Vol.070 — AI Native, the Japan Edition (Recruitment System): “Delete It and It Gets Cheaper. But You Can’t Delete It” — Four Days
Date: 2026-09-19 | Newsletter
Key Summary
Vol.070 is the sixth instalment of “AI Native, the Japan Edition”, and this time the subject is recruitment — the first four days of it. It did not start with AI. It started with a price-rise notice. The company runs hiring on a cloud applicant-tracking system: job postings go up, around ten recruitment agencies upload candidates’ CVs into it, and the HR officer schedules interviews from there, with agency correspondence also running through the system’s own mail. It costs roughly ¥1 million a year, and the price tier is set by how much candidate data is stored. More applications pushed the company over a tier boundary, and the bill was going up by about 40 percent. The obvious fix — delete old data and drop a tier — is the one thing they cannot do, because that stored data is what everyone, agencies included, uses to check whether an applicant has applied before. Delete it and it gets cheaper; delete it and the work stops. Zenta’s point is that nobody explains this at signing: the more data you hand over, the more convenient and reassuring the service becomes, and that reassurance quietly turns into the justification for the price. Lock-in arrives not as a feature but as the shape of where your data lives. The leads this time are not a new hire but two people: a DX consultant whose day job is digital-transformation work for outside clients, assessed at L2 on the firm’s AI-use ladder and with no experience building a skill, and the HR officer who runs hiring every day — one person who knows the work, one who knows how systems are put together. Day one built nothing; the two wrote out a month of the hiring process end to end, and the stuck points were plain: every CV opened one by one, ten agencies handled one by one, every Zoom link issued by hand. The pain centred on screening and interview scheduling. Day two listed every function of the current system across seven groups and marked each keep, drop or partial. “Data volume and limit management” — the very entry point of the price rise — got a cross: bring the system in-house and that management task disappears altogether. A new pain surfaced too: interview feedback taken verbally and retyped later, marked “NG” in the notes. The same day, the DX consultant built a skill for the first time — CV screening that reads a hundred sample PDFs and filters by education and average tenure, cutting screening effort by 40 to 70 percent — then a skill that drafts new job postings from old ones (“that was easy”) and a prototype that issues Zoom links. Someone who had never built a skill built three the next day. On day three Zenta stopped setting homework, because a person who can already see what to build only slows down when handed assignments; he answered questions instead. There was one: where the data should live. They tried OneDrive, SharePoint and Google Sheets and chose Sheets, because everyone, inside and outside the company, could use it the same way. The best moment that day was not technical. Watching the build, the HR officer proposed a feature nobody had asked for — a view of each agency’s track record. An agency that sends many CVs looks busy, but a stack of poor-fit CVs only adds checking work and fills no roles. Volume is not value. From that point the system stopped being something built for HR and became something HR was building. On day four the consultant added the remaining functions alone. The screens are three, not one all-purpose view: one for HR, one for the hiring managers and the CEO, one for the external agencies — the same data, but each role needs to see different things and must not see others. Then came the hardest question of the four days: “So how do we make this work in production?” Building, Zenta writes, is no longer the hard part. Deployment is — a bundle of questions about who logs in, what each role sees and does not see, where the data sits, and when and how outside agencies are brought in. Here the functions, skills, connections and screens go into one GitHub repository, connected to Render, with a login for each role — the same approach used to run the AI webinar host. It goes live in October, with agencies briefed from mid-October. If a DX consultant is to become an AI consultant, Zenta argues, the line runs through three steps: break the work down and sort what to keep and drop; build the skills and screens yourself; and deploy them as something someone uses every day. The first two are now within anyone’s reach; only the third remains. He closes as an owner: his long-standing advice not to marry any one LLM — design so you can always switch — applies just as much to cloud subscriptions. Whether ¥1 million a year is too much is not the question. The problem is being unable to refuse a price rise, and you cannot refuse because your data is on their side and your work depends on it. Build the ability to leave, and staying becomes a choice you can make with confidence. What the AI Native project aims for, more often than not, is exactly this: bringing what you pay for every month back to your own side — and making it something the people who know the work can build themselves. Do it simple — your usual work, in your usual words.