Been sitting on this for a bit after watching an AI feature at my last job basically die a slow death post-launch, and I think the model-failure explanation is usually a red herring tbh. Concrete version of what I mean. We had an agent doing first-pass triage on inbound support tickets, routing + drafting a suggested reply for a human to approve. Launched, looked great for like 6 weeks. Engineering was watching latency (fine, consistently under 2s) and error rate (also fine, sub 1%). Product was...