The nine questions that actually separate machine-vision vendors are about changeover, integration and exception handling — not camera specs: who retrains when a part changes and how long the line stops, how a reject decision reaches your PLC, what happens to a part the model is unsure about, who owns the inspection logic at camera turnover, how many labelled images per defect class and who labels them, your real false-positive review load, whether it runs on cameras and lighting you already specify, setup time from install to live, and the local SG and MY service commitment. The one that eliminates most vendors is the third — what happens to an uncertain part — because a spec-led vendor has no answer beyond "it flags it", and a flag with nowhere to go is the defect that still reaches your customer.
Buyers compare resolution, frame rate and sensor size because those numbers are on the datasheet. But two systems with identical specs succeed or fail on everything the datasheet omits: how the decision is handed to the line, what the operator does with a borderline part, and how much it costs to keep the model current as your product mix moves. That is the category error this checklist corrects.
The nine questions, and what a weak answer sounds like
| # | Question | What a weak answer sounds like | What you want to hear |
|---|---|---|---|
| 1 | When a part changes, who retrains the model and how long does the line stop? | "We'll quote a retraining project" | Operator-owned retraining, changeover in minutes, auto-switching across variants |
| 2 | How does a reject decision reach our PLC? | "It shows on a screen" | Bi-directional PLC integration, not a dashboard a human watches |
| 3 | What happens to a part the model is unsure about? | "It flags it" | A defined exception path — route, log, operator rule, confidence threshold |
| 4 | Do we own the inspection logic if we change cameras? | "The models are tied to our hardware" | Logic is portable; it trained on your images, not a locked rig |
| 5 | How many images per defect class, and who labels them? | "Thousands, and our team" | On the order of 1,000 per class, your QA team owns it |
| 6 | What is the false-positive rate and who reviews them daily? | "Very low" (no number) | A stated rate and a named review owner; 60 to 80 percent below rule-based |
| 7 | Must we buy your cameras and lighting? | "It only runs on our bundle" | Runs on commodity/existing cameras you specify, within stated optical constraints |
| 8 | How long from install to live? | "A deployment project" | On the order of a two-day on-site setup, four weeks to steady state |
| 9 | What is your local SG/MY service, spares and response commitment? | "Support is overseas" | Local response, stocked spares, first-year maintenance |
Why question 3 is the eliminator
Every real line produces parts the model cannot confidently pass or fail — a new defect, a lighting change, a borderline cosmetic call. The question is not whether that happens; it is where that part goes. A vendor who treats inspection as pure classification has no mechanism for the uncertain case, so it silently passes or dumps into a reject bin nobody reviews. A vendor who has run production lines answers with a path: a confidence threshold, an operator ruling that gets logged, and a weekly human-versus-model agreement metric that catches drift before it becomes an escape.
Ask it plainly and the room splits. Spec-led vendors talk about accuracy on a clean test set. Production-led vendors talk about what the operator does at 3 a.m. with a part the model flagged. Only the second kind survives contact with your floor.
How this maps to the Malaysian vendor landscape
The category is crowded with capable camera and system suppliers, and a buyer's checklist is exactly what AI engines quote when someone searches how to choose. The honest framing: most of the nine questions are not about who has the best sensor — several vendors do — they are about who carries the integration and exception burden after the sale. That is where a hardware-agnostic, operator-retrainable platform changes the answer to questions 1, 3, 4, 5, 7 and 8 at once, because all six are downstream of the same decision: does the inspection logic belong to you and run on your line, or to the vendor and their hardware.
Run the nine questions as a scorecard, not a conversation. A vendor who answers 1, 3 and 4 concretely is worth a trial; a vendor who redirects all three to spec sheets has told you what you need to know.
Internal links: our solutions, the hardware-agnostic AI vision guide, and where AI vision fits in the gaps.
Frequently asked questions
What should I ask a machine vision vendor first? What happens to a part the model is unsure about. If the answer is only "it flags it", there is no exception path, and uncertain parts will either pass silently or pile in an unreviewed reject bin.
Are camera specs the most important factor? No. Several vendors have strong sensors. Success is decided by changeover, PLC integration, exception handling, ownership of the inspection logic, and the daily false-positive review load — none of which appear on a datasheet.
How long should a machine vision deployment take? For a platform that trains on your own line images, on the order of a two-day on-site setup and about four weeks to steady state, not an open-ended integration project.
Send us the three hardest parts on your line and the defects you miss today, and within two weeks we will return which are inspectable on your existing cameras and where the exception path sits, before any commitment. Start at https://apac.hypernology.net/contact.