Japanese equipment maker Kuken and AI developer Recursive announced a drawing-search system on November 4, 2026. Recursive reports a 4.5-minute average lookup in a Fukuoka factory trial—a useful case for maintenance teams trying to navigate old technical records.
An OpenAI customer account published on November 9 describes an archive spanning 70 years. The task is finding records for replacement parts, not creating new designs.
For manufacturers considering a similar project, the key question is whether staff reach the correct record with less total effort—not simply whether AI returns a plausible drawing.
A similar drawing is only a candidate
According to OpenAI, Kuken’s cooling-tower specifications vary by customer and installation date. Engineers can search using numbers, images or PDFs, then check equipment specifications and replacement history before choosing a drawing.
Recursive says its application extracts part numbers from uploaded parts lists and retrieves candidate pages, including near-matching identifiers and different revisions.
That creates a practical tradeoff: a wider search can uncover a hard-to-find record, but similarity cannot establish that a part fits. A useful pilot should count an engineer-confirmed drawing for the specific installation as success. An obsolete revision returned quickly should not earn the same credit.
This is a different assignment from the new-design work explored in RohitAI’s analysis of AI engineering trials: the output here should be a traceable existing record, not a newly generated design.
A search-time result is not a repair-time result
Recursive contrasts the trial average with earlier searches taking tens of minutes to an hour, sometimes half a day. Its roughly 90% annual reduction in search labor is a projection based on prior-year activity, not a measured full-year result.
The published accounts do not give the sample size or clarify whether all preparation and final checking are included. These are supplier and customer reports, not independent performance audits. They do not establish faster completed repairs or a financial return.
Drawing checking needs a separate evaluation
OpenAI also describes a prototype that compares drawings with specifications. It attributes to Kuken a result in which about 90% of raised flags were genuine discrepancies.
That percentage answers “How often was a warning justified?” It does not answer “How many mistakes did the system miss?” Nor is it a drawing-search accuracy score. A separate evaluation needs known discrepancies, including those the tool fails to flag.
Recursive says handwritten notes and layout errors are not fully covered; engineers make the final judgment.
What a useful archive pilot should measure
For a manufacturer evaluating this approach, an informative comparison would use the same authorized lookup tasks with AI and with conventional document search. Include older revisions, poor scans and cases where no matching record exists. Have engineers establish the correct answers before testing.
Correct identification: track wrong revisions, unresolved searches and valid no-match results—not just whether a candidate appears.
Time to a checked record: include preparing the query, reviewing candidates and resolving ambiguity; compare newer and experienced staff.
Extra value from AI: compare against text recognition and part-number indexing, so improvements from organizing the archive are not automatically credited to the model.
Keep original pages available for review and preserve document permissions. A tool that needs constant help from the most experienced engineer may speed up retrieval without making newer staff more independent; record that review effort too.
Recursive announced a target of full deployment at Fukuoka and Chiba by the end of 2026. The announcements do not confirm completion, offer a standard product price or disclose implementation costs. A buyer would need those costs alongside its own lookup results before deciding to expand.