Start with pressure, not technology
Look for enquiries that arrive while the team is unavailable, repetitive questions, delayed callbacks or administrative load that interrupts higher-value work.
Write down one failure mode in plain language. “We miss complete after-hours intake” is more useful than “we need an AI agent”.
Check repeatability
A credible first workflow has predictable inputs, a small number of permitted outcomes and named exceptions.
If every case depends on expert judgement, begin with intake and routing rather than the judgement itself.
Find the authority boundary
List what the role may capture, classify or draft. Then list what must remain human: pricing, diagnosis, public claims, payments, sensitive decisions and exceptions.
If nobody owns escalation, the workflow is not ready for a pilot.
Define evidence
Choose measures such as intake completeness, response time, accepted handoffs, escalation accuracy and QA error rate.
Avoid promising revenue or savings before a baseline exists.