The unreasonable effectiveness of giving an agent a refund button.
What happens when you stop treating AI as a read-only oracle and start letting it take real actions — with real consequences.
Building agentic workflows is one of those things that sounds straightforward until you actually try it. You wire up a model, give it some tools, write a system prompt — and then watch it hallucinate its way through your production database.
The setup.
We had a simple goal: let the support agent issue refunds under $50 without human approval. The reasoning was sound — 80% of refund requests were under that threshold, and the median resolution time was 4 hours. An agent could do it in seconds.
The interesting question wasn't can it use the tool. It was what does it do on the day it shouldn't.
What actually happened.
Three things surprised us:
-
The agent was more conservative than humans. It asked clarifying questions we never thought to require. "Was this a duplicate charge or a change of mind?" matters for fraud detection — the agent figured that out from the policy docs alone.
-
Edge cases surfaced faster. Within the first week, the agent flagged 12 cases that didn't fit any existing policy. A human would've just picked the closest match and moved on.
-
The refund rate dropped. Not because the agent was stingy — because it resolved issues before they became refund requests. Turns out, half the "I want a refund" messages were really "I can't find my order."
The takeaway.
Giving an agent real authority — not just advisory power — changes how it behaves. It becomes careful. It asks better questions. It treats the tool like something that matters.
The best agents aren't the ones that do the most. They're the ones that know when not to act.
This is the part nobody talks about in the "AI agent" discourse. The interesting problem isn't capability. It's judgment.