The most common question before an AI-in-operations pilot is: is our data good enough?
The honest answer: maybe. But the better question is: good enough for which use case?
AI-based root-cause analysis doesn't need the perfect digital factory. It needs a clear loss, a meaningful quality or outcome measure, relevant process data, and a team that can validate hypotheses in operation.
1. A target metric that really counts
Without a good target metric there's no good root-cause analysis. Scrap rate, first-pass yield, moisture, torque, surface defects, line stops or complaints can all work. What matters is that the measurement is stable, can be placed in time, and is operationally relevant.
If the quality metric itself varies, is poorly defined, or arrives too late, that's where the work has to start.
2. Process data tied to time or object
Root-cause analysis needs connection. Which parameters belonged to which part, batch, order, material lot or time window? Without that link you get analyses that look mathematical but don't hold up in the plant.
Typical data sources are PLC, MES, SCADA, historian, test systems, lab values, maintenance data, recipes, shift notes and ERP context. Not everything has to be perfectly integrated on day one. But for the bottleneck you've chosen, the chain has to be plausible.
3. Process knowledge, not just data volume
More parameters don't automatically make an analysis better. Good teams know which parameters are physically related, which are only proxies, and which measurement is unreliable in daily use.
So Jidokai should always be told with a practitioner setup: AI finds patterns, people know the process reality, and together they build a better hypothesis.
4. One line, one loss picture, one owner
The best start is rarely a corporate programme. Better is a lighthouse line with a clear pain: recurring scrap, rework, drift, micro-stops or slow complaint analysis.
You also need an owner who can make decisions. Otherwise AI turns into analysis without execution.
The takeaway
Data maturity isn't an abstract score. It shows up in whether a concrete problem can be understood better and solved more stably with the data you already have.