The exception queue is the real interface for industrial AI
An industrial AI system can reduce a shift's alerts to three cases and still leave the operator with hours of investigative work. The product has to carry evidence, uncertainty and ownership into the decision.
These are companion notes. The full essay lives on The Turing Pilgrim, my Substack publication.
Why this matters
A compressor inspection recommendation is useful only when the person receiving it can understand the work behind it. A change in vibration and a recent sensor repair point toward different explanations. The case should make both visible, alongside the checks already performed and the reason the agent stopped.
That handoff has a capacity cost. Twenty cases taking fifteen minutes each require five hours of review before anyone carries out a physical inspection. Product teams need to see the time spent understanding cases, the information reviewers still have to find and the work waiting for a specialist. A smaller queue alone cannot tell them whether the system is helping.
The same record needs to survive the next shift. Evidence, comments, ownership and reasons for deferral should travel with the case. Acknowledging an alert is different from establishing that the underlying condition has changed, so closure needs a reason and supporting evidence. The article develops this as a product design proposal using hypothetical examples, without claiming reliable automated diagnosis.
Key takeaways
- Carry the observations, their timestamps, the agent's completed checks and its unresolved questions with each recommendation.
- Measure review effort and waiting time. Keep the original age and history when a case is returned for more information.
- Make the next responsible person and outstanding decision explicit across operator, planner and technician handoffs.
- Review a sample of suppressed or automatically resolved cases as well as escalations. Queue size alone can hide missed uncertainty.
- Start with one equipment class and a bounded decision. Evaluate corrections before turning them into broader operating policy.
Who should read it
Product and operations leaders deciding how industrial AI should hand work back to a person, especially where maintenance, control-room and specialist teams share responsibility.
Related field notes
The next AI shift is delegation
Define the boundaries around what software may do on our behalf.
Why industrial AI needs its own org chart
Consider who owns the operating model as AI becomes part of industrial work.
When AI learns physics
Explore how physical context changes the questions an industrial model can answer.
Working through this decision?
A short note on the AI bet you are weighing is enough to start. I usually reply within one business day.