Kuken Industries: AI-assisted engineering drawing search

What the source reports

Kuken holds a 70-year archive of cooling-tower installation drawings. Customer-specific changes and installation dates make part-name or model searches ambiguous, and engineers historically relied on experienced colleagues to locate the right record. Kuken and Japanese AI startup Recursive built an OpenAI API-powered search app with input from Kuken engineers. Users can enter a drawing or part number, or upload an image or PDF; the app reads listed part numbers and returns possible drawings, including similar designs. Engineers check equipment specifications and past replacement records to identify the relevant drawing. OpenAI’s account says Kuken engineers report finding a drawing in 4.5 minutes rather than nearly an hour, while the stated 90% annual reduction in search hours is a projection. The account separately describes a drawing-checking tool in development; in prototype tests, Kuken says about 90% of flagged issues were genuine discrepancies. That prototype result is not a production result for the search app.

Reported technology: OpenAI API-powered search app, built with Recursive. Scope and outcomes are source-reported.

What the enterprise environment would need

The following is a proposed project design for an enterprise doing similar work. Published details are marked; the remaining systems, data and scenarios would be agreed for your project.

Records and access

Proposed: load a bounded set of fictional cooling-tower installations, dated drawing revisions, replacement histories and part references into a read-only practice workspace.

People and search

Proposed: let an engineer search by part number or drawing image; return a ranked candidate list with source drawing, installation and revision links. The agent does not choose a compatible part.

Independent check

Proposed: compare retrieved candidates against a separately prepared engineer-verified mapping for the exact installation, revision and replacement history.

Owner decision and limits

Proposed: an accountable engineer accepts or rejects each match. No authoritative drawing is changed and no repair, purchase or manufacturing action is triggered.

Ambiguity and recovery

Proposed: include duplicate part numbers across installations, similar equipment, superseded drawings, low-quality scans and missing metadata; hold uncertain matches for owner review and preserve the query evidence.

A proposed simulation

An illustrative practice design derived from this work, not an executed assessment.

Proposed: retrieve the correct historical installation drawing from a bounded fictional equipment archive and show why each candidate matches.

Review gate: An engineer verifies the exact installation, drawing revision and replacement history against an independent reference before accepting a candidate.

Challenge the result: Two installations share a similar part number but use different revisions. The system must show both source records, avoid asserting a match, and return the case to engineer review.

What needs checking

• Source boundary: OpenAI reports Kuken and Recursive built a search app using Kuken’s drawing archive and engineer input; engineers can search by identifiers or image/PDF and check possible matches against equipment specifications and replacement records.
• Outcome boundary: the 4.5-minute comparison and 90% annual-search reduction are Kuken/OpenAI-account figures; the annual reduction is a projection and methods/sample details are not published.
• Separate prototype: drawing comparison is described as under development; the roughly 90% genuine-discrepancy figure is from prototype testing and does not establish production performance.
• Proposed: test exact-installation retrieval against an engineer-verified reference, with similar installations, ambiguous part numbers, stale revisions and unreadable scans. Designed, not executed.

Review environment fidelity, human capability and agent performance separately. The enterprise’s accountable owners decide acceptance of the shared workflow and any operating use.

How an enterprise could build similar work

A proposed build order for an enterprise doing similar work, written by Nuvepro from the published account and our interpretation of the environment and checks above. It is not how Kuken Industries built or operates this work.

1. Select a bounded drawing archiveUse fictional installation metadata, part references, dated revisions and replacement history.
2. Prepare known-match queriesCreate part-number and image/PDF searches with an engineer-verified expected drawing for each installation.
3. Return source-linked candidatesShow candidate drawings, installation identifiers and revisions; preserve ambiguity instead of selecting a compatible part.
4. Check retrieval independentlyCompare the exact installation and revision with a separate engineering reference, including duplicate identifiers and poor scans.
5. Engineer accepts or returns the resultRecord the owner decision and re-run the corrected query where evidence was incomplete. Keep the workspace read-only.

Your systems, data, rules and accountable owners decide the real project. Agree acceptance criteria, escalation owners and operating responsibility with the people accountable for the work before anything runs on live records. Treat the first run as a rehearsal on test records: compare the draft output with an independently prepared answer, and record what the people reviewing it change.

Questions about this example

How does Kuken search its drawing archive?

OpenAI says engineers can enter a drawing or part number, or upload an image or PDF. The app returns possible matches; engineers check equipment specifications and past replacement records.

Did Kuken measure a 90% time saving?

Kuken reports finding a drawing in 4.5 minutes instead of nearly an hour. The 90% reduction in annual search hours is a projection; the article gives no independent evaluation or measurement method.

Is drawing review in production?

No production use is reported for the separate drawing-comparison tool. OpenAI describes it as under development and cites Kuken prototype testing.

What would a similar business environment need?

A proposed, unexecuted exercise would use fictional installations and drawings, source-linked candidate retrieval, an independent engineer-verified reference and owner review.

Shape this around your enterprise

Bring a similar workflow and the people responsible for it. Together, define the records and systems to use, the result your team will build or configure, and how its owners will check it. Nuvepro scopes the simulation, environment and preparation. Agree any existing-agent evaluation, implementation assistance and production responsibilities separately.

Sources and limitations

Curation review: 10 November 2026. This is the date this source example was reviewed, separate from source publication dates.

• This is an OpenAI customer story, not an independent engineering or productivity evaluation.
• Kuken’s reported timing comparison and projected annual time reduction lack a published sample, measurement protocol or independent verification.
• The source does not provide the search model configuration, archive coverage, ranking quality or error rates.
• The drawing-review prototype and its reported test result are a separate development-stage effort, not production evidence for the search tool.
• The proposed practice environment is fictional, unexecuted and not a reproduction of Kuken’s systems.

1. Kuken Industries searches 70 years of drawings in minutesSource publication: 2026-11-09

More software engineering examples

• AI chip verification workflows: OpenAI exampleUnnamed semiconductor company
• AI energy model customization: Exowatt exampleExowatt
• AI agent evaluation task design: BrendanFoody accountBrendanFoody evaluation-project account (organization unnamed)
• AI content platform migration: Wimbledon exampleAll England Lawn Tennis and Croquet Club (Wimbledon)

All 14 software engineering examples