The agent feature we shipped, then quietly killed 3 months later

Some lessons only land when you have to retire something you were proud of.

What we shipped: a fully autonomous agent that could approve low-risk customer service refunds without human review. Cost-justified, well-scoped, passed every test.

What killed it 3 months later: edge cases compounded. Customers learned how to phrase requests to trigger autonomous approval. Refund fraud crept up 4x. The cost of the fraud exceeded the labor savings.

The lesson, autonomy without adversarial testing is a liability. We shipped supervised approval instead. Slower, way less fun in demo. Still in production today.

What's an AI feature you shipped, then had to quietly walk back?