No plant measures everything. Some variables are technically hard to capture. Others get forgotten, documented by hand, or only become visible once an experienced operator names them. Material behaviour, tool condition, cleaning, ambient conditions, handling, shift practice: not all of it sits cleanly in the system as a parameter.
Does that make AI-based root-cause analysis worthless?
No. But it has to be used with humility.
Incomplete data is normal
In theory, you'd want every relevant cause available as a clean variable. In practice, there are gaps, proxies, measurement errors and time lags.
A good analysis therefore doesn't pretend to be omniscient. It shows which measured factors stand out, which hypotheses become plausible, and where further observation or data collection makes sense.
Data gaps can become visible
Sometimes the most valuable thing an analysis produces isn't the final cause. It's the realisation: we're missing a critical measurement.
When a team keeps getting stuck at the same point, that can justify an investment decision: an additional sensor, better lot linkage, cleaner shift notes, a different inspection frequency, or a new field in the MES.
Process knowledge stays decisive
Unmeasurable or unmeasured factors often show up first at the gemba. An experienced practitioner asks: what actually changed? Who's running the process differently? Which cleaning was unusual? Which material lot felt different? Which maintenance isn't logged in the system?
AI can structure the search space. People have to name the blind spots.
How to handle it cleanly
Good Jidokai communication should say:
- AI supports hypotheses, but doesn't replace validation.
- Data gaps are treated as part of the learning.
- New measurements are prioritised by value, not by interest in the technology.
- Decisions stay anchored in the operating model.
The takeaway
Incomplete data is no reason to do nothing. It's a reason to design root-cause work as a learning system: measure, understand, test, improve, sharpen.