An article in Supply Chain Management Review by Sara Hsu, an associate professor of supply chain management at the University of Tennessee, argues that focused AI applications deliver much better returns than large language model AI that attempts to operate entire supply networks autonomously. In short, AI-based systems that extend familiar processes rather than replace them are better. Leading supply chain system vendors are talking about Agentic AI orchestration to manage an end-to-end supply chain. I agree with Ms. Hsu; we aren’t there yet.

There are several examples of AI-based solutions that provide strong ROI by extending traditional operations. One I wasn’t familiar with was AI machine vision systems that inspect products at the end of a manufacturing line. I talked to the CEO of Overview , Chris Van Dyke, and one of their customers, SKF, to learn more about this.

Overview was founded in 2019, before the current wave of AI rhetoric became so prevalent. Mr. Van Dyke worked at Tesla for almost a decade. He helped ramp up the Tesla battery factory in Reno and had “lived through the challenges of building highly automated factories. I also lived in San Francisco and had friends in the AI space.”

He saw the improvements in computer vision and realized that AI-camera systems for defect detection would be a great application for manufacturing. He also realized there was little competition. “So I left Tesla with another colleague, and then we joined with two software engineers who had AI expertise, and started the company with a combination of a deep understanding of manufacturing and expertise in computer vision.”

For the first few years, it was a consulting services-based firm. The company launched its first product in 2023 and has grown quickly since then. Overview now has over 70 employees worldwide and is on track to generate $25 million in revenue this year. Their solutions are deployed in hundreds of factories across 20-plus countries. Their value proposition is that the system excels at defect detection, is easy to use, easy to set up and embed in an existing process, and runs on the edge without needing an internet connection.

The AI model is trained to detect a wide range of physical defects in manufactured parts. These include, but are not limited to: surface cracks, scratches, or dents; misalignments or dimensional deviations; missing components; contamination or foreign objects; and color or texture anomalies. One model does all of these things; it is not necessary to have separate systems working together.

Some supply chain solution providers claim to be AI-native and started around the same time as Overview. Their claims are often dubious. Overview really is AI native. Tesla was trying to develop autonomous cars. They had “cameras in the cars that were looking at the road “ that were really doing a lot of sophisticated interpretation and inference,” Van Dyke explained. “The cameras you could buy in the factory were just much further behind.”

One claim, verified by their customer SKF, is that their learning models can be trained in an hour. Users label images to distinguish between “good” and “bad” parts.

In Karen Hao’s best-selling nonfiction book Empire of AI: Dreams and Nightmares in Sam Altman’s OpenAI , she describes the armies of low-paid workers in Kenya and Venezuela who helped train the models. It took years to train the initial large language models.

But at Overview, they are using ResNet models for deep learning-based inspection. This architecture uses “skip connections” to train very deep neural networks without losing data or accuracy. Van Dyke also points out that the training speed reflects the fact that this is a narrow use case. The camera system also uses a powerful NVIDIA chip to enable edge computing and rapid training.

The company also does not use a hardware-as-a-service pricing model that locks customers into annual fees. They also do not use token pricing, which can get pricey. This is an upfront, one-time purchase. The smart cameras are priced between $5k and $15k, while newer PC-based systems cost $25k or more. One thing to note: on some lines, you might need 4 or more cameras working together, so the system price would be higher.

So, what do their customers say?

SKF and AI-based Product Inspections

I talked to Dave Thomas, a controls and automation engineer at SKF (Nasdaq: SKF A and SKF B), to hear the voice of the customer. SKF is a global Swedish engineering company with around 140 factories worldwide. The company specializes in bearings, seals, lubrication systems, and condition monitoring services. Mr. Thomas works in a factory in Flowery Branch, Georgia.

SKF had been looking for a vision solution for years and had tried several manufacturers. Mr. Thomas wanted a solution that performed very accurate inspections and was easy to set up. With other solutions, you “basically had to be a data scientist in order to maintain the systems long term.” Before implementing Overlook, a project “would involve many hours of setup, programming, and modification in order to accommodate a new product line.” SKF also is pursuing a digital manufacturing strategy, and AI fits into that vision.

The Overview AI system allows users to highlight specific features, such as color or shape, to train the model. Because the system looks for specific features rather than whole models, many different parts can often use the same trained images if they share similarities.

The local facility currently has three active camera systems. Another plant is also using this solution. SKF plans to add nine more systems.

The initial use case was fairly simple. Set-up and implementation only took about a day and a half. Training involves capturing images from a variety of positions and angles, then telling the system whether the product is good or not.

But they are now using Overview in more complex ways. One system uses four cameras performing 32 different functions. This system requires various lighting setups and part rotation. “Obviously, the higher the complexity of the system,” Mr. Thomas pointed out, “the longer it takes for the implementation. We’ve since moved on to significantly higher degrees of automation and complexity. The Overview system has adapted quite well to very complex systems.”

The system verifies that specific features are on a product. Both humans and the camera system are used to search for defects. If either a person or the system identifies defects, the product flunks the inspection. Longer term, they want to solely use camera systems for inspection.

SKF aims for a payback period of less than 24 months for their capital investments. They will achieve this. More important than short-term ROI, however, is protecting the company’s reputation by preventing defective components from reaching customers. Mr. Thomas praised Overview for its excellent customer support, on-time deliveries, and regular software updates.

Moving forward, the global manufacturer plans to integrate the vision system more deeply into their MES and other systems.

SKF is also moving toward focused AI to predict maintenance, quality, and setup issues. Other future digital goals include more streamlined communication between global factories and integrated data sharing across maintenance, scheduling, and ordering systems.