Conclusion

To put it simply, the era has arrived where AI can work 24 hours a day at the cost of one inspector.

How much do factories pay for visual inspections annually?

For one part-time inspector at 1,200 yen per hour, working 8 hours a day for 250 days, that totals 2.4 million yen per year. If there are two inspectors, the cost rises to 4.8 million yen. Yet, even then, the chances of missing defects are not eliminated. The human error rate is generally said to be between 0.5% to 1%. If you process 1 million items, that means 5,000 to 10,000 items could be either missed or over-detected. The costs associated with rework and customer complaints do not show up in the numbers.

Now, three technologies have come together simultaneously. Perceptron’s factory-specific visual AI model, IBM’s Granite 4.2 (a lightweight model compatible with local inference), and Liquid AI’s Pipette (a model performance evaluation tool for edge devices). All share the common feature of “not needing to connect to the cloud” and “not requiring a massive GPU server.”

The question is simple: Will this combination truly change the inspection lines in small factories? How much can be achieved, and at what cost?

The Significance of the Three Technologies Coming Together—Exceeding the Practical Line of ‘Small AI’

Perceptron: Factory-Specific Visual AI

Perceptron offers a visual recognition model tailored for factory environments. The difference from generic image recognition AI is clear; it is optimized for common factory issues such as lighting variations, metal reflections, and fine scratches. Importantly, this model is designed to operate on edge devices (small computers placed on-site), eliminating the need to send data to the cloud.

What does this mean? Since inspection images do not need to be sent externally, there is no need to worry about non-disclosure agreements (NDAs) with clients. It doesn’t matter if the factory’s internet connection is weak. Additionally, there are no cloud usage fees.

IBM Granite 4.2: A ‘Decision-Making AI’ That Operates Locally

Granite 4.2 is a language model released by IBM under the Apache 2.0 license, available in three sizes: 3B, 8B, and 30B parameters. Notably, the 3B and 8B models can perform inference on edge devices of the NVIDIA Jetson class.

What happens when combined with visual inspection AI? Perceptron determines, “Is this a scratch?” while Granite 4.2 generates a report stating, “This scratch pattern is likely due to mold wear in process A. There has been a 15% increase in similar detections over the past 30 days.” This allows for not just inspection but also cause estimation and trend analysis to be completed on-site.

Previously, such analysis would have required large-scale AI in the cloud or hiring a specialized data scientist. This meant monthly cloud API costs in the tens of thousands of yen or hiring personnel with salaries exceeding 6 million yen. Now, it can run on a local small device.

Liquid AI Pipette: Pre-Validation of ‘Which Device It Will Run On’

Pipette is an open-source benchmarking tool that allows users to measure the inference speed, accuracy, and power consumption of AI models on their existing edge devices.

While it may seem unremarkable, this could be the most critical factor for small and medium-sized enterprises. Why? Because they can answer the question, “Will this AI model run on our factory’s PC?” on their own before implementation. This eliminates the risk of waiting two months and paying 500,000 yen for a quote from a system integrator only to be told, “The specifications were insufficient.”

Calculating the Break-Even Point—’How Many Months Until It Pays Off’

Let’s move from abstract discussions to concrete calculations.

Assumptions

Comparison: Current Costs of Visual Inspection

Break-Even Point

Annual cost after AI implementation: Initial investment of 500,000 yen + monthly cost of 50,000 yen × 12 months = 1.1 million yen (first year); from the second year onward, it will be 600,000 yen/year.

Difference from visual inspection costs: 2.2 million yen savings in the first year. From the second year onward, 2.7 million yen/year savings.

Payback period: Approximately 3 months.

Initial investment of 500,000 yen ÷ monthly savings (approximately 220,000 yen) = 2.3 months. By the third month, it is projected to turn a profit.

Of course, this is the maximum value based on the premise of “completely replacing one inspector.” In reality, many cases start with a “semi-automation” approach where a person makes the final confirmation of the AI output. Even so, if the inspector’s working hours are halved, that results in an annual saving of 1.65 million yen. The payback of the initial investment can be achieved within four months.

The True Value Lies Beyond ‘Cost Reduction’

While cost calculations are straightforward, the essence goes beyond that.

It can operate 24 hours a day. Human inspectors are limited to 8 hours. AI does not stop. If inspection processes can be incorporated into unmanned night operations, production volume could potentially increase by 1.5 to 2 times. The impact on revenue growth far outweighs cost savings.

Elimination of Dependency on Individuals. The criteria for determining scratches that “only that veteran can distinguish” are retained as data within the model. Even if a person leaves, the knowledge does not disappear.

Data Accumulation. Inspection results are automatically recorded, visualizing trends in defect rates, time-based patterns, and deterioration patterns by mold. This is an asset that could never have been obtained through visual inspection. It can also be used directly as evidence for ISO audits.

A Structure Where Small and Medium Enterprises Can Compete with Large Corporations is Emerging

This is the most important point I want to convey.

Large corporations have already invested tens of millions to hundreds of millions of yen in building inspection systems. Therefore, switching is not easy. There are concerns about compatibility with existing systems, internal approvals, and vendor contracts. They move slowly.

On the other hand, small and medium enterprises have “not implemented anything yet.” This allows them to introduce the latest lightweight AI from scratch. With an initial investment of 500,000 yen and a monthly cost of 50,000 yen, they can start as early as next month with the president’s decision.

This is similar to how cloud accounting penetrated small and medium enterprises before large corporations’ ERP systems. The absence of legacy systems becomes the greatest weapon.

So, What Should Be Done?

I propose three steps.

1. First, validate with Pipette (Cost: Zero). It’s open source. You can test how fast the visual AI model runs on your existing PCs or devices. This will clarify whether it will run on your hardware.

2. Test with just one line (Cost: 500,000 yen + 50,000 yen/month). There’s no need to implement it across all lines. Start with the line that has the highest defect rate and run it in parallel with visual inspections. Within a month, you can gather data on the AI’s detection accuracy and missed detection rates.

3. Make decisions based on data (Cost: Zero). Look at the data from the parallel operation and decide whether to expand or withdraw. Make judgments based on numbers, not intuition.

What is needed for the first step is not a budget of tens of millions of yen or expertise in AI, but simply the decision to “give it a try.”

The technology is ready. Prices have dropped. Now, it’s a matter of whether to act or not. That difference will determine competitiveness three years from now.