The natural response to an incident is to automate detection. We had the idea within the hour: an agent that watches expiry dates and warns us in advance.
We didn't build it.
The process had three layers, and we were about to automate the third one. The first is automatic renewal with a payment method that actually works. The second is a written record of what expires and when. The third is a probe that verifies. An agent that warns you about something that could have renewed itself isn't automation — it's expensive theater staged on top of a broken process.
What we did instead: fix the first layer, which is where the real failure was; write the second one down in a file; and add the third as a few lines inside a check that already existed.
The uncomfortable part: we sell process redesign with AI, and our first instinct was to throw an agent at a problem a working payment card would have solved. The urge to automate is strongest right after something breaks, because building something feels like fixing something.
Eliminate before you automate. And before you eliminate, find out which layer is actually failing.
We'll tell the whole incident in a few weeks.