A small-factory executive builds his own apps
【月300時間以上削減】町工場の専務が生成AIでアプリを自作！製造業の現場はどう変わった？
B · Useful, limited Interview on a manufacturing-AI training channel with the managing director of a small Japanese parts maker who built his own shop-floor apps; concrete and first-hand, but self-reported with no measured outcomes.
Summary based on the full transcript·Original: about 18 minutes to listen
In short
An 18-minute Japanese interview (Genba AI channel, 26 October 2026) with Yuta Okada, managing director of Okada Kenma, a parts maker in Ishikawa that makes bushings, pins and hydraulic-cylinder parts for construction machinery. With no prior web-app knowledge he built the company's internal apps with generative AI: inspection records, procedures, defect costs, machine utilisation and attendance. The channel title claims "300+ hours a month" saved; in the interview he says the saving cannot be measured and is not his goal. His thesis is that every firm gets the same models, so the lasting difference is accumulated, linked company data and context.
Key numbers
- about 90%share of the firm's output that is parts for construction machinery
- close to 300work procedures registered in the AI-scored procedure app in one year
- 70 pointsAI score threshold that procedures are improved toward, one by one
- about two eveningstime he spent at home building an attendance and leave system to replace an external SaaS
As reported by the source.
Japanese·Video·opens youtube.com
What it reports
- He moved from no-code tools to AI-built web apps because no-code could not handle the complex screen layouts the shop floor needed. All shop-floor staff carry iPads.
- A portal ("Okada Board") holds the apps. The most used is an inspection-record app: measurements and times are entered at the machine, drawings are shown, and defect reports are linked to the records.
- A procedure app ("Knowledge Board"): staff register work procedures from iPads, an AI scores each against set criteria, and the company improves them toward a 70-point bar. Points convert to cash rewards. Close to 300 procedures were collected in a year.
- Defect types and their costs are aggregated for quality meetings. Signal-tower data from machine tools feeds utilisation charts; acting on that data is the next step.
- He built an attendance and leave system in about two evenings to replace an external SaaS, because labour data split across SaaS products cannot be used in real time and packages did not fit the company's leave and overtime rules.
- On savings: if each person saved 30 minutes, that would add up to hundreds of hours a month, but it cannot be measured, and time saving is "not the purpose at all".
- Waiting for a quantified ROI before starting means falling behind. In small firms the executive should build, because shop-level or DX-office builds tend to optimise locally.
- Security: give the AI common attack classes as build-time context, then run a separate review agent. He builds with Claude and has Codex do the security review as a third party.
How they got it
A recorded interview at the plant, with screen demos of the apps (demo data). The interviewer's questions are not audible. The digest is based on cleaned auto-captions; the channel closes with an ad for its own training and a free "time-saving diagnosis".
Limits
- Self-reported by the person who built the apps; no usage, defect or time data.
- The "300 hours a month" in the video title is the channel's framing, not a figure the guest gives.
- Auto-captions garble names and some numbers (the procedure count is captioned as "300 yen").
- The channel sells AI training to manufacturers, so the framing favours adoption.
- One small firm; the executive-as-builder model depends on one person's skill and time.
Why it matters for companies
The case shows a small manufacturer replacing SaaS and spreadsheets with owner-built apps whose main value, in his account, is a single linked data base rather than saved hours. That argument applies to larger firms too: fragmented tools fragment the context any model can use. The practical points are copyable: start with low-risk forms (lunch orders, daily equipment checks), score documents against explicit criteria, and use a second model for security review. The risk is key-person dependence: one executive now owns the code, the data model and the security checks.
Key passages in translation
Japanese original
コンテキストをどれだけしっかりと貯めてきているかっていうところが今後の決定的な差になると思ってんですけど
I think how thoroughly you have accumulated context will be the decisive difference from now on.
Japanese original
時間は間違いなく作業はされてますけど、測り用がないって感じですし
Time is definitely being saved, but there is no way to measure it.
Japanese original
個人的にAIを導入するのに、時間削減っていうのは全然目的にしてないことなので
Personally, saving time is not at all the purpose of introducing AI.
Summary based on the full transcript. Read the original for the full argument and data.Watch the original on youtube.com