When AI Visual Inspection Is Worth It, and When It Is Not

---

Overview

01

The honest version of the AI vision question

Most articles about ai visual inspection answer a question nobody asked. They explain what it can do, and skip whether you should.

The technology works. Cameras plus a trained model catch defects that tired eyes miss, at speeds no human matches.

That does not make it right for your line. Plenty of inspections fail the fit test, and the cost of finding out after the pilot is high.

This is the decision framework we use with clients before anyone buys a camera. For the wider service view, see digital quality inspection.

---

Fit Test

02

Five conditions that decide it

Score your case against these before anything else. Three or more misses means the answer is no, at least for now.

One more thing decides more projects than any of the six: who signs off a reject. If nobody will accept a machine deciding, the system becomes an advisor and the payback maths change completely.

---

Wrong Tool

03

When something simpler is the right answer

Two cases come up often, and in both a model is the expensive way to be worse.

You need a measurement. If the drawing says 12.50 with a tolerance of five hundredths, that is a gauge job. Vision can measure with calibration and good optics, and a classifier trained on photographs cannot.

The rule is simple and fixed. Presence or absence of a bright component on a plain background is rule-based machine vision. It has been solved for thirty years, it is cheaper, and it explains its decisions.

Question	Rule-based vision	AI vision
Setup effort	Thresholds, tuned by an engineer	Image collection and labelling, then training
Handles new defect types	Only if someone codes the rule	Yes, once examples exist
Tolerance to lighting drift	Low	Higher, if trained for it
Explains a reject	Yes, the rule that fired	Partly, through a heat map
Best fit	Fixed, high contrast checks	Varied surfaces and defect types

---

Real Cost

04

The model is the cheap part

Budgets get built around the algorithm. The algorithm is rarely the problem.

Lighting and fixturing usually take longer than training. Controlled, repeatable illumination is what makes a mediocre model work and a good model reliable.

Then comes the data. Collecting and labelling enough defect images is weeks of somebody’s attention, and the rare defects are the ones you have fewest pictures of.

Rare defects deserve their own plan. If a critical fault happens twice a year, you will not collect enough examples in a pilot, so either produce samples deliberately or accept that the model covers the common faults and people still watch for the rest.

Last is the plumbing: triggering to the line, writing the verdict where the MES can read it, driving the reject mechanism, and storing the image with the part record. The pipeline is the deliverable, not the model.

---

Failure Modes

05

Four ways these projects go wrong

None of them are exotic. All four are avoidable at design time.

1. Drift nobody watches. Lighting ages, a supplier changes a coating, and accuracy slides quietly. Without monitoring, the first sign is a customer complaint.
2. Labels of poor quality. Two inspectors disagree on what counts as a defect, the training set contradicts itself, and the model learns the confusion.
3. Skipping shadow mode. A system that rejects parts on day one has never been checked against ground truth. Run it alongside people until the numbers agree.
4. No owner for retraining. New products arrive, defects change, and nobody has the job of updating the model. Twelve months later it is switched off.

There is a quieter fifth. Operators stop trusting a system that cries wolf, and once they start waving parts through, the recorded pass rate stops meaning anything.

The same list applies to automated visual inspection bought as a finished product. Buying rather than building moves who does the work, not whether it has to be done.

---

Action

06

A two week feasibility, before any purchase

You can answer the fit question cheaply, and long before a vendor quotes.

1. Write the defect list. Name every defect type you want caught, with how often each one occurs. Rare and critical is a different problem from common and cosmetic.
2. Photograph a week of production. Good parts and bad, in the conditions the line actually runs, not in a lightbox.
3. Check visibility by eye. If you cannot see the defect in your own photographs, a model will not either. This step ends a surprising number of projects.
4. Price the whole loop. Camera, optics, lighting, fixture, edge hardware, integration and the annual maintenance, not just the software licence.
5. Compare against the cheap option. A second pair of eyes at the pack station, or a rule-based check, with the same payback maths.

If the fit test passes, the next question is scope. Start with one defect type on one line rather than the whole inspection station, because a narrow model reaches useful accuracy far sooner than a broad one.

Dev Station builds inspection systems where computer vision inspection is one component beside capture, records and reporting, and we say no when the fit test fails.

Most of that work is for manufacturers in the United States and the United Kingdom, so image retention and model records follow SOC 2 or HIPAA on one side and GDPR with ISO 27001 on the other.

Our engineers work from Vietnam with overlap into US Eastern, US Pacific and UK GMT hours, and we invoice in USD or GBP. Send us your defect list and a folder of part images and we will come back with a straight answer on feasibility.

---

FAQ

07

Frequently asked questions

What manufacturers ask before committing to a pilot.

When is AI visual inspection worth it?

When volume is high, the defect is visible to a camera, presentation is stable, labelled examples exist or can be collected, and an escape actually costs money. Miss several of those and a trained inspector or a rule-based check will serve you better.

How is it different from machine vision?

Rule-based machine vision follows thresholds an engineer sets. An AI model learns from labelled images, so it copes with varied surfaces and new defect types, at the cost of needing data and explaining itself less clearly.

How many images does a model need?

It depends on how varied the defect looks, so treat any single figure with suspicion. The practical constraint is usually the rare defects, where you may have a handful of examples and need months of collection or deliberately produced samples.

Can it do dimensional checks?

Vision systems can measure with proper calibration and optics, but a classifier trained on photographs cannot. If the drawing states a tolerance, plan for a gauge or a calibrated measurement system and use ai defect detection for the things that are judged rather than measured.