The useful question
“Where can we use agents?” invites a technology inventory. “Which decisions or handoffs are expensive because context changes?” points toward valuable work.
Bound the first workflow
Define available tools, permitted data, reversible actions, approval gates, time and cost limits, and a clear owner for exceptions.
Evaluate traces, not demos
A polished demonstration says little about unusual inputs, tool failure, stale context, or repeated action. Representative traces make those risks observable.
Earn autonomy
Begin with recommendation or draft modes. Expand authority only when evaluation and operating evidence show that the next boundary is acceptable.