When a plant-floor AI pilot deserves production budget
A budget memo for deciding when an industrial AI pilot has crossed from controlled experiment to production workflow.
Most plant-floor AI pilots do not fail in the model. They fail when nobody owns the exception after the model is wrong.
Audience
Operations leaders evaluating AI pilots in manufacturing, industrial service, and asset-heavy environments.
Surface
Plant-floor exception handling.
Decision
Decide whether the pilot is ready for production budget or should stay in controlled evaluation.
The expensive miss
The demo measures prediction quality. The operation feels exception cost. The budget owner inherits the gap between the two.
A maintenance copilot can summarize technician notes. A vision model can flag a quality defect. A scheduling assistant can suggest a better dispatch sequence.
None of that matters if the workflow breaks when the output is uncertain.
- A supervisor still checks the same queue twice.
- A technician still writes the real context in a side channel.
- A planner still trusts the senior person's memory over the system.
- A vendor dashboard shows activity, but not avoided downtime.
Where AI belongs
AI belongs where the workflow already has repeatable judgment and messy data exhaust.
The useful surface is not autonomy. It is a narrower exception loop.
- Classifying maintenance notes before backlog review.
- Extracting warranty patterns from service records.
- Flagging dispatch conflicts before the day starts.
- Summarizing shift handoffs into structured exceptions.
Where AI should stay out
AI should stay out of final authority when the cost of a miss is safety, compliance, or customer downtime.
If the system cannot show its source, owner, and next action, it is not ready for budget expansion.
- Approving safety-critical work without accountable review.
- Replacing technician signoff.
- Changing dispatch priority without operational owner approval.
- Creating procurement evidence without source traceability.
Decision rule
Fund the workflow only when these are true.
- The output changes one named workflow.
- A human owner is assigned to every exception.
- The team can measure avoided delay, reduced rework, or cleaner escalation.