From building automations for SMBs and a few mid-market clients, the failure mode I see most isn't the AI - it's the surrounding plumbing. Brittle webhook chains, no idempotency, no retry budgets, no observability. Teams treat the LLM like a black box and forget basic distributed systems hygiene. The other big one: zero data quality validation upstream, so the AI is "wrong" because the input is garbage. The model is rarely the bottleneck.