Buying the camera is rarely the hard part. Getting an inspection system to hold up on a running line, across shifts, part variants, a new material lot, and an operator who loads the fixture slightly differently at 2am, is where vision projects go quiet. That work belongs to whoever integrates the system, which is why the integrator you pick shapes the outcome more than the sensor you pick.
This guide covers what a machine vision system integrator does, when you actually need one, the questions worth asking before you sign, and how AI vision has changed what a good partner looks like.
A machine vision system integrator designs, builds, installs, and validates a camera-based inspection system on your production line. Bring one in when parts vary, cycle times are tight, or the line has to act on the decision in real time. Evaluate on proven application evidence, hardware independence, and who owns the system after handover.
What does a machine vision system integrator actually do?
A machine vision system integrator takes an inspection requirement and turns it into a working station on your line. The camera is one line item in that scope. The rest of the work decides whether the station survives production.
A full scope usually covers:
- Application study. Which defects matter, how often they occur, what a false reject costs you, and whether the defect is even visible to a camera at line speed.
- Optics and lighting design. Lens selection, working distance, and the lighting geometry that makes the defect appear. This is where vision projects are won and lost, long before any software runs.
- Mechanical integration. Mounting, part presentation, guarding, encoder or trigger placement, and the reject mechanism.
- Inspection logic. Rule-based tools, trained models, or a mix, plus the pass and fail thresholds you agree on.
- Controls handshake. Talking to the PLC, handling trigger timing, logging results, and pushing data to MES or a dashboard.
- Validation and handover. Factory acceptance, site acceptance, operator SOPs, training, spares, and a support agreement.
If a quote covers hardware and software with a single line for "integration", that is the part of the project you know least about and pay for twice. Our guide to inspection systems in manufacturing breaks down how these pieces fit together.
Do I need an integrator, or can my engineering team handle this in-house?
Your team can handle it when the imaging problem is stable. Fixed part presentation, one variant, a high-contrast defect, an off-the-shelf smart camera, and a maintenance engineer who owns the station long term. Plenty of presence-absence checks and barcode verifications belong exactly there.
Bring in an integrator when any of these are true:
- Parts arrive in random orientation or free flow rather than a fixture.
- The surface is shiny, curved, transparent, or textured, so lighting has to be engineered.
- The defect class is fuzzy and operators themselves disagree on borderline parts.
- The decision has to be made inside the cycle and drive a reject.
- The plant is regulated and the system needs a validation trail.
AI vision moves that line. When the inspection logic learns from images instead of hand-tuned thresholds, some variation gets absorbed by the model rather than engineered out with fixtures and light tents. That reduces mechanical work and shifts effort towards image collection and labelling. Our AI vision inspection setup guide walks through what that looks like in practice. For a view from outside vendor material, the A3 note on how to select a systems integrator and Quality Magazine on whether you need one at all are both worth twenty minutes.
Why did our last vision project pass the trial and then fail on the line?
Because the trial ran on a curated sample under controlled light, and production does none of that. This is the most common failure story in machine vision, and it is a scoping failure rather than a hardware failure.
What breaks a validated station after go-live:
- Ambient light changes when the shutter door opens or afternoon sun hits a skylight.
- A new supplier lot arrives with a different surface finish, so a threshold tuned on the old finish starts flagging good parts.
- A part revision adds a feature the recipe was never shown.
- Operators load slightly off-centre once the line speeds up.
- Lens fouling, vibration, and slow focus drift that nobody has on a PM checklist.
Ask every integrator on your shortlist a direct version of this: how does your system behave on a part it has never seen? A rule-based recipe will usually fail it or pass it silently. A trained model should flag it as low confidence and route it for review. The cost of getting this wrong shows up later as scrap, sorting, and customer complaints, which we cover in the cost of poor quality breakdown.
What should I ask a machine vision integrator before signing?
Seven questions separate a partner from a supplier. The wording matters less than what you listen for in the answer.
| Question | What a strong answer sounds like | Warning sign |
|---|---|---|
| Have you solved this exact defect before? | A named application, sample images from that line, and a reference you can call | "Vision can do anything, send us parts" |
| Whose hardware are you tied to? | Camera, lens, and lighting chosen per application, with the reasoning explained | The same brand and model on every project |
| What accuracy do you commit to, and how is it measured? | False accepts and false rejects quoted separately, against an agreed sample set | One headline accuracy number with no test protocol |
| What happens when the part changes? | A defined retraining or re-recipe process, who runs it, and how long the line waits | A change request and a fresh quote every time |
| Who owns the images and the trained models? | You do, and they are exportable | Both live inside the vendor platform permanently |
| What exactly is handed over? | SOPs, operator training, spares list, escalation path, response times in writing | A PDF manual and a support email address |
| Can any of this run on our existing cameras? | Yes for process and workflow monitoring, with the limits stated honestly | Every requirement needs new hardware |
Ask for the reference call. An integrator with a genuinely comparable installation offers it before you finish the sentence.
What drives the cost of a machine vision project?
The inspection specification drives cost far more than the hardware list. Two projects with identical cameras can differ by an order of magnitude in engineering effort.
The variables that move the number:
- Defect complexity. One obvious defect on a flat surface is cheap. Twelve cosmetic classes that look similar to each other is a research project.
- Cycle time. Faster lines force better lighting, faster sensors, and tighter triggering.
- Number of stations and views. Every extra angle multiplies optics, mounting, and calibration work.
- Part presentation. Fixtured parts are predictable. Free-flowing parts on a conveyor are not.
- Surface behaviour. Specular, transparent, and dark textured surfaces all need engineered lighting.
- Environment. Washdown, dust, heat, and vibration change enclosure and mounting specifications.
- Data and traceability. Storing every image against a batch ID for audit is a different build from a simple pass or fail signal.
- Validation burden. Regulated plants add documentation effort that has nothing to do with the camera.
When you compare quotes, insist on a split across hardware, engineering, and post-handover support. Budgets usually break on the third line, because a system nobody maintains gets bypassed within a year and then quietly switched off.
Should I hire a traditional vision integrator or an AI vision vendor?
It depends on whether your inspection problem is measurable or judgemental. Rule-based vision is excellent at measurement. Trained models do better where a human inspector would say "that one looks wrong" and then struggle to write the rule down.
| Dimension | Traditional rule-based integrator | AI vision vendor |
|---|---|---|
| Best fit | Gauging, measurement, presence and absence, barcode and OCR, tight tolerances | Cosmetic defects, surface variation, assembly correctness, mixed part families |
| How the logic is built | Engineered tools and thresholds tuned per part | Models trained on labelled production images |
| A new defect type appears | Recipe rework, often a site visit | Add examples and retrain, sometimes by your own team |
| Lighting tolerance | Low, so lighting must be locked down | Higher, though good lighting still helps |
| Time to first useful result | Slower to build, then very stable | Faster once images exist, improves with data |
| Ongoing ownership | Usually the integrator | Can sit with your quality team if models are exportable |
| Where it struggles | Variation it was never tuned for | Sub-pixel metrology and thin training data |
Plenty of plants need both. Metrology stations stay rule-based, while cosmetic inspection and assembly verification move to trained models. A comparison of the two approaches on real production problems sits in KOMPASS versus traditional inspection methods, and the two-layer view of product inspection against process monitoring is in machine vision for quality control.
How do I know the system will still work a year from now?
Get the maintenance model in writing before the purchase order, because that is the only point in the deal where you still hold the money.
A workable ongoing model includes:
- A monthly review of false reject rate, not only defect capture, since operators bypass stations that cry wolf.
- A defined trigger for retraining or re-recipe, tied to part changes and supplier changes.
- Image retention with enough history to retrain later.
- Named people trained to label and approve changes.
- Spare camera, lens, and light on the shelf, with lens cleaning and lighting checks inside the PM schedule.
- An escalation path with response times, and remote access agreed with your IT team in advance.
Systems that verify process rather than product need a different maintenance rhythm again, and many of them run on cameras you already have. That approach is covered in real-time production monitoring using existing cameras, and the wider quality goal it serves is explained in right first time manufacturing.
What does a realistic project sequence look like?
A single-station project moves through five stages, and skipping any of them is how teams end up with an expensive camera pointed at a problem it cannot see.
- Feasibility. Sample parts, including known bad ones, imaged under candidate lighting. The output is evidence the defect is detectable, or an honest no.
- Offline proof. Logic or a model built on real plant images, scored against a hold-out set you agree on.
- Pilot. One station running live in shadow mode, flagging without rejecting, while you compare its calls against your inspectors.
- Acceptance. Agreed pass and fail criteria tested at line speed, with false accepts and false rejects counted separately.
- Ramp and handover. Operator training, SOPs, documentation, and the support agreement starting from a defined date.
Shadow mode is the stage teams most often cut to save time, and it is the one that earns operator trust. A station the line believes in gets used. A station that rejects good parts in week one gets bypassed and never recovers.
Frequently Asked Questions
1. What is the difference between a machine vision integrator and a machine vision manufacturer?
A manufacturer builds components such as cameras, lenses, lighting, or vision software. An integrator combines those components into a working inspection station on your line and takes responsibility for the result. Some companies do both, which is worth knowing, because a vendor that manufactures hardware has a reason to recommend its own hardware.
2. Do machine vision integrators work with existing production lines?
Yes, and retrofits are the majority of projects. The constraints are physical space around the conveyor, available trigger signals, network access, and how much line stoppage you can allow for installation. Process monitoring systems have the lightest footprint, because they can often run on existing CCTV coverage instead of new inline hardware.
3. How do I evaluate an integrator's accuracy claim?
Ask for accuracy split into false accepts and false rejects, measured on a defined sample set that includes borderline and rare defects, at production speed. A single accuracy percentage with no test protocol behind it tells you very little. Then ask what happens to that number when a new part variant arrives.
4. Can one vendor handle both product inspection and process monitoring?
Some can, though the two are different problems. Product inspection checks the item that leaves the station. Process monitoring checks whether the operator and the sequence followed the standard. Plants often need both, and running them as one project keeps the data in one place instead of two disconnected dashboards.
5. What should a machine vision scope of work document contain?
Defect definitions with images, cycle time, part variants covered, lighting and mounting responsibility, acceptance criteria for false accepts and false rejects, PLC and MES interfaces, data ownership, training deliverables, spares, and support response times. Anything left out of that document becomes a change order later.
Choosing well comes down to evidence
The integrator worth signing will show you images from a line like yours, tell you which defects they cannot catch, quote false rejects alongside detection rates, and explain what happens the day your part changes. The one to avoid promises everything and specifies nothing.
Jidoka Technologies builds AI vision systems for manufacturing plants, with KOMPASS inspecting what leaves the line and Nagare verifying how it gets made, often on cameras that are already installed. If you want a straight answer on whether your inspection problem is a vision problem at all, book a conversation with our team and bring your worst parts.
