How to frame queries and answers from customer Agent utterances using Deep Learning SOTA

I am working with smart-reply use case for Async chat customer and agent chat transcripts.

The one chat bubble looks like this:

From one chat bubble how can I identify intent queries/Question by Customers and Responses from Agents. In between chat there is lot of unnecessary conversation, so how can I remove such things.

Final Data-set:

cust_query                              Agent_response
Pending Payment                        You've reached PayPal messaging! Hi, Person This is Person from Google and I'm happy to assist you today. I see here that you're reaching us out about a pending payment. I see here in my end that the tracking number that you provided us is invalid.I suggest you double check the tracking number and upload it again
Can I cancel the transaction        I’m transferring this message to Dispute Department. They’ll review our previous conversation and get back to you. The response may not be immediate, but you'll get a notification via email, App and within your Google account as soon as they reply.

Can we utilise any transformer based models to get this thing.