Kuken Industries finds spare-part drawings in 4.5 minutes instead of an hour using the OpenAI API

It is a familiar story at any company more than ten years old: the knowledge exists, but it is buried in an archive, and only someone who has worked there for twenty years can find what is needed. Japan's Kuken Industries, one of the country's largest cooling tower manufacturers, tackled this problem using the OpenAI API. This is a customer story published by OpenAI itself, so the figures should be read with the source in mind. But the use case is thoroughly practical and relevant well beyond manufacturing.

The problem

Cooling towers remove heat from circulating water: they cool shopping centers, factories, and equipment in data centers. When a part in one of these installations needs replacing, a modern spare does not always fit the older design. Without the original drawing, engineers sometimes have to design and manufacture a replacement from scratch.

The problem is that Kuken has spent 70 years customizing equipment for individual customers. Specifications depend on the customer and installation date, while a search by model or part name returns several drawings at once. Back when the archive was on paper, a search took half a day. Digitization did not make things any easier: engineers still had to know which name or number to enter and where the file was stored. New hires asked experienced colleagues for help, interrupting their own work.

How the solution works

Kuken worked with Recursive, a Japanese AI startup and official OpenAI partner. Together, they built a search system and a mobile app for engineers working on site. They based the work on the actual archive and the engineers' own requests.

Before choosing a model, the team tested several options on different drawings, comparing how accurately each extracted details and how much it cost. OpenAI's models performed well on both criteria. The article does not specify which models were used or what they were compared against.

In the app, users can enter a drawing or part number, or upload an image or PDF containing a drawing or spare-parts list. A screenshot of the list eliminates the need to type in each number manually: the app reads the numbers and returns possible matches, including drawings of similar designs. The final choice remains with a person, who checks the tower's specifications and records of previous replacements. AI narrows the search here; it does not make the decision.

Results and caveats

According to Kuken, finding the right drawing now takes 4.5 minutes, down from almost an hour. The company expects to reduce the time spent searching by around 90% over a year. That is a forecast, not a confirmed annual result. Engineers can also view drawings directly in the app, reducing printing costs. Experienced employees spend some of the time freed up on mentoring, while new hires handle a wider range of tasks independently.

The next step is a tool for checking drawings prepared by partners. Engineers currently compare these against specifications by eye, looking for mismatched parts or quantities. The prototype is intended to highlight these discrepancies. According to Kuken, around 90% of the issues it identified turned out to be genuine discrepancies. The tool is not yet in production use, so this is a test result.

Lessons to take away

First, the value often lies in routine searches through accumulated documents, where both text and images matter, rather than in a "smart chat." If you have an archive of scans, PDFs, or screenshots, this may be the shortest route to something useful.

Second, choose a model by testing it on your own documents. Kuken compared options for data-extraction accuracy and price using actual drawings, rather than public benchmarks. For documents with complex layouts, this is the only honest way to assess them.

Third, design the input around users' actual habits. Being able to paste in a screenshot of a parts list may seem like a small detail, but it is precisely what removes the chore of manual entry.

Finally, leave verification to a person. The system returns candidates, including similar drawings, and an engineer confirms the match against the specifications. In tasks where a mistake means a defective part, this approach makes more sense than full automation.

The limitations are clear, too. This is a vendor marketing case study, the metrics are reported by the company, and nothing is said about the architecture, data volume, or proportion of unsuccessful searches. Nevertheless, the article is useful as an example of applying multimodal models in an industry with legacy archives.