Faster, but broken, is still broken.

AI agent cut my working time in half. But without a human, it would have broken the entire system.

I was building an automation: WooCommerce → Mollie → E-boekhouding. The task looked simple — customer returns a product, status changes, a chain of actions kicks off. In practice — total uncertainty: which mutations in EPR were already created manually, which weren’t, and what a “correct” refund processing even looks like.

After doing chargebacks completely by hand, I wanted to speed things up — so I brought in the Codex CLI agent (by the way, I’ve heard it’s smarter than Claude Code — is that actually true?). We worked together: querying data, analyzing fields, figuring out what to look for and where. Not “write me a workflow” — more like a joint investigation of the system.

What happened:

→ Time cut in half → Agent built a working workflow → During testing I realized — run it without review, and it would create hundreds of duplicate mutations that could only be cleaned up manually What saved me: I already knew the database structure from the previous workflow. I could see when the agent was going off track and course-corrected deliberately. Without that context — I would have trusted it and ended up with a mess.

The final solution was also mine: create the old mutations manually one time, route all new refunds through the new automated path. Simple, clean, works. AI agents speed things up. But in processes where a mistake means hours of manual cleanup, they don’t replace expertise — they amplify it. You need a human who understands the system to steer the wheel.

And by the way — if you need something like this, you know who does it.