Start with real enquiries
Choose a small sample: one that became a customer, one that went quiet, one that booked but never attended, one duplicate, and one that arrived outside working hours. Reconstruct what happened from the actual records. Do not rely on what the process is supposed to do.
Find the first point without an owner
For each enquiry, record where it arrived, who saw it, who owned the next action, and how that person knew. A shared inbox is a location, not an owner. If two people thought the other was responding, you have found a process problem before you have found an AI problem.
Separate speed from usefulness
A fast acknowledgement can help, but it is not the same as a useful reply. Ask whether the response answered the question, collected the right detail, or made the next step clear. Measure meaningful progress as well as time to first contact.
Check the stop rules
Follow-up needs an end condition. A reply, opt-out, booking, or manual intervention should change the sequence. If your current process cannot reliably recognise those events, fix that before adding more messages.
Build the smallest useful pilot
Choose one source, one CRM, and one accountable person. Define the required information and the expected next action. Test duplicates, missing details, and an unavailable connection. Keep a record of exceptions so your team can improve the rules.
Read the result honestly
Compare similar enquiries over a sensible observation period. More booked calls may reflect better leads, a different offer, or seasonality as well as faster follow-up. Report response time, follow-up completion, and commercial outcomes separately. A tiny sample is a learning signal, not a case study.
A reliable handover has an owner, the right context, a clear next action, and a visible exception path.