Visual AI SOP Selection: 12 Questions to Settle Before You Commit

Vision AI SOP projects rarely fail on the algorithm. They fail because nobody settled the site conditions and the acceptance criteria before the purchase order went out. Split the decision into twelve questions — four on site conditions, three on detection capability, three on integration, two on cost and ownership. If any group cannot be answered, the project belongs in assessment, not in procurement. All twelve can be settled in half a day on site plus one round of historical data, and the usual outcome is not “roll it out everywhere” but “pilot one station”.

Group one: site conditions decide whether it can be done at all

• What exactly is being judged? Is it “was the step performed” or “was it performed correctly”? Detecting that a part appeared and disappeared needs one class of model; judging angle, position and sequence needs another, and the cost differs accordingly.
• What is the cycle time in seconds? Cycle time sets the exposure time available per frame, and with it the compute budget. A station making one unit every three seconds and one making a unit every thirty seconds cannot use the same frame rate or the same model complexity.
• What is the space like, and what gets in the way? Will the operator’s hands cross the target? Will a neighbouring robot or a fixture block the view periodically? Can a bracket be bolted down, or does the camera have to hang from the ceiling?
• Can the lighting be controlled? Ceiling lights, side windows and the illumination from the next station all change, and day shift does not look like night shift. Whether a dedicated light can be fitted decides whether the imaging stays stable.

Group two: detection capability decides how accurate it can be

• Which is worse, a false alarm or a miss? A miss sends a defect downstream or out of the door, and quality owns the consequence. A false alarm stops the line for a check, and production owns that one. Quality and production have to settle this at kick-off; leaving it to acceptance guarantees an argument. The trade-off itself is explained in how false alarms are reduced.
• What false-alarm rate is actually tolerable? Not “as low as possible” but “how many times an hour before the operator stops looking up”. Past that number the alarm gets muted and the system is effectively dead. The reasoning is set out in how to bring false alarms down.
• How are borderline cases judged? A screw half driven, a connector not fully seated, a part resting on the edge of the detection box. Each needs a stated rule and a stated record, or two people will reach two different conclusions about the same event.

Group three: integration and ownership decide whether it gets used

• How does the alarm reach the operator? A screen pop-up, a stack light, an audible tone? A notification that arrives more than two seconds late has lost its power to correct anything, and survives only as evidence for assigning blame.
• Does it write back to MES, and which fields? Operation number, timestamp, serial number, result, exception snapshot. Get the field definitions agreed with IT first or the interface will be reworked twice, and the agreement is easier with how vision SOP connects to MES in front of you.
• How long is data kept, and where? Retention is usually set by customer audit requirements, commonly one to three years. That number drives local storage and network budget, and it has to be confirmed up front rather than patched afterwards.
• What does it cost per year after the build? Lamps age, lenses need cleaning, detection rules follow process changes, software gets updated. This recurring cost is normally estimated as a percentage of the initial spend; leave it out and the second-year budget has nowhere to come from.
• Who owns it on site? Who adjusts a threshold when a false alarm appears, who updates a rule when the process changes? Without a named owner the system quietly becomes nobody’s equipment, and this is how error-proofing measures decay.

When not to start at all

Three situations are worth setting aside. The first is when the root cause sits in the SOP itself — if the standard has no judgeable wording, no number of cameras will decide it, and rewriting the work instruction is the more useful move. The failure modes are set out in why SOP execution fails. The second is a station producing only a few hundred units a year where the missed step can be reworked, because the payback stretches out of reach. The third is a station with no stable power or network position, or with dust and oil mist heavy enough that the lens needs wiping daily. None of these is a technology problem; they are condition problems, and changing supplier does not change them. If all twelve questions can be answered, the next step is a single-station pilot to measure real false-alarm and miss rates, checked against the visual SOP site assessment checklist — and only then a budget conversation, which is covered in cost structure and ROI.