Kuken Industries searches 70 years of drawings in minutes

Kuken Industries, which makes air conditioning and ventilation equipment and holds one of the largest shares of Japan's cooling tower market, built a drawing search tool on the OpenAI API with Japanese startup Recursive. Engineers can enter a part number or upload an image or PDF of a drawing or parts list, and the app returns likely matches. Kuken says finding a drawing now takes 4.5 minutes, down from nearly an hour, and projects a roughly 90% reduction in hours spent searching each year.

This matters because replacement parts made today don't always fit older towers. Without the original drawing, engineers may have to design a part from scratch. Better retrieval means faster, cheaper repairs on equipment that cools buildings, malls and data centers.

We like this story because it's unglamorous. No frontier benchmark, no agent swarm, just a company with a hard archive problem and a tool that makes it smaller. This is the kind of AI adoption we want to see more of: specific, measurable and centered on the people doing the work.

The most interesting part is the knowledge transfer angle. Kuken says newer engineers used to depend on veterans to know where to look and how to read a drawing. Now they can search on their own, and experienced engineers spend some of the time saved mentoring. That's a hopeful pattern. AI here isn't replacing expertise; it's making institutional memory accessible so expertise can spread.

We'd still offer some caution. The numbers are Kuken's own and partly projected, and the 90% figure is an expectation over a year, not a measured result yet. Search returns possible matches, and engineers still verify against tower specifications and past replacement records. That human check is important. A wrong drawing for a replacement part on industrial equipment isn't a typo; it's wasted fabrication or a bad fit.

The next step raises the stakes. Kuken is prototyping a tool to check partner drawings against specifications, and says about 90% of the issues it flagged in testing were genuine discrepancies. That's promising, but it tells us about precision, not about what the tool missed. For review tools, the dangerous errors are the quiet ones that never get flagged. I'd want to know the miss rate before engineers start trusting it to replace eyes-on review.

Our read: this is a good model for small and mid-sized manufacturers. Pick a bounded, painful task, test models on your own documents for accuracy and cost, keep humans in the verification loop, and measure what happens. If it scales across industry, the risk is subtle deskilling, where newer engineers learn to search but not to read. Kuken's emphasis on mentoring suggests they see that risk too.

Kuken's archive spans more than 70 years of customized installations, so specifications vary by customer and date. When the archive was on paper, a search could take half a day. Digitizing helped, but engineers still needed to know which name or number to look up. Kuken and Recursive tested several models on accuracy and cost before choosing OpenAI's.