AI can now draft a genuinely good customer support reply in seconds. The open question is not whether AI should help your agents — it clearly should — but how much control to hand it. Fully autonomous AI that replies to customers unsupervised is tempting on a spreadsheet and risky in practice. Agent-assist — where AI drafts the reply and a human approves it — captures most of the speed with almost none of the risk. Here is how that model works and why it wins for most teams.
Agent-assist vs full automation
Think of a spectrum. At one end, agents write every reply from scratch — accurate but slow. At the other, AI auto-resolves tickets with no human review — fast but exposed to hallucinated answers, off-brand tone, and edge cases the model misreads. Agent-assist sits in the high-value middle: the AI does the heavy lifting (finding the answer, drafting the reply), and the agent does the judgment (reviewing, editing, approving). You get speed and a safety net.
How AI suggested replies work in Relay
When a ticket arrives, Relay drafts a reply grounded in your knowledge base and past conversations — not freeform generation, but an answer built from your own content. The agent sees the draft, can adjust tone or length, edit anything, and then approves and sends. The AI also helps with summaries and translation. The defining rule: nothing reaches the customer until a human approves it.
The impact on response time
Most of an agent’s response time is not typing — it is finding the right answer and figuring out how to phrase it. AI drafts collapse that step: the answer is already on screen, sourced from your content, ready to review. Teams using grounded AI drafts commonly see meaningful reductions in handle time and first-response time. Be wary of inflated claims (you will see "90%+ faster" marketing); the honest story is that removing the search-and-draft step is a large, repeatable win without promising magic.
Why human-in-the-loop matters
Three reasons to keep a person in the loop. First, accuracy: a grounded model still occasionally misreads intent, and a wrong answer sent confidently is worse than a slightly slower correct one. Second, brand voice: your tone is a differentiator, and a quick human pass keeps it consistent. Third, trust: customers increasingly sense unsupervised bots, and an agent-approved reply reads as a real answer from a real company. The review step costs seconds and prevents the failures that make AI support infamous.
Suggested replies + macros + automation
AI drafts are one layer of a faster workflow, not the whole thing. Macros give one-click answers to the most repetitive questions; automation routes and prioritizes tickets so the right agent sees them first; and the knowledge base both deflects routine questions and grounds the AI drafts. Together they compress response time end to end — before, during, and after the reply.