A plant manager returns from a trade fair with ten AI ideas and a stack of brochures. Six months later two pilots are running, nobody can say what they are worth, and the losses that hurt most have not been touched.
The problem is not the technology. It is the order of the questions.
Start from the loss
Before any tool, list the five largest operational losses of the plant in numbers: scrap, rework, complaints, unplanned downtime, excess stock, time spent searching. In a plant of 420 people molding and assembling plastic housings, such a list might read: EUR 180,000 a year in escapes and complaints, EUR 180,000 in molding scrap, 310 hours of unplanned downtime, EUR 95,000 in stock above target.
Then ask for each loss whether faster detection, better prediction or faster access to knowledge would reduce it. Only then ask which kind of AI fits.
Five families
Nearly every factory use case belongs to one of five families:
- Vision inspection: cameras check parts, labels and assembly completeness. Needs labeled images and stable lighting. Benefit: fewer escapes, less manual checking.
- Predictive maintenance: sensor patterns warn before a failure. Needs vibration, temperature or pressure history together with failure records. Benefit: repairs planned into breaks.
- Forecasting: demand, scrap, energy or lead time predicted from history. Needs two to three years of clean data and known drivers. Benefit: lower stock, fewer rush orders.
- Assistants: language models answer questions from manuals, standards and repair notes. Needs current documents with an owner. Benefit: faster troubleshooting and onboarding.
- Planning and optimization: algorithms suggest sequences and schedules. Needs reliable master data and written rules. Benefit: shorter changeovers, fewer delays.
Knowing the family tells you early what data you need and what kind of benefit to expect.
Four signals of a strong use case
Repetition: the same check or decision happens hundreds of times a day. A hidden pattern: experts suspect a relation but cannot see it by hand. A clear reaction: a result leads to a defined action such as reject, repair or reorder. A baseline: the loss is counted today.
A use case that misses two of these signals is usually an experiment, not an improvement.
Where AI is the wrong tool
If the cause is known and a rule fixes it, write a standard. If a defect becomes visible with good light, install the light. If a wrong answer can hurt a person, use engineered safety functions, not a learning model.
AI earns its place where people cannot keep up with the volume or the complexity. Everywhere else, simpler tools are cheaper, more robust and easier to maintain.
The plant with ten ideas does better with one well chosen use case, a business owner who wants the result, and a loss that everybody can see shrinking.