I currently head a business unit — dozens of people, revenue in the hundreds of millions of yen — while also personally delivering software development work in parallel. By any conventional standard this is too many hats. Yet it holds together.

The reason is simple: I moved most of my work as a business unit head onto a suite of automation harnesses I built myself. Python scripts — roughly 10,000 lines across about 30 of them — written in the gaps between work, over about two months.

There is no shortage of abstract debate about whether AI will take managers’ jobs. This article is the opposite: a first-person record of a manager moving his own job to AI, one task at a time, and reporting what moved and what remained to the end.

What actually got automated

Looked at closely, the operations of a business unit form a single pipeline. In my case, from upstream down:

  • Attendance and workload collection — automatically gathering members’ attendance data and aggregating it into per-project workloads
  • Billing and internal paperwork — flowing workload aggregates into billing, and formatting routine approval documents
  • Data entry into management systems — automatically posting aggregates into the PMO’s tracking system
  • Report generation — producing weekly reports, and board- and sales-meeting materials (PPTX/Word), straight from the aggregated data
  • Monitoring — a resident watchdog that alerts Slack on numeric anomalies and stalled items
  • HR evaluation entry — automatically posting the evaluations I write for dozens of people into the HR system

Laid out like this it is almost deflating: the “work” portion of running a business unit turned out to be variations on collect → aggregate → format → transcribe → report. What the human had been doing was moving fixed-format information, at a fixed cadence, to a fixed destination. That is precisely what a harness does best.

How it was built — a harness, not a proper system

The key was not trying to build a respectable internal system.

Each piece is a plain Python script written by Claude Code. When it breaks, I fix it. When the business changes, I rewrite it. There is no requirements document. The only rule I kept was this: every time a task comes up, ask “will this come up again next month?” — and if so, script it immediately.

In about two months the scripts numbered around 30, and at some point the recurring work had become the exception. This approach only works at AI-driven development speed; estimated as traditional “internal tool development,” none of these would have been worth starting.

One more observation: automation had a verification side effect. In the process of translating aggregations into code, ambiguities in metric definitions and structural errors that had been buried in manual spreadsheet work surfaced more than once. That story runs deep, and I plan to write about it separately.

What could not move — judgment, accountability, people

So what remained? It split cleanly into three kinds.

Judgment. Evaluation entry can be automated, but the evaluations themselves I write. Setting prices, declining deals, assigning people — AI can help lay out the options, but choosing stayed my job.

Accountability. Board materials can be auto-generated, but explaining those numbers at the board meeting and committing to hitting them is me. Approvals are the same: the harness formats an approval document; it does not approve. Move that across, and you are no longer a manager.

People. Negotiation, evaluation interviews, the conversation you have when someone seems off. I never considered moving these — if anything, with the busywork gone, there is more time for them.

In other words, what remained is exactly what has always been called “the real job of management.” Which also means: most of a manager’s working hours were being consumed by transport work that was never the real job. In my case, once that transport work vanished, heading the unit and shipping software stopped being mutually exclusive.

Implications for organizations

Generalizing from this experience:

A manager’s workload is a mixture of “judgment, accountability, people” and “collect, aggregate, format, transcribe, report” — and the latter can be almost entirely eliminated with an AI harness. In most organizations, the majority of a manager’s time goes to the latter.

What follows is not “cut the managers.” In organizations that stacked management headcount on top of transport work, one manager with a harness can carry a span of control that used to take several — a redesign of management span, not a reduction exercise. “Player-coach” has long been a euphemism for someone stretched too thin; once the busywork lives in the harness, playing while managing simply works.

There is one precondition: the manager — or someone right next to them — must be able to turn operations into scripts. That, I believe, is why AI-driven development adoption is not just a topic for the engineering department.

I am happy to discuss specifics — architecture, and which tasks to start with — within permissible limits, in a consultation.