Many companies have successful AI pilots and few AI applications that matter. The step from one working station to a capability across lines and sites is where most of the value is, and where most efforts stall.
A model is a piece of equipment
After go-live, a model needs what any equipment needs: monitoring, care, spare parts and a plan for changes. Monthly monitoring with clear thresholds detects drift before customers do. When a new material supplier changed the gloss of housings, detection dropped below threshold within a month. Retraining with 300 new images restored it.
Budget this care. Model operation typically costs 10 to 15 percent of the initial investment per year, mostly in internal time.
Package what worked
The second line should not be a new project. Write down the scaling package after the pilot:
- hardware list and mounting drawing,
- lighting rules and the defect catalog,
- model card and retraining procedure,
- training for operators and team leaders,
- a business case template with the values measured on the first line,
- a checklist for the parallel run.
With the package, the next line in the example cost EUR 85,000 instead of EUR 120,000 and took twelve weeks instead of a quarter of experimentation.
Measure the benefit, not the accuracy
A use case is a success when the operational KPI moves. Track it monthly against the business case:
| Measure | Baseline | Target | Month 6 |
|---|---|---|---|
| Escapes to the customer (ppm) | 420 | 150 | 160 |
| Inspectors at the end of the line | 6 | 4 | 4 |
| Net benefit per year (EUR) | 138,000 | run rate 135,000 |
The run rate is the average of the last three months times twelve. It tells the sponsor early whether the case holds.
A twelve-month roadmap
- Months 1 to 3: losses, portfolio, data readiness, first pilot with stop criteria, lean basics on the candidate processes.
- Months 4 to 6: pilot decided, governance with model cards, data collection for the next use cases.
- Months 7 to 9: second and third pilot, monitoring routine, skills for practitioners.
- Months 10 to 12: scaling package rolled out to further lines, benefits reported with the plant KPIs, maturity check repeated.
Build skills inside
Partners help to start. Over time, each site needs one or two people who can operate models, run monitoring and retrain, and a wider group of process experts who can prepare data and labels. Without them, every change becomes a purchase order and every drift waits for an external appointment.
The pattern behind success
Plants that scale AI well share a pattern: they start from losses, they stabilize processes first, they decide pilots on a date, they involve people early, and they treat models like equipment. None of this is new. It is lean, applied to a new tool.