Your LLM feature ships. It works great. Three months later: your cloud bill is $140K/month vs $12K projected. P95 latency is 8 seconds — users think it's broken. API error rate is 3.1%. No SLOs were defined, so no one knew any of this was unacceptable until it was catastrophic.