Kuken Industries uses the OpenAI API to search 70 years of its cooling tower drawings. According to OpenAI’s write-up, published on 9 November 2026, an engineer can now find the drawing for a repair in 4.5 minutes. It used to take nearly an hour.
I couldn’t load the full article, so I’m working from OpenAI’s summary. I won’t guess at how Kuken built it or which parts of the setup did the work.
The excitement is about the model
The easy reading of this story is “AI cut a one-hour search to under five minutes.” That gives the credit to the model. It also makes the lesson sound like “buy the API and wait for the results.”
I’m wary of that reading. My first question about any AI story is whether there was a real need. Here there clearly was. An engineer needs a drawing to do a repair. Nearly an hour of looking for it, every time, is a real cost in a real job. Nobody has to be talked into that problem.
The second question is what actually changed. The summary names one new thing, which is a model that can search. It names one old thing, which is 70 years of cooling tower drawings. Kuken has held that archive for decades, and no AI company supplied it.
What didn’t change
The drawings didn’t change. Seven decades of work on cooling towers still sit behind that search. A competitor with the same API access but a thin or scattered archive would not get a 4.5 minute answer. They would get a fast search over very little.
I hold one position above most others. Context is the product. The quality of what goes in decides the quality of what comes out. Kuken’s result is a good example, because the input here is an asset the business spent 70 years building. The model is the part anyone can rent next week.
The people in the story haven’t changed either. Engineers still decide which drawing matches the repair, and they are still responsible for the work. The search only got them to the right page sooner.
What I’d ask an operator
Don’t start by asking which AI tool to buy. Ask which records your staff keep digging through.
Most businesses I can think of have some version of this:
- A repair shop with years of job cards and photos of past faults.
- A clinic with old treatment notes and referral letters.
- A service business with past quotes, site visit reports and email threads.
- A marketing team with campaign briefs, results and approved copy spread across drives.
Someone on your team goes looking for these every week. They ask a colleague who has been there longest, or they scroll through folders and hope. That repeated digging is the signal. It shows a need that already exists, and it shows you own the raw material.
Putting AI on that search comes before anything flashier. A chatbot with a new personality, or an agent that “runs your business”, has to earn its place against a plain need like this one. Finding the right record fast is a need nobody argues about.
The unglamorous work is the real work
An archive only helps if it can be read. Drawings that are scanned badly, named inconsistently or filed in the wrong place will limit what any model can find. I can’t tell from the summary how much cleanup Kuken did. For your own records, I’d expect that part to take real effort, and I’d plan for it.
This is the same reason I say the technology is the easy part. Getting people to put records somewhere findable, and keep doing it, is harder than calling an API. It’s a habit change for the team, and it needs someone to own it.
You also have to check the answers. When the search returns a drawing for a repair, a person should confirm it is the right one before anyone acts on it. Where a wrong answer is costly, a human checks. That doesn’t change because the search got faster.
Where this shows up in my own work
My GEO work starts from the same place. Before anyone asks what to publish, I ask what the business already knows that it hasn’t written down clearly. AI search engines quote specific facts. A business that has its details in order is easier for them to name than one that hides them.
What I’d do this week
- Ask your team what they spend time hunting for. Write down the three answers you hear most.
- Pick the one with a clear cost, such as an hour lost per job, and a record set you already own.
- Check the state of those records. Are they complete, readable and in one place?
- Fix the worst gaps first, then point an AI search at them.
- Time the search before and after, the way Kuken did, so you know what changed.
- Keep a person in the loop to confirm each answer before it is used.
Source: OpenAI News: Kuken Industries searches 70 years of drawings in minutes