Anomaly detection sounds good. A system notices that the process is behaving unusually and flags it earlier than a team would catch it by hand.
But an early signal is only the start. Value only appears once it's clear what happens next.
The new wall of alarms
Plenty of digital systems generate hints. Too many hints. Operators and shift leads quickly learn which alarms they can ignore. If AI just produces more notifications, it doesn't improve operations. It adds noise.
A good signal therefore has to be actionable: what's unusual? Why might it matter? What should be checked now? Who decides? When do we escalate?
Standard work starts at the next step
An AI hint should be embedded in a clear response logic. For example:
- Parameter drift detected
- Shift lead checks three defined points
- Operator confirms the condition at the process
- Engineering validates the hypothesis on a repeat
- The standard is adjusted once the pattern is confirmed
That logic has to be simple enough to work in shift operation.
Learning, not one-offs
When the same anomaly comes back, the team shouldn't start from scratch every time. That calls for a rule: a monitor, an SOP, a checklist, a training point or an A3.
AI can help spot the repetition. Lean makes sure it turns into standard work.
Where the senior practitioner comes in
The practitioner asks the uncomfortable questions:
- Is the signal actually important?
- Who can respond safely?
- Which countermeasure is reversible?
- Which change needs sign-off?
- How do we avoid local one-off logic?
Without those questions, Jidokai stays a technical system. With them, it becomes part of the operating system.
The takeaway
AI makes anomalies visible earlier. Lean turns them into better work. The bridge between the two is standard work.