Kuken’s AI drawing search cuts lookup time to 4.5 minutes

By James Calloway

Kuken Industries says its field app, built with Recursive through the OpenAI application programming interface (API), finds a replacement-part drawing in 4.5 minutes, down from nearly an hour. Kuken expects the tool to cut hours spent searching for drawings by about 90% over a year.

Kuken makes air-conditioning and ventilation equipment for large buildings, shopping centers and factories. The company says it holds one of the largest shares of Japan’s cooling-tower market. According to Kuken, cooling towers release heat from circulating water into the air, cooling large buildings, malls and equipment in data centers.

Kuken framed the project as preserving practical knowledge, not merely speeding up retrieval:

“I believe OpenAI’s technology can help manufacturers like us pass skills on to the next generation. That’s long been a challenge for us.”

The 4.5-minute result and 90% projection come from Kuken. Kuken’s account names no specific model, archive count, independent test or full cost comparison. It establishes a specific workflow saving, not a verified return on artificial intelligence (AI) investment.

Why does the cooling-tower archive matter?

When a cooling tower needs a replacement part, engineers search drawings spanning 70 years of installations. Newer engineers previously asked experienced colleagues where to look and how to read the records. That dependency could slow both maintenance and the transfer of practical judgment.

Kuken wanted engineers at every experience level to search independently, learn from earlier decisions and spend more time on the repair. The field app is meant to turn an archive of installation records into a working reference for daily maintenance. Takeaway: The project targets a dependency on veteran staff, rather than a generic chatbot use case.

What made the archive so difficult?

Kuken adapted its equipment to individual customers over more than 70 years. Specifications therefore vary by customer and installation date, and a search by model or part name can return several drawings. Paper searches could take half a day.

Digitization removed some friction, but engineers still needed the correct name or number and the file’s location. Newer staff often spent time reconstructing those details. A vague part search also risked landing on the wrong installation record.

Parts made today do not always fit older cooling towers, and each drawing records what was actually installed. Without the original drawing, Kuken says engineers may have to design and manufacture a replacement from scratch. Takeaway: Missing or incorrect records can create extra engineering and manufacturing work.

How does the OpenAI-backed app work?

Kuken connected with Recursive, a Japanese AI startup and an official OpenAI partner. The companies built the search system and its field app together, drawing on Kuken’s archive and input from its engineers.

They tested several AI models across a range of drawings, comparing how accurately each extracted details and what each cost. Kuken says OpenAI’s models performed well on both measures. The OpenAI API also gave the team flexibility to build search directly into the app.

Conversations with field engineers shaped the interface. Engineers can enter a drawing or part number. They can also upload an image or PDF (Portable Document Format) document of a drawing or parts list.

Pasting a screenshot of a parts list lets the app read the numbers without manual entry. The system returns possible matches, including drawings for similar designs. It does not present one guaranteed answer.

Kuken engineers still verify each result by comparing the match with the cooling tower’s specifications and records of past replacements. The app accelerates candidate search, while staff retain responsibility for confirming the correct document. Takeaway: The system shortens discovery, but engineers still make the final engineering judgment.

What has changed for engineers?

Kuken reports several practical effects from using and testing the tool. Its measured and projected results include:

• A projected reduction of about 90% in the hours engineers spend searching for drawings each year.
• More time for experienced engineers to mentor newer colleagues.
• Lower printing costs because engineers can review drawings in the app.

The company says engineers now spend less time locating a drawing and more time studying it. Both newer engineers and veterans report having room to consider design improvements and explore ideas they previously had little time to pursue.

Newer engineers can handle a wider range of jobs independently. Veteran colleagues can spend more time teaching and taking on engineering work of their own.

Kuken quoted an engineer:

“People now have time to do work they’ve wanted to try and take on new challenges. That’s been a real source of motivation.” Kuken did not identify the speaker.

Kuken provides no measured before-and-after figures for mentoring, printing or design work. Kuken also does not publish the number of drawings tested or the model-by-model accuracy and cost results. Takeaway: The evidence supports a plausible time saving, while the organizational gains remain largely self-reported.

Can the system check partner drawings?

Kuken is developing a separate tool to check drawings prepared by partners. Engineers now compare each partner drawing with specifications by eye, looking for details or quantities that do not match. The proposed system would flag those differences so engineers could review drawings faster and catch mistakes they might otherwise miss.

In prototype testing, about 90% of the issues the AI flagged were genuine discrepancies, according to Kuken. Kuken expects substantial reductions in checking time if the tool enters production, but provides no drawing count, starting time or figures for missed discrepancies. Takeaway: The checker’s 90% result is encouraging, but its test data cannot establish production reliability.

What does this prove?

Each drawing preserves an engineer’s decisions for a particular installation. Kuken aims to make those decisions easier to find and understand so it can pass experience to the next generation of engineers.

The practical case is narrower than the company’s aspirations. This is an archive-search and document-check system built around proprietary records, not a claim that AI can automate cooling-tower engineering. Its value rests on whether engineers trust its matches, whether incorrect flags remain rare and whether the projected savings persist after deployment.

Kuken has not set out a production timetable for the partner-checking tool in its account. The next test is whether wider use preserves the reported 4.5-minute lookup and 90% reduction in search time. Takeaway: Wider deployment will show whether a reported 4.5-minute search can become a lasting way to transfer engineering knowledge.