I gave an AI full CEO control of a $70 budget and 30 days. Here's what actually broke.

A developer handed Claude Code full CEO control of a ¥10,000 (~$70) autonomous business experiment with a 30-day profit goal, limiting human involvement to identity verification and payments. The agent chose to sell a ¥500 digital download of its own prompts, decision logs, and financial data, but the project hit two stacked Windows Task Scheduler bugs — StartWhenAvailable set to False and an unpathed python executable — that silently blocked posting, and after fixing them the GitHub Traffic API confirmed zero site visits for 14 consecutive days. The AI also flagged that its reused X account carried a real person's prior personal posts and declined to rename the handle on its own, asking the human instead.

I'm running an experiment: Claude Code an AI coding agent is the CEO of a tiny business. Initial capital: ¥10,000 ~$70 . Goal: turn a profit in 30 days, fully autonomously. My job as the human is limited to identity verification and payment — everything else market research, pricing, building the storefront, marketing, customer acquisition is the AI's call. No existing audience, no ad budget, no shortcuts. After evaluating 10 monetization ideas LINE stickers, note.com articles, BOOTH digital templates, Kindle, freelance gigs... , the AI picked the one thing it could run end-to-end via API with zero manual browser work: The product itself is meta: it's a ¥500 digital download containing the actual prompts, decision logs, and financial data from this exact experiment. Day 1 opened with "why hasn't anything posted since yesterday." The Windows Scheduled Task that was supposed to fire at 21:00 had a LastRunTime that looked like it had never run at all. Two separate bugs stacked on top of each other: StartWhenAvailable was False on the trigger — if the machine was asleep at the scheduled time, the task silently gave up instead of catching up. Execute="python" with no path. Inside the Task Scheduler's execution context, that resolved to nothing — ERROR FILE NOT FOUND 2147942402 , failing silently with zero log output. Fix was mundane: absolute path to python.exe , and flip StartWhenAvailable to True . Verified with a manual Start-ScheduledTask and a LastTaskResult of 0 . Once posting was reliable, the real number showed up: 0 site visits, 14 days straight , confirmed via the GitHub Traffic API — not a sampling issue, an actual zero. Stripe's checkout/sessions list was empty too. Not "traffic but no conversion" — no traffic at all. Digging into why turned up something I didn't expect: the "brand new" X account reused for this project had prior history — a real person's personal posts coffee, a dog, daily life in Hokkaido sitting right above the business content in the same timeline, under a username that still didn't match the "AI CEO" branding. The AI flagged this as a trust problem but explicitly declined to rename the handle on its own, treating it as identity-adjacent and outside its authority — and asked instead. Following along and yes, this post is itself part of the low-effort distribution experiment here: https://tomotomosy12030703-lgtm.github.io/ai-shacho-lab/ https://tomotomosy12030703-lgtm.github.io/ai-shacho-lab/