AI Visual Inspection for Small Factories: How a Pilot Runs

AI visual inspection puts a camera over the production line and a model on a small local box that flags defective parts in real time. For a small factory, the realistic first project is one line, one product and one defect type, run as a paid pilot for about six weeks on a Jetson-class device. Published results are mostly vendor claims, such as a 10% to 20% scrap cut in Siemens's Procter and Gamble case, so measure the client's own baseline first.

This guide is part of the edge AI series. It covers three factory jobs: inspection, safety and maintenance.

Why do factories want AI on site instead of in the cloud?

Three reasons come up in every plant.

• Speed. A part on a conveyor is past the reject gate in a second. In Covision's deployment at GKN, NVIDIA reported 200 milliseconds per piece, with one on-premises GPU handling images from four lines.
• Data stays in the plant. Describing its Amberg factory, Siemens wrote that "the important production data stayed inside the production critical level. No interaction with the cloud was needed."
• Cloud services close. Amazon shut down its Lookout for Vision inspection service on October 31, 2025. A model on a box the factory owns keeps working whatever a cloud vendor decides.

What are the three factory jobs a local box can do?

Start with the job where the client can name the cost of the problem. "We scrapped 3% of last month's run" is a project. "We want AI" is not.

How does AI defect detection work?

A fixed camera with controlled lighting photographs each part. A model on the box classifies it as normal or defective, then sends a signal to a light, a reject arm or the line controller.

There are two ways to train the model:

1. Object detection, usually a YOLO model. You label photos of each defect type and train the model to find them, so you need a real stock of defect photos. Covision told NVIDIA it trains on "hundreds of example images" in about an hour.
2. Anomaly detection, using a tool such as Intel's Anomalib (Apache 2.0). You train only on good parts, and the model flags anything that looks different. Anomalib's authors note that supervised approaches are often ineffective here "due to a lack of sufficient representative samples in the anomalous class". Most small plants have exactly that problem: defects are rare.

For a first pilot, anomaly detection is usually the faster start, because the client can give you 500 photos of good parts this week and may only see 20 defects a month.

Classic rule-based vision with OpenCV (Apache 2.0 since version 4.5) still has a place. Measuring a hole diameter or checking a label is present is often simpler with rules than with a model.

What does the software licence cost?

Check this before you quote, because it can wipe out your margin.

Ultralytics YOLO. The most-used detection models are free under AGPL-3.0 only if you open-source your whole project. Ultralytics' licence page says "An Enterprise License is required if you want to use Ultralytics YOLO without open-sourcing your entire project", and lists internal business tools and "embedded deployments in hardware, edge devices, robotics, cameras, or appliances" among the uses that need one. Its FAQ adds that this applies even if you train your own model from scratch or use YOLO only internally. The price is by quote.

Roboflow. Its pricing page has a free tier, a Core plan at $39 a month and custom Enterprise pricing. Running models on your own hardware with Roboflow Inference is free on every plan, but Roboflow notes that Ultralytics models need a special licence to use downloaded weights commercially.

Apache 2.0 tools. Anomalib and OpenCV carry no such condition. Detection models with permissive licences exist too. Check each model's licence file, not the blog post about it.

Put the licence cost in the quote as its own line. A client who later learns about an AGPL obligation from someone else will not hire you again.

What results can a factory expect?

Nobody publishes independent averages for small plants. What exists is vendor case studies, which show what is possible in a good deployment:

Use these to explain what the technology does. Do not promise a client any of them. Measure their current scrap rate, false reject rate and inspection labour hours before the pilot, and report against those.

Can AI detect missing PPE and unsafe behaviour?

Yes, and it usually runs on the cameras the site already has. Protex AI, for example, flags workers "not wearing required personal protective equipment (PPE)" from existing CCTV, and a partner blog reported that Marks and Spencer saw a 40% fall in unsafe events within two weeks and 80% in three months. Intenseye's case studies page claims a 25% drop in one customer's recordable incident rate. Both are vendor-reported.

Two cautions. Staff need to know what the cameras look for and what happens with an alert. And in the EU, the AI Act has banned emotion recognition in workplaces since 2 February 2025, except for medical or safety reasons. A system that detects a hard hat is fine. One that claims to read a worker's mood is a different legal question. In the UK, the ICO says a DPIA is needed if you use CCTV to monitor staff. The camera rules are covered in more detail in AI cameras for shops and restaurants.

How does predictive maintenance work on a small budget?

Sensors on motors, pumps and fans record vibration and temperature. A model learns what normal looks like and warns before a bearing fails.

The best-known numbers come from the US Department of Energy's O&M best practices guide: "Past studies have estimated that a properly functioning predictive maintenance program can provide a savings of 8% to 12% over a program utilizing preventive maintenance alone." The same guide quotes unnamed surveys claiming a 35% to 45% cut in downtime. Those surveys are not identified, so use the 8% to 12% figure. The guide is from 2010; PNNL's 2021 update repeats it.

The guide also warns that "much of the equipment requires cost in excess of $50,000". That was written about traditional vibration analysis kit. Cheap sensors and a small local box bring the starting cost down, but the time cost stays: a model needs weeks to months of normal running before its warnings mean anything. Sell maintenance as a slow-burn retainer, not a quick win.

How does a six-week inspection pilot run?

1. Week 1: baseline and site visit. Record the current scrap rate, false rejects and inspection hours for one product. Photograph the line, check lighting and power, and agree in writing what "good" and "defect" mean.
2. Week 2: capture. Mount the camera and lighting. Collect several hundred photos of good parts and every defect you can get. Ask the quality team to keep rejected parts in a box for you.
3. Week 3: train. Train an anomaly model on good parts, or a detection model if there are enough defect photos. Test it on photos it has not seen.
4. Week 4: shadow mode. Run the model on the line with no control over the reject gate. Log every decision and compare it with the human inspector.
5. Week 5: tune. Adjust thresholds with the quality manager. Decide together how many false rejects are acceptable for each missed defect.
6. Week 6: report. Show the numbers against the week-1 baseline and propose the full install with a retainer. If the numbers are poor, say so. A pilot that ends in "not yet" still pays and keeps your name clean.

Charge for the pilot. A free pilot gets treated as a free pilot. Pricing is in how to price local AI installs.

Who pays for a factory pilot?

The factory, but it may not pay alone. In the US, the NIST Manufacturing Extension Partnership awarded more than $30 million in September 2026 to help small and medium manufacturers adopt advanced technology including AI, through a network of about 1,400 advisers. In Singapore, Enterprise Singapore's EDGE Grant replaced PSG and EDG from 30 September 2026 and supports up to 50% of qualifying costs for SMEs. Its FAQ says work started before the application is not supported, so the client applies first and you start after.

Knowing these schemes is a sales skill. The client sees you understand their budget, not only the technology.