You don't have to leave a conversation to use an AI agent on it. The composer can be pointed at an agent instead of at the customer, so you can ask for a draft, an investigation, or a lookup with the whole thread already in front of it.
In the composer, switch the destination from replying to the customer to asking an agent. The agent appears as a pill, the send button changes to name it, and anything you type goes to the agent rather than out.
Agents with quick actions offer them as buttons here — “Draft customer reply” and the like — so the common requests are one click rather than a sentence you have to phrase.
Drafting a reply you then edit, which is faster than starting from an empty box on a long thread.
Looking something up in a system the agent has a tool for, without switching to another tab.
Asking a specialist agent about a case outside your area — a billing question landing with a technical teammate.
Starting a code investigation on a bug report, so the thread carries the finding.
A drafted reply can be taken into the composer rather than copied by hand, and it arrives wrapped in the greeting and closing your project uses — so it reads as your team wrote it, with names filled in.
Read it before sending. A draft is confident by construction; whether it's right depends on the content it drew on, and you're the one who knows this customer.
It works in the other direction too. An agent that hits something it can't decide stops and asks, and the conversation appears in the Needs my input view until someone answers.
Check that view daily if you run agents at all. A run waiting on a question is a customer waiting too, and it's the one queue that doesn't announce itself.
