Finding one engineering drawing used to take Kuken's engineers nearly an hour. Back when the archive was on paper, it could take half a day. It now takes 4.5 minutes.
The interesting part is not the AI. It is what the AI was pointed at.
A cooling tower company and its memory
Kuken Industries makes cooling towers and ventilation equipment and holds one of the largest shares of Japan's cooling tower market. OpenAI published the case study this week.
The problem is specific. When a tower needs a replacement part, modern parts often do not fit older installations. Engineers need the original drawing. Over 70 years, Kuken customised equipment for individual customers, so one model name can match many drawings with different specifications.
Newer engineers had to ask senior colleagues where to look and how to read what they found. The knowledge lived in people, and people retire.
Kuken worked with Recursive, a Japanese AI startup, to build a search tool and field app on the OpenAI API. Engineers can type a part number or upload a screenshot of a parts list, and the app returns matching drawings. Kuken projects a 90% reduction in the hours spent searching each year.
The model was the easy decision
One sentence in the case study deserves more attention than the headline number. The team tested several AI models on a range of drawings to compare extraction accuracy and cost, and picked the one that performed best on both.
In other words, the model was a procurement decision. A test, a comparison, a choice. It could be revisited next year if something cheaper or better appears.
The archive could not be swapped. Seventy years of installation decisions, customer variations and repair records belong to Kuken alone. No competitor can buy them, and no model vendor can train them in.
I made a version of this argument in the model is not your moat. Kuken is a clean example of it.
Your company has the same archive
Most marketing and sales leaders read a case study like this and think it does not apply to them. They do not have cooling tower drawings.
They have something equivalent. Ten years of proposals, with notes on which ones won. Campaign results, including the ones nobody wants to talk about.
Sales call recordings. Customer support tickets. Briefs, post-mortems, pricing exceptions, the email thread where someone explained why a client churned.
That is cold data in the most literal sense. It sits in Google Drive, in a CRM nobody updates, in the heads of the three people who have been there longest. When one of them leaves, a piece of the company's judgment leaves with them.
At difrnt, the most valuable internal tool we have built is not clever. It lets anyone search past proposals and campaign results by client type and problem. A new account manager can see how we solved a similar brief three years ago in a few minutes, instead of finding the one colleague who remembers.
The mentoring effect nobody budgets for
The detail I liked most in the Kuken story is what happened to the senior engineers. They used to spend their time guiding juniors through each search. Now they spend some of the saved time mentoring and doing their own engineering work.
That is the real return. Search time is the visible saving. The invisible one is that experienced people stop being a lookup service and go back to the work only they can do.
Kuken is already extending the idea. A prototype checks partner drawings against specifications and flags mismatches. In testing, about 90% of the issues it flagged were genuine discrepancies. That is the next step for any archive: from finding the past to checking the present against it.
Where to start
Pick one archive with a clear, repeated question attached. "Which proposal did we send to a company like this?" "What did we try last time this metric dropped?" A narrow question beats a company-wide knowledge base that tries to answer everything and answers nothing well.
Then decide who can see what. Proposals and pricing exceptions are sensitive, and a search tool that exposes every client's terms to every new hire creates a different problem. Kuken's archive is technical, so access is simpler. Most commercial archives need permissions designed before launch, not after the first awkward discovery.
Clean only what that question needs. Kuken did not reorganise 70 years of records first. It made the existing records searchable by the identifiers engineers already use, part numbers and screenshots.
Measure time to answer before and after. Kuken's numbers are convincing because they are simple: nearly an hour, then 4.5 minutes.
If you want this built without assembling a team yourself, difrnt.ai builds custom AI solutions of exactly this kind, end to end. Whoever builds it, the order matters. Archive first, model second.
Everyone can rent the same model. Nobody can rent your history.