In customer service, AI works best on three tasks:
- sorting and labelling incoming messages
- finding the right policy document or the previous agreement
- drafting a reply
The employee decides, corrects and sends. With that division of labour the customer gets an answer in minutes instead of hours, without a single unchecked message ever leaving the building.
Start with one well-defined piece of customer contact, for example the first assessment of product complaints by email. Write down which step AI can take, which one the human keeps and where they meet, and agree the KPIs up front: how quickly the customer should have an answer and how satisfied they are. Measure those two before and after.
A chatbot connected to a mailbox is almost never the answer; the bottleneck is nearly always in looking up, sorting and retrieving what was agreed last time.
For knowledge that lives in documents, a connection between those documents and an AI model is usually enough: a week or so of building, not a large custom project. Build the simplest version that can work and let your sceptical colleagues actually use it for four weeks. They will find the mistakes your enthusiasts miss. And ask your people what they think, because if they do not use it the numbers mean nothing.