A pipeline can be green and still fail the business. The job ran, but the data arrived late. The table exists, but partitions are missing. The API responded, but the feature values were stale. These are reliability failures, even when the scheduler says success.
AI systems make this sharper because model behavior depends on the data path. Freshness, completeness, correctness, and recovery time should be first-class signals, not side observations buried in logs.
SLO thinking gives data teams a better language. Instead of asking whether the pipeline ran, ask whether the system delivered usable data within the promise users and downstream systems actually depend on.