Suggested Replies

What are suggested replies?

Suggested replies are draft responses an AI generates for a support agent to consider, rather than sending automatically. When a message arrives, the system proposes one or more replies; the agent can send a suggestion unchanged, edit it, or ignore it and write their own. You will also see them called suggested responses or smart replies, and the defining feature is the same in every case: the AI drafts, but a person decides.

This keeps the speed of automation and the judgement of a human in the same workflow. The agent is never handed a blank page, but nothing reaches the customer without a person signing off.

How suggested replies work

To draft something useful, the system reads the incoming message, works out what the customer is asking, and draws on the relevant source material (help articles, past replies, account context) to compose a response. Understanding the request draws on intent recognition, and keeping a longer exchange coherent draws on dialogue management, so a suggested reply fits the conversation rather than answering in a vacuum.

The best suggestions are grounded in your own content and written in your team's tone, so the agent's edit is light: a quick read, a small tweak, and send. That is where the time saving comes from: not from removing the agent, but from removing the blank page.

Suggested replies and agent-facing assistance

Suggested replies sit deliberately on the agent-facing side of support. A customer-facing chatbot answers the customer directly; suggested replies instead assist the human agent, inside the tools they already use, and leave the send button to them. That distinction matters for tickets that need a person's judgement, a careful tone, or clear accountability for what is said. The AI accelerates the reply without taking over the decision.

Because a person reviews every draft, suggested replies are well suited to exactly the conversations you would not want fully automated. The result is faster replies with the same human oversight, which is why teams measure their impact partly through first response time.

How to get the most from suggested replies

Feed the system good source material: a current, well-organised knowledge base gives it better raw material to draft from. Let it learn your tone, so drafts need less editing. And treat the agent's edits as feedback: the patterns in what they change point to where your content or guidance could be clearer. Used this way, suggested replies get sharper over time while the human stays firmly in the loop.