Learning from the KAUCHE case study: Operational design for auto-merging 83% of PRs solely with AI using Claude Code Action

The impact of "83% merged without human review"

On April 7, 2026, an engineer at Kauche published an article on Zenn titled "Enabling 83% of All PRs to Be Merged with AI Review Alone," which garnered widespread attention on Zenn Trending.

In typical development organizations, the golden rule is that PRs must pass human review before being merged. This initiative directly overturns that conventional wisdom, and the impact comes from the fact that the figure is not "around 10%" but rather "83%."

Drawing from Kauche's case study, this article outlines the mechanics of automated review operations using Claude Code Action and highlights the essential points to keep in mind when adopting it internally.

Overall picture of the mechanism

Core components

Kauche's workflow consists of three simple components.

1. Claude Code Action — An AI review agent running on GitHub Actions
2. GitHub Actions workflow — Triggers reviews upon PR creation
3. Automated Approve → auto-merge pipeline — Merges directly once the review passes

When a PR is created, Claude Code Action automatically reviews the diff and automatically approves it if there are no issues. Once CI passes, it is merged directly into the main branch via GitHub's auto-merge feature.

Important prerequisites

For this operational approach to succeed, several prerequisites must be met.

In Kauche's case, a distinctive feature is that they automatically update review rules every night. When a mistake overlooked by AI review is discovered, it is reflected in the next day's review prompt, creating a system whose accuracy improves as it runs.

Concrete implementation examples

Example GitHub Actions workflow

Below is a minimal configuration example for introducing Claude Code Action based on Kauche's case study (this may differ from Kauche's actual setup).

# .github/workflows/ai-review.yml
name: AI Code Review

on:
  pull_request:
    types: [opened, synchronize]

jobs:
  claude-review:
    runs-on: ubuntu-latest
    permissions:
      contents: read
      pull-requests: write
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0

      - name: Run Claude Code Action
        uses: anthropics/claude-code-action@v1
        with:
          anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
          review_prompt_file: .github/ai-review-rules.md
          auto_approve_low_risk: true

Providing review context

Passing repository rules and coding standards to Claude Code Action is critical. Prepare a Markdown file such as the following.

<!-- .github/ai-review-rules.md -->
# レビュールール

## 必ずチェックすること
- シークレットがハードコードされていないか
- SQLインジェクションのリスクはないか
- フロントエンドで未エスケープの出力がないか
- 新規依存の追加には理由コメントがあるか

## 自動Approveしてよい条件
- ドキュメント・README・コメントのみの変更
- devDependenciesのマイナーバージョンアップ
- CIワークフローの改善PR（テスト系のみ）

## 必ず人間レビューに回すもの
- データベーススキーマの変更
- 認証・認可まわりの変更
- 外部API呼び出しの新規追加

Enabling auto-merge

On the GitHub side, enable the auto-merge feature in repository settings.

# GitHub CLIからの設定例
gh repo edit --enable-auto-merge

By configuring gh pr merge --auto to be called by colleagues or CI upon PR creation, auto-merge runs as soon as CI passes.

Steps for internal adoption

Recommended phased rollout

There is no need to aim for 83% right from the start. Expanding the scope step by step is the practical approach.

Establish an observation period of 1 to 2 weeks in each phase to monitor whether overlooked bugs occur.

Required internal consensus

Before launching an auto-merge workflow, the following topics should be agreed upon internally.

1. Locus of responsibility: Who is accountable for bugs overlooked by AI review?
2. Emergency rollback procedure: The incident response workflow when problems are discovered
3. Retention period for review logs: Whether to retain AI review results as audit logs
4. PR categories requiring mandatory human review: Handling of security, database modifications, authentication, and similar changes

Drawing from perspectives in the Web Security Fundamentals Guide, it is safest to operate with mandatory human review especially for security-related PRs.

Cost estimation

Claude Code Action API usage fees run approximately a few cents to several tens of cents per PR (depending on PR size). Even for a team with 1,000 PRs per month, costs frequently fall within several tens to several hundreds of dollars monthly, delivering an overwhelmingly positive return on investment compared to reducing engineer review hours.

However, monitoring AI review costs is necessary. Because large PRs or tasks with massive diffs can consume tokens rapidly, cost optimization perspectives like those introduced in Leveraging New Gemini API Tiers are also beneficial.

Caveats and pitfalls

1. Risk of accumulating "unseen bugs"

AI reviews can produce subtle mistakes that slip through due to pattern-matching limitations. Establish monthly codebase audits (sampling reviews conducted by humans) to uncover patterns that AI tends to overlook.

2. Prompt injection

In theory, attacks that attempt to manipulate AI review outcomes through malicious commit messages or code comments are possible. Particular caution is required for PRs from external contributors, and the baseline rule should be to exclude external PRs from auto-approval.

3. Review quality obsolescence

AI review quality varies significantly depending on prompt tuning. Kauche adopted "nightly automatic updates" precisely to prevent this obsolescence. Run weekly to biweekly prompt review cycles internally as well.

4. Degradation of human review skills

Leaning too heavily into full automation carries the risk that developer review skills will decline. Organizations should also consider intentionally retaining human reviews as training grounds for junior engineers.

Comparison with similar approaches

Several other tools and services are also participating in the shift toward automating code reviews with AI.

What makes Kauche's case study unique is that it goes all the way to automated Approve + auto-merge. Many other tools stop at "review assistance," maintaining the boundary where final decisions are left to humans.

For an overall comparison of AI coding tools, please also refer to the In-Depth Comparison: Cursor 3 vs Claude Code.

Conclusion

Kauche's case study of "merging 83% of PRs with AI review alone" represents an important milestone signaling an evolution from AI coding to AI review operations.

• Overall picture: Claude Code Action + GitHub Actions + auto-merge
• Conditions for success: Comprehensive CI, low-risk PR classification, and continuous prompt refinement
• Phased adoption: Documentation → dependency bumps → tests → feature additions
• Caveats: Audits for overlooked bugs, prompt injection safeguards, and preventing skill degradation

There is no need to expand automation across the entire organization from day one. You can gain an operational feel by starting with documentation updates alone. While referencing Kauche's case as a target vision of what is achievable, build momentum step by step from a realistic starting point for your team.

For broader Claude Code utilization, please see Claude Code Workflows; for second-opinion techniques in the AI coding era, refer to Rubber Duck Mode in Copilot CLI.

---

References

• Enabling 83% of All PRs to Be Merged with AI Review Alone — Zenn
• Claude Code Action — Anthropic GitHub
• GitHub Auto-merge Documentation