0 Employees, 60 Clients — A Japanese Tax Accountant's AI Automation

Japanese CPA and tax accountant Kento Hatakeyama recently published a long post on X that garnered over 2 million views.

His firm has zero employees, yet serves 60 client companies. By industry standards, this workload requires 6 staff members and over 30 million yen (~$200K) in annual personnel costs. Using Claude Code, he built a comprehensive automation system that saves over 24 hours per month — roughly 300 hours per year.

It’s an impressive case study. But what’s even more interesting is that the system he spent months building from scratch maps almost perfectly to features Tetora already provides out of the box.

What He Built

1. Nightly Auto-Journaling — Two-Stage AI Classification

A scheduled task runs at 21:00 every night, fetching unprocessed transaction records from the freee API for all 60 companies. Account classification uses a two-stage approach:

• Stage 1: Keyword dictionary matching — 14 account categories, 100+ keywords each. Fast and zero API cost
• Stage 2: Claude API fallback — Only unmatched transactions go to AI. A confidence threshold filters low-confidence results for human review

A mature design: don’t waste AI on what rules can handle.

2. Five Services Connected via MCP

freee (accounting), Gmail, Google Calendar, Notion, Slack — all connected through MCP with Claude Code as the central orchestrator.

3. Skills: Accumulating “Business Patterns”

Repetitive business patterns defined as Claude Code Skills:

/freee-check    → Check unprocessed entries
/mtg-followup   → Post-meeting notes & action items
/ipo-analysis   → Analyze newly listed companies

The more you use them, the more Skills accumulate, the faster you work.

4. CLAUDE.md as the Business Manual

Journal entry rules, tax classifications, security policies, output paths, decision boundaries — all written into CLAUDE.md. A senior employee’s SOP in a format AI can read.

5. Automated Task Logging

After each task: estimated manual time, actual AI time, and time saved are automatically logged. Monthly summaries are auto-generated.

6. Multi-Company Data Isolation

Data for all 60 companies is strictly isolated by company_id. Transaction details are logged only in company-specific log files.

His Custom Build vs Tetora Out-of-the-Box

The key difference: Hatakeyama spent months building schedulers, logging, and Skill management infrastructure from scratch. With Tetora, you skip all of that and start writing business logic immediately.

Why “Field Knowledge” Is the Real Core

What matters is knowing “what should be automated.” The only person who can make that judgment is you — the one working on the ground every day. — Kento Hatakeyama

Hatakeyama emphasizes at the end of his post: it’s precisely because he’s not an engineer that Claude Code works so well. Tax accountants know the “business patterns” — journal entry rules, filing procedures, month-end checkpoints. This practical knowledge, accumulated over a decade, is something AI cannot generate on its own.

This is exactly Tetora’s design philosophy.

Tetora doesn’t require programming skills. What you need is domain knowledge:

• You know which processes can be standardized → Write them in Workflow YAML
• You know where the decision boundaries are → Write them in SOUL.md
• You know the repeating work patterns → Define them as Skills
• You know what to automate and what needs human eyes → Set the permission level

Engineers use AI to build “technically impressive things.” Professionals use AI to build “practically correct things.” Tetora makes the latter possible — without building infrastructure from scratch.