Ambitious AI programmes stall because they try to change how a business thinks before changing how it works. The reliable starting point is far more mundane.
Look for repeated, rule-shaped work: qualifying inbound enquiries, extracting fields from documents, drafting first-pass responses, routing requests to the right owner. Each of these has clear inputs, a checkable output and an obvious fallback to a human.
Automate one of them end to end, measure the hours returned, and use that evidence to fund the next one. Capability compounds; grand plans rarely do.