> ./posts
-
AI automation is easy to demo and hard to operationalise
The prompt is maybe 10% of the work — the other 90% is audit logs, retries, idempotency, permissions, and the workflow design that decides what happens at 2am when something fails.
-
How I would design an AI intake workflow for insurance brokers
Provenance per field is the difference between automation and theatre — without a clickable source span, brokers re-read the document anyway.
-
Why production voice AI fails after the demo
Most production voice AI fails not because the model is wrong, but because the state machine, latency budget, and barge-in handling were never designed for real callers.