The Two Wrong Camps
On AI, mid-sized manufacturing splits into two camps: one believes the vendors' revolution promises, launches a mega-project, and fails against the reality of its data landscape. The other has heard exactly those stories and never touches the topic — while competitors quietly accumulate experience.
Both camps make the same mistake: they treat AI as an all-or-nothing decision. The realistic middle consists of small, sharply scoped projects with a named result.
Four First Projects That Have Proven Themselves
1. Analyze the production data you already have. Your controllers, MES, and historian have been producing data for years that nobody reads. A retroactive analysis answers concrete questions — where does the scrap come from? What preceded the last failure? — without a single day of new data collection. Effort: ~3 weeks, fixed price.
2. Early downtime detection on one critical machine. Not the whole hall — one machine whose failure truly hurts. Possible from week one without historical data; the full module takes 7 weeks.
3. Make maintenance documentation searchable. Twenty years of maintenance logs, inspection reports, and vendor PDFs in which nobody finds anything anymore — locally searchable, without a single document leaving the building. Unglamorous, but one of the fastest tangible effects in daily operations.
4. Analyze energy and production data together. Savings potential rarely sits in the energy meter alone — it appears when consumption and production context are examined together.
Three First Projects That Regularly Fail
The all-at-once rollout. "We are digitalizing the entire production" fails not on technology but on organizational capacity: too many fronts, no single measurable result — after a year the budget is gone and the proof is missing.
The chatbot without a data foundation. An assistant that can access nothing operational answers questions Google answers too. Data foundation first, interface second.
Full autonomy. Systems meant to intervene in production on their own are not a first project — the trust simply isn't there at the start, and rightly so. The proven path: let the system detect and explain first, keep humans intervening. Autonomy is an expansion stage, not the ticket in.
Four Rules for the Start
- One problem, one machine, one result. Not a program — a project.
- Fixed price and fixed duration — open ends kill trust in first projects, on cost too.
- The result must belong to you. Models and data stay with you — otherwise you are buying dependency, not capability.
- Data stays in the building. For compliance reasons — and because your process knowledge is your capital.
If you don't know which of the four first projects is right for you: that is exactly the question the AI Readiness Assessment answers — €9,500 flat, 1–1.5 weeks, open-ended. Or simply tell us what costs you money.