on one vendor.

It depends on your role and the purpose: whoever places an agent on the market or puts it into service under their own name, including for their own use, has different obligations as a provider from a company that only uses it; whoever builds and deploys it in-house is both. An agent in a high-risk area has different obligations from one in purchasing, and since 2 August 2026 every agent that communicates with people has to disclose that it is an AI (Article 50). Connected products add the CRA, machinery adds the Machinery Regulation. In the audit we prepare these questions technically so that your legal team can decide them before the architecture is fixed.

The audit has a fixed price, set after a 60-minute scoping call. What moves it: the number of processes and systems in scope, the number of interviews, a briefing for the board or supervisory board, and the days on site.

No. For ten years I was responsible for delivery and go-to-market, at times with some 40 people, and on FOTA at OEM scale I was the Business Owner responsible for the customer. I could do it. That is precisely why I do not bid: whoever picks the partner and signs off the work must not earn from the contract.

Because the pre-study is not free then; it is paid for out of the implementation sold afterwards. Whoever earns from the implementation does not write the specification against its own interest. I do not sell implementation; I sign it off against the specification. The price of the pre-study is fixed beforehand — not afterwards, in the margin.

From recommendations, never, in any form. The complete answer needs the other half of my work, though: I advise industrial-tech scale-ups on go-to-market, sometimes with a success-based component or advisor equity. So the rule has no exceptions: no company that pays me that way, or that I hold a stake in, enters a selection process I am running. Where that overlaps, you hear it from me before it matters, and I step out of the selection.

None: no non-compete, no ongoing contracts, no commission. What I learned at Bosch stays confidential.

Directly with me. No associates, no junior layer. Where specialised depth is needed I draw on a small trusted network of senior freelancers — by exception, with a clear rationale.