Can AI handle real system administration tasks?

A sysadmin tests Claude Code on a real 13,000-line Ansible automation system (316 files) managing laptops, desktops, and servers. Over several weeks, Claude made 58 changes and generally understood the architecture well — correctly identifying where changes belong in existing playbooks rather than creating redundant ones. However, it did attempt to create a new playbook for a service that already existed, requiring human correction. Key lessons: always code-review AI output as you would a junior developer's work, avoid feeding sensitive data to AI tools, and don't let your own skills atrophy through over-reliance.