One production problem. One measurable pilot.

Start with the production problem.

Describe the defect, delay or equipment issue in operational language. Perhaps labels are missed, packages need counting, stoppage reasons are recorded inconsistently, or maintenance tickets arrive without useful context. Name the location, the person responsible and the decision they need to make.

Choose one process and one measurable outcome for the first version. A vision pilot might measure missed defects and unnecessary flags separately. A maintenance workflow might measure how many alerts arrive with enough information for the technician to act.

Show the conditions, not just the perfect example.

For machine vision and image recognition, describe product variation, lighting, camera position, movement and occlusion. A useful sample set includes acceptable products, known defects and borderline cases across relevant operating conditions. Agree on which examples will be held back for evaluation.

For a conveyor request, include the product range, approximate throughput and where observation is possible. The first question is whether the problem can be reliably observed. Selecting a model comes after understanding what the camera or sensor can actually capture.

Distinguish maintenance alerts from predictions.

A scheduled reminder, a threshold alert and a prediction of failure are different projects. If the immediate problem is missed inspections or incomplete work orders, a better maintenance workflow may be the useful first step.

Predictive maintenance requires evidence about equipment condition and outcomes. Ask what sensor history, repair records and failure labels exist before promising predictions. NIST’s work on prognostics and health management explicitly addresses assessing equipment health and evaluating methods for manufacturing. See NIST’s manufacturing health management program.

Keep the pilot observable and reviewable.

Start by recording observations and proposed alerts alongside the existing process. Have operators review disagreements and track nuisance alerts as well as missed events. Agree on what evidence would justify further investment and what would cause the pilot to stop.

Any change that controls machinery needs the site’s responsible engineering and safety review before deployment. A monitoring pilot should not be treated as an approved machine safeguard. Make the scope clear so the specialist knows whether the task is reporting, inspection assistance or a control-system change.

Include installation constraints, connectivity, on-site support and ongoing maintenance in the brief. A useful deliverable is a measured pilot with findings and an operating plan, followed by an explicit decision about the next stage.