Separate past and future
Use a later holdout period and prevent closely related records from the same lot leaking between sets.
MANUFACTURING / PROVE
Manufacturing
AI discovery and PoC.
Select one line or asset group and one goal: quality investigation or anomaly notification. Can the team assemble useful investigation evidence sooner than with the current workflow?
Explore the delivery scope ↓OPERATING MODELILLUSTRATIVE
- 01Production & inspection
- 02Align lots and timestamps
- 03Compare process conditions
- 04Investigate on the floor
Conceptual data-to-operation flow, adapted to the agreed project scope.
THE DECISION TO MAKE
Can the team assemble useful investigation evidence sooner than with the current workflow?
Select one line or asset group and one goal: quality investigation or anomaly notification.
A PoC should produce evidence for the next investment decision. Business owners, data owners and users agree on normal tasks and exceptions, rather than treating a working screen as proof of value.
INDUSTRY-SPECIFIC TESTS
Include the conditions that matter.
Check equipment differences
Review results by operating condition, product change and maintenance state.
Reproduce on the floor
Ask operators to inspect source records and explain the alert and next action.
BASELINE & ACCEPTANCE
Choose measures for the selected topic.
Production and quality data
| Candidate measure | Method |
|---|---|
| Traceability | Traceable lots divided by in-scope lots |
| Reporting effort | Active time from source collection to reconciled report |
| Defect rate | Agreed defect count divided by inspected quantity, with a fixed retest rule |
Maintenance and anomaly AI
| Candidate measure | Method |
|---|---|
| Alert workload | Notifications requiring review per asset per day |
| Actionable lead time | Time available before intervention is needed |
| Useful detections | In-scope events detected within the agreed window |
Set targets from your own baseline. Hold evaluation evidence separate from tuning and fix the assessment period and scope. Record failures, abstentions and untested conditions alongside successes.
GO / REVISE / STOP
A decision to stop is useful evidence.
If traceability is poor, alert review increases effort, or events are insufficient to evaluate, prioritize data capture or rules-based monitoring.
Proceed to build
Core criteria pass, with owners, interfaces and remaining issues understood.
Validate further
There is a promising signal but insufficient coverage; define a bounded follow-up.
Choose another approach
Value does not exceed the baseline; compare data cleanup or improvements to existing tools.
A defined engagement and usable deliverables.
The PoC scope defines the workflow, source material, environment, evaluation and prototype/report deliverables. Production development is a separate decision.
- Workflow and data definitions with assumptions
- Reproducible prototype, evaluation and exceptions
- Evidence for implementation, further testing or stopping
QUESTIONS BEFORE YOU START
Questions before starting
Can we start without sensor data?
Yes. Start with production and inspection reconciliation. Predictive maintenance needs a separate assessment of sampling and event history.
Do you guarantee a defect reduction?
No fixed reduction is promised. First validate traceability and investigation effort, then evaluate quality interventions through plant trials.
RELATED FIELD GUIDE
MANUFACTURING / GUIDE
Evaluating manufacturing AI: lots, time and false alarms before accuracy
Before commissioning factory AI, define how evaluation data is separated and how operators will review results. This guide explains common pitfalls and a practical evaluation brief.
Read the guide ↗YOUR OPERATION / OUR STARTING POINT
Start with the work you do today.
Tell us about the workflow, systems and reporting or documents that need attention. We will define a suitable phase, deliverables and assumptions.
Manufacturing — discuss your project How scope and estimates work ↗