Title

How Much Does It Cost to Transfer a Veteran’s Knowledge to AI? — Analyzing the Real Costs and Limitations of Knowledge Extraction, Knowledge Graph Creation, and AI Inspection

Main Text

The Question is Simple: “If that person leaves, can this site continue to operate?”

It’s too late to panic after a veteran submits their resignation. For small and medium-sized manufacturers, the loss of a skilled worker’s tacit knowledge is equivalent to not just “losing a person” but to having “know-how evaporate entirely.”

So, can we digitize a veteran’s knowledge using AI? We will examine three recent studies from the perspective of small and medium-sized manufacturing, focusing on the “real costs” and “limitations.” To conclude, there are usable aspects, but it’s not magic. Moreover, if you misallocate your budget, you could end up with a total loss.

—

1. Knowledge Extraction for Troubleshooting Guides — How Many Hours Can AI Reduce from 300 Manual Hours?

The troubleshooting manuals available on-site. Those stacks of paper that say, “If this symptom occurs, check here” in flowchart format. The problem is that digitizing and structuring them is incredibly labor-intensive.

Recent research (arXiv:2601.22754) has validated a method for automatically extracting knowledge from troubleshooting guides using Vision Language Models (VLMs)—AI models that understand both images and text.

What Was Found

• Two types of VLMs were compared. The accuracy was evaluated using standard prompts and extended prompts optimized for troubleshooting-specific layouts.
• There is a trade-off between layout reading accuracy and semantic correctness. Models that can correctly interpret flowchart arrows and branches tend to misinterpret contextual meanings, and vice versa.
• In other words, we are not yet at the stage of “fully automated and perfect.” Human review is essential.

Cost Perspective for Small and Medium Enterprises

When manually structuring a troubleshooting guide, it takes about 30 minutes to 1 hour per page. If we convert that to an hourly wage of approximately 5,000 yen, for a 100-page manual, it would cost 250,000 to 500,000 yen. If this applies to multiple pieces of equipment, the annual cost for knowledge organization alone could reach 1,000,000 to 3,000,000 yen.

Using VLMs, the extraction process itself can be completed in a few hours. The API usage fee would be several thousand to tens of thousands of yen for 100 pages. However, the reality is that reviewing and correcting the output results takes 30-50% of the total time. It’s reasonable to think of it as “reducing 60-70% of manual work.” In other words, a task that used to cost 3,000,000 yen could be reduced to around 1,000,000 yen.

Here’s the critical point: The knowledge written in “paper manuals” is just the tip of the iceberg. The true tacit knowledge of veterans—such as distinguishing abnormalities by sound or judging the condition of materials by touch—is not documented in manuals. VLMs can only extract what is “already documented.” Misunderstanding this could lead to the false impression that “tacit knowledge has been digitized by AI,” when in fact only the surface has been captured.

—

2. Reliability of Manufacturing Knowledge Graphs — The Question of “Is the Data We Input Really Correct?”

Knowledge graphs manage knowledge in a graph structure (a network of nodes and relations). In manufacturing, connecting the relationships between equipment, processes, causes of defects, and countermeasures allows for an immediate response like, “If this defect occurs, the cause is this, and the countermeasure is that.”

However, in small and medium-sized manufacturing, where data sources can be disparate, it often becomes unclear: “When was this data collected, by whom, and under what conditions?”

The reliability infrastructure proposed in arXiv:2608.21418 directly addresses this issue.

Four Reliability Features

1. SHACL Validation — Automatically checks the format and constraints of the data, filtering out garbage data like “a string in the temperature field.”
2. PROV-O Provenance Tracking — Records when, by whom, and from which system the data was entered, clarifying accountability.
3. Dual Time Series Management — Separates management of the “time the fact occurred” and the “time recorded in the database,” maintaining a history even if corrections are made later.
4. Graph-Native Decision Objects — Embeds records of decisions made based on this data within the graph.

Cost Perspective for Small and Medium Enterprises

To be honest, building this reliability infrastructure to full specifications requires an investment of 3,000,000 to 5,000,000 yen, including design, development, and data migration. Additionally, annual operational and maintenance costs would be around 500,000 to 1,000,000 yen.

However, consider the cost of a database without reliability. Responding to issues based on incorrect data can lead to misjudgments and halted production lines. It’s not uncommon for a single misjudgment to result in losses of hundreds of thousands to millions of yen. If this happens three times a year, the investment in reliability infrastructure could be recouped within a year.

That said, it’s unrealistic for a factory with fewer than 30 employees to suddenly implement a full-scale knowledge graph. It’s better to start by establishing a habit of recording “who, when, and under what conditions.” Reliability infrastructure is a topic for later.

—

3. AI Visual Inspection — The Reality That “Implementation Is Not the End”

A study validated a CNN-based visual inspection system for garment sewing lines (arXiv:2608.21426), focusing on detecting defects such as jump stitches.

Key Points of the Results

• High accuracy in detecting defects was achieved for specific fabrics.
• However, the accuracy significantly drops when the color, texture, or pattern of the fabric changes. In some cases, it becomes unusable for fabrics not included in the training data.
• A fundamental issue is the lack of defect samples (lower defect rates result in fewer training data).

Cost Perspective for Small and Medium Enterprises

The system construction costs range from 2,000,000 to 3,000,000 yen. Including hardware such as cameras, lighting, and PCs may increase the total cost.

However, please calculate calmly. If the cost reduction is only several hundred thousand yen per year, it will take 5 to 10 years to recoup the investment. This is not feasible for small and medium enterprises.

AI visual inspection is truly worthwhile under one of the following conditions:

• When the damage from inspection errors is significant. For products where a single complaint can cost 1,000,000 yen, preventing just a few issues a year can cover the costs.
• When it is difficult to secure inspectors. In factories in rural areas where hiring is challenging, this becomes an investment in “business continuity” rather than just cost reduction.

Conversely, if the motivation is merely to “somehow improve efficiency,” it’s better not to implement it.

—

So, What Should We Do?

The three studies reveal the following structure:

Three Things Small and Medium Manufacturers Should Start Doing Today

1. First, take photos of the veteran’s manuals with your smartphone. That alone is enough preparation for VLM. The API cost is just a few thousand yen. Try it out and see for yourself whether it’s usable.

2. Create a system to record “who, when, and under what conditions.” Excel is fine. Before building a knowledge graph, first establish a foundation of data reliability. This can be started at zero cost.

3. Evaluate AI inspection based on “damage amount.” While “automating inspection with AI” sounds appealing, if the return on investment doesn’t add up, it can be postponed. Calculate the damage amount per inspection error and determine how many issues need to be prevented annually to break even.

—

The Real Challenge Lies “After Implementing AI”

We extracted knowledge with AI, created graphs, and automated inspections. But that’s not the end.

The key is to deliver the extracted knowledge in a “usable form” to the site. Create a system to “continuously update” the data entered into the graph. Ensure that the accuracy of the inspection AI is “continuously improved with feedback from the site.”

If this becomes dependent on individuals, we will end up back at the same problem.

The strength of small and medium enterprises is their speed of decision-making. While large corporations are going through bureaucratic processes, they can try, fail, and correct quickly. With the dramatic decrease in AI implementation costs, the ability to cycle through “test small first → measure effects numerically → decide whether to continue or stop” will determine whether the site can continue to operate after the veterans are gone.