An AI support agent looked green for 3 weeks while misrouting refund requests
Warm-Reaction-456 · reddit · 2026-07-26
A support-mail agent looked green in the dashboard for 3 weeks, but it repeatedly sent welcome emails to customers who were actually asking for refunds.
The key failure was that the system was treated as "handled" as long as it took an action, not whether it took the right action. Nobody was reading the full outputs; they only checked summaries. The author says this is the wall the agent market is heading toward: "autonomous" is not the same as "unsupervised."
One concrete example was a German customer, where a withdrawal request can trigger consumer-law deadlines immediately. The bot’s cheerful onboarding reply became a compliance problem.
The team kept the agent, but changed the workflow:
- anything involving money or cancellations goes to a human
- someone now reads 15 full threads every Friday
- the review costs about 40 minutes a week
The takeaway is that agent systems need real auditing, not just pass-rate dashboards.
More from coding & agent
- Inference providers may need smarter cache windows than a fixed five minutes — lucasmeijer · 2026-07-26
- Coding agents are already doing far more than coding, from browser automation to campsite search — yacineMTB · 2026-07-26
- A coding agent used remote browser control to build a campground search site — yacineMTB · 2026-07-26
- How one developer stopped being vague with an AI coding agent — rseroter · 2026-07-26
- Karpathy says AI coding is about knowing when to delete the worktree and say no — lemire · 2026-07-26
- Productive Engineer Shares AI Coding Workflow: Never Reads Code, Relies on Agent Plans — danshipper · 2026-07-26