Find the journey that keeps restarting
A customer calls to ask about a bill. An employee checks one system, then another, then sends a message to a specialist. The customer calls again because nobody owns the next step. The problem is larger than the conversation: the work has no reliable path from question to resolution.
That kind of recurring journey is a better starting point than a general promise to automate a department. Look for requests with enough volume to matter, a recognizable sequence, and an outcome people can name. Interview the team doing the work and follow a few real cases from first contact to closure.
Map the answer and the action
For each step, ask where the answer comes from, who may see it, what action is allowed, and which exceptions require judgment. A voice agent may be able to explain a verified balance but should not invent a settlement. An employee agent may assemble an inventory recommendation while the merchandise owner makes the transfer decision.
If the required facts are fragmented or unreliable, the first release may be a verified view for employees. If the facts are sound but the queue is the bottleneck, a bounded agent may be the right next layer.
Release one complete path
Choose one request type, one set of source records, and one defined handoff. Test ordinary cases, conflicting data, missing information, and requests outside the agent's authority. A narrow workflow that finishes safely teaches more than a broad assistant that answers many questions but owns no outcome.