Health check fails at 2am. Left wakes a human and holds the rollback until someone acks. Right rolls back on its own in 41 seconds and you read about it at 9.
Same pipeline, same failure, drawn in the style of my CapCut World LA film.
Which one would you let touch production, and what would you need to see in the logs before trusting it?
The log line I check first is the error rate during those 41 seconds. A rollback that completed cleanly but still burned 38% of requests on the way is not clean, it just finished.
Same support inbox, one difference. Left keeps a human on every send. Right lets the AI send on its own once the draft confidence passes 90, refunds and cancellations excluded.
I built both in Relay this week and I still cannot decide.
Would you flip that switch for your own customers, and at what number?
I'd flip it per topic, not per score. A 90 on a question the AI has answered hundreds of times is safe. A 90 on a question it has never seen is still a guess. Can Relay set a different threshold per topic?