Short answer: Buy a machine vision system the way you would hire an inspector: with your own parts, under your own lighting, against error definitions you wrote down first. The suppliers worth shortlisting — the vision specialists listed in the directory's machine vision section, from camera makers such as Basler to inspection-system houses such as Cognex — expect these questions. The ones who resist them are answering a different question than the one you asked.
Question one: will you test our parts, not your samples?
Demonstration parts are chosen to be demonstrable. Your parts are not. Ask the supplier to run a sample test — a feasibility study — on parts you supply, including the awkward ones: the reflective finish, the borderline defect, the variant that only fails on Fridays. Ask how many parts the test should include to mean anything, who labels the ground truth (which parts are genuinely good and genuinely bad), and whether you get the images and results, not just a summary score. A supplier's sample-test process is the single best preview of their production behaviour; treat the test itself as the first deliverable and judge it as one.
Define the two errors before anyone quotes a rate
Vision systems make two kinds of mistake: accepting a bad part (a false accept, an escape) and rejecting a good one (a false reject). They trade against each other, and a single accuracy figure hides the trade. Ask each bidder to state expected performance as the pair — escapes and false rejects, separately — measured on your sample set, and ask what each error costs in your process: an escaped defect may mean a customer return or a safety issue, while false rejects mean scrap and re-inspection labour. Then write the pair you are buying into the acceptance criteria, with the sample set it was measured on. Suppliers who quote one heroic number are quoting the average of two numbers you needed separately.
Whose problem is the lighting?
More vision projects fail on light than on algorithms. Ambient light changes — skylights, shift changes, a new lamp over the line — alter what the camera sees, and a system tuned in a dark lab drifts in a bright plant. Ask who owns the lighting design, whether the enclosure and light sources are inside the quoted scope, and what happens to performance when the plant's lighting changes anyway. Ask, too, about the mundane enemies: oil mist on the lens, vibration at the mounting point, parts arriving hotter, wetter or dustier than the samples. The answers divide suppliers into those who have installed systems in factories like yours and those who have installed them in brochures.
Software: licences, recipes and who may change what
Ask how the software is licensed (per camera, per line, per site), what a new product variant — a new recipe — costs to create and who is allowed to create it, and whether your engineers can adjust thresholds and add parts without a supplier visit or a licence dongle. Ask what is logged: a vision system that records every image and decision is an audit asset; one that discards them leaves you arguing from memory when a customer disputes a shipment. And ask the exit question early: if you change suppliers in five years, what happens to your recipes, your image archive and your trained models?
Interfaces are a question, not an assumption
The vision system must talk to the robot, the PLC and the reject mechanism, and the handshakes are where integration hours hide. Name the exact interfaces in the enquiry — the fieldbus or protocol, the trigger source, the reject confirmation signal, the data your line control expects — and ask the supplier to confirm each in writing rather than with the phrase 'standard interfaces'. Ask also about timing: at your line rate, how much time does the full cycle of trigger, exposure, processing and decision actually take, and what degrades first when the line speeds up?
The shortlist version
Compressed into a first meeting: Will you test our worst parts? What are your escape and false-reject rates, separately, on our samples? Who owns lighting and enclosure? What does a new recipe cost, and can our engineers make one? What exactly leaves your system, in what format, when we part ways? Which interfaces will you confirm in writing? Six questions, and the quality of the answers — specific, written, a little reluctant about the hard parts — is itself the feasibility study of the supplier.
Frequently asked questions
What is the most important question to ask a machine vision supplier?
Whether they will test your own parts — including your worst, borderline and reflective ones — and report escapes and false rejects separately on that sample set. A demonstration on the supplier's chosen parts proves only that the demonstration works.
What is the difference between a false accept and a false reject?
A false accept lets a defective part pass; a false reject scraps or diverts a good part. The two trade against each other, so a single accuracy figure is meaningless — always ask for the pair, measured on your parts.
Why does lighting cause so many vision project failures?
Because systems are often tuned under controlled light and then installed where skylights, shift patterns and new lamps change the scene. Ask who owns lighting design and enclosure in the quoted scope, and what happens to performance when plant lighting changes.
What should I ask about vision software licensing?
How it is licensed (per camera, line or site), what a new product recipe costs and who may create one, whether your engineers can adjust the system without a supplier visit, what images and decisions are logged, and what you keep if you change suppliers.