Three different jobs can arrive in the same chat
“Do you accept batteries on this lane?” needs a service rule. “Where is my shipment?” needs a specific record. “Can it arrive tomorrow?” also involves a commitment. Giving all three to a conversational model does not establish where its information comes from.
Choose a frequent question with an answer you can check
I would start by sorting a week of anonymised enquiries. The aim is to see which information colleagues repeatedly look up.
- Service information: who approved the routes, cargo rules and document requirements?
- Tracking: which record can be retrieved, and how is the customer verified?
- Quotations: which fields are needed, and who decides the price?
Keep the first task small
One lane’s document requirements can be enough to start. Compare the original question, current reply, source and cases that still need a colleague. If two documents disagree, fix that disagreement first rather than asking AI to guess.
Keep the unresolved questions
Speed is only part of the review. Did the customer need to ask again? Did a colleague repeat the same checks? The unanswered cases can say more about the next useful improvement than a fluent reply.
