1 Hour Reduced to 4.5 Minutes: What Changes with This?

“Where’s that drawing again?”

Have you ever considered how much time has been spent on this one question in the manufacturing industry? Seventy years’ worth of design drawings, stacks of paper, old CAD data, and the knowledge in the minds of retired veterans. For small and medium-sized manufacturers, the act of “searching for past drawings” has been both a task and a symbol of dependency on individuals.

Cooling tower manufacturer Kuken Industries has solved this problem with AI. By utilizing OpenAI’s API, they made a database of 70 years’ worth of drawings searchable, resulting in a reduction of search time from about 1 hour to 4.5 minutes per query, achieving approximately a 93% time savings.

The essence of this story is not merely that “searching has become faster.” It is that knowledge does not disappear even when veterans leave.

Why Older Companies Benefit More from AI

Here, we want to consider the intrinsic value of the fact that there is “70 years’ worth of data.”

Startups that have only been around for a few years do not possess this asset. Design drawings accumulated over 70 years, past failure cases, and custom specifications for each customer—these are all things that cannot be obtained without time. In other words, the length of a company’s history itself becomes a competitive advantage, provided that it is in a “searchable state.”

Until now, only large corporations have been able to meet this condition. Building a database could cost tens of millions of yen, requiring dedicated IT departments and long-term contracts with external vendors. Small and medium-sized enterprises could not afford this.

However, with the use of OpenAI’s API, the situation has changed dramatically. In the case of Kuken Industries, they converted existing drawing data into a vector database and established a system that allows for natural language searches. What once cost tens of millions of yen can now be achieved at a fraction of the cost, even when considering API usage fees and construction costs.

The question is simple: Is the “past asset” lying dormant in your company in a searchable state?

Shipping 20 Times More with the Same Team: What CodeRabbit Proved

Another structurally interesting case is that of CodeRabbit, which provides an AI code review tool.

They utilized OpenAI’s model to increase the number of software shipments by 20 times with the same engineering team. They did not hire more people. By having AI take on “necessary but non-creative tasks” such as code reviews, bug detection, and documentation generation, engineers were able to focus on their core development work.

Translating this into the manufacturing context, it looks like this:

• Engineers who used to spend 1 hour searching for drawings can now do it in 4.5 minutes.
• The 55 minutes saved can now be spent on new designs or improvement proposals.
• Without increasing headcount, both the quality and quantity of output have increased.

For small and medium-sized enterprises, “increasing productivity without hiring more people” is a survival strategy, especially in regions facing labor shortages. There is no need to hire 100 people like large corporations. If the productivity of the existing five can be tripled, it is equivalent to the work of 15 people. AI enables that leverage.

Let’s Talk About Costs

“So, how much does it cost?”—this is always the first question from small and medium-sized business owners.

First, the API usage fees. For OpenAI’s GPT-4o, the cost is $2.50 per million tokens for input and $10 per million tokens for output (as of June 2025). When converted to the cost per drawing search query, it amounts to a few yen to several tens of yen. Even with 1,000 searches a month, the API costs would be in the range of a few thousand to tens of thousands of yen.

Moreover, it is noteworthy that the trend in model costs is decreasing. Meta’s Muse Ember, which has a 1 million token context window, is priced at about $0.14 (approximately 21 yen) per task. For applications that require processing large amounts of drawing data at once, the emergence of such models will further drive down costs.

On the other hand, what about construction costs? If you build a RAG (Retrieval-Augmented Generation) system in-house, outsourcing it typically costs several hundred thousand to over a million yen. This is not the world where you first spend three months on requirements definition, six months on development, and incur costs of 5 million yen, as was the case in past system development.

Calculating the return on investment is also straightforward.

Assuming an engineer’s hourly wage is 3,000 yen. With three searches for drawings a day, each saving 55 minutes, that results in about 2.75 hours saved per day, or 55 hours a month over 20 working days. This equates to a monthly labor cost reduction of 165,000 yen. If there are three engineers, that amounts to 500,000 yen a month, or 6 million yen a year. If the construction cost is 1 million yen, it can be recouped in just two months.

This calculation is conservative. In reality, if you include opportunity losses such as “couldn’t find it even after searching” or “asked the veteran but they were unavailable,” the effects would be even greater.

Three Conditions for Small and Medium-Sized Enterprises to Turn the Tables with AI

From the cases of Kuken Industries and CodeRabbit, we can identify the conditions for small and medium-sized enterprises to achieve results with AI.

1. Having “Accumulated Assets”

Seventy years of drawings, past estimation data, and customer interaction history. Data that has accumulated over a long period is the foundation for AI utilization. New companies cannot replicate this.

2. Clear Identification of “Bottlenecks of Dependency”

Companies where “I can only find out by asking that person” will see a greater effect from AI implementation. Conversely, if the tasks are not dependent on individuals, introducing AI will not result in dramatic changes. AI is effective where there is pain.

3. Willingness to “Start Small”

It is better to test one business case for two weeks than to plan a company-wide implementation and deliberate for six months. Kuken Industries also started with a single use case: drawing searches. A successful experience can change the atmosphere within the company.

“Is This Enough?” as the Starting Point

Finally, I want to pose one question.

If a veteran employee in your company were to leave next month, what would come to a halt? The list of those “things that would stop” is your priority list for AI implementation.

In an era where you can search 70 years’ worth of drawings in 4.5 minutes, will you continue to rely on “I can ask that person”? Even though there is a system that can produce 20 times the output with the same team, will you keep lamenting that “we don’t have enough people”?

Technology does not wait. However, small and medium-sized enterprises have their own ways to fight. There is no need to launch a company-wide DX project like large corporations. Just try one thing tomorrow.

Whether it’s drawing searches, estimation creation, or automating meeting minutes—anything will do. Start by handing over one task that is dependent on individuals to AI. That will be the first step towards turning the tables.