> ./posts
-
AI automation is easy to demo and hard to actually run
The prompt is maybe 10% of the work. The other 90% is the boring stuff: audit logs, retries, idempotency, permissions. Plus the workflow design that decides what happens when it fails at 2am.
-
How I would actually design AI intake for insurance brokers
Without a clickable source for every field, brokers re-read the document anyway. That's the gap between automation and theatre.
-
Why production voice AI breaks once it leaves the demo
Most voice AI demos work. Most production deployments break, and usually not at the model. It's the state machine, the latency budget, or the barge-in handling — none of them designed for real callers.