In this article:
- What is an AI email assistant?
- Should the AI draft replies or send them?
- Which support emails should you give to AI first?
- How do you set up an AI email assistant step by step?
- What should an agent check before sending an AI draft?
- How do you know the AI drafts are helping?
- What do customers think about AI in support?
- Which AI model should draft your replies?
- How LiveAgent helps with AI email replies
What is an AI email assistant?
An AI email assistant is software that writes replies to customer emails for your support team. It reads the incoming message and the earlier thread, looks up facts in your help articles or saved answers, and returns a reply in your chosen tone. The useful ones work inside the help desk, not in a separate chat window.
The difference from a plain AI email generator is context. A generator turns a short prompt into a new email, which works for a sales follow-up. A support reply has to match what the customer actually asked, what your policy says today and what the agent already promised two messages ago.
Help desk vendors now ship this as a built-in feature. Help Scout calls it AI Drafts : an agent clicks a quick link in the reply box and gets a draft built from past conversations and Docs articles.
Zendesk’s auto assist suggests replies that need agent approval before they reach the customer. We checked both vendor pages on 8 October 2026.
Should the AI draft replies or send them?
Start with drafts. In draft mode, the AI writes and a person decides, so a wrong answer costs you a few seconds of editing instead of an unhappy customer. Move a topic to send mode only after the drafts for that topic need almost no changes for several weeks.
The two modes solve different problems, and you can run both side by side:
| Draft mode (agent sends) | Send mode (AI agent sends) | |
|---|---|---|
| Who presses send | A human agent | The AI agent itself |
| What it saves | Typing and searching time | The whole reply, including wait time |
| Main risk | Agents approve drafts without reading them | A wrong or off-tone answer reaches the customer |
| Good fit for | Mixed inboxes, new setups, sensitive topics | Narrow, repeated questions with one correct answer |
| Examples | LiveAgent Answer Composer, Help Scout AI Drafts, Zendesk auto assist | An AI agent with LiveAgent’s Reply via email tool |
The risk in draft mode is quiet. After a few good weeks, agents start clicking send without reading, and draft mode turns into send mode with a human signature on it. The review checklist further down exists to stop that.
Which support emails should you give to AI first?
Give the AI the emails your agents answer the same way every time. Order status, password resets, opening hours and “how do I” questions have one correct answer that lives in your help center. Keep refunds, complaints, legal requests and anything involving money or personal data with a person.
A quick way to sort your inbox is to look at the last 100 email tickets and put each one in a bucket:
| Email type | Example | Suggested mode |
|---|---|---|
| Repeated how-to question | “How do I add a second user?” | Draft first, send later |
| Status question | “Where is my order?” | Draft first, send later if your data is connected |
| Account change | “Please change my billing email” | Draft, agent confirms |
| Complaint or bad experience | “This is the third time it broke” | Human writes, AI can suggest a summary |
| Refund, credit, contract | “I want my money back” | Human only |
| Legal or data request | “Delete all my data” | Human only, follow your process |
If more than half of your sample lands in the first two rows, an AI email assistant will save real time. If most emails are complaints or one-off cases, the AI will still help with summaries and tone, but expect smaller gains.
How do you set up an AI email assistant step by step?
Set it up in six steps: pick the topics, fix the knowledge behind them, set the tone, run drafts only, review the edits and then widen the scope. Plan an afternoon for the first three steps and two to four weeks for the review, which is the step that is easiest to skip.
Step 1: Pick three to five topics
Choose the topics from your inbox sort above. Fewer topics make the review easier, because you can tell quickly whether the drafts are right.
Step 2: Fix the knowledge behind those topics
The AI repeats what it finds. Read the help articles and saved replies that cover your chosen topics, and update old prices, renamed menus and changed policies before the AI starts quoting them.
Step 3: Set the tone and any fixed rules
Pick a formality level that matches how your team writes today. Add short instructions for things that never change, such as “never promise a delivery date” or “always link the relevant help article”.
Step 4: Run in draft mode only
Turn the assistant on for your chosen topics and ask agents to generate a draft on every matching ticket. Agents edit and send as usual.
Step 5: Review the edits every week
Once a week, read ten sent replies next to their original drafts. Every repeated edit is a fix for your knowledge sources or your instructions, not a reason to give up on the tool.
Step 6: Widen the scope or move a topic to send mode
Add new topics one at a time. Move a topic to send mode only when its drafts go out with no meaningful edits, week after week.
Join our newsletter
The latest tips and deals delivered straight to your inbox.
What should an agent check before sending an AI draft?
An agent should check five things before sending: the facts, any promise, the tone, the recipients and the personal data in the reply. Facts and promises cause most of the damage, because a confident wrong answer is worse than a slow right one.
- Facts: does the answer match the current help article, price or policy?
- Promises: does the draft commit to a refund, a date or a fix your team cannot guarantee?
- Tone: does it fit this customer’s mood? A cheerful reply to an angry customer reads badly.
- Recipients: is the reply going to the right people on the thread, and only to them?
- Personal data: does the draft repeat account details or order data the customer did not ask about?
This list should take about a minute per ticket. If it feels slow, that is a sign the topic is not ready for send mode yet.
How do you know the AI drafts are helping?
Track two numbers: how often agents send drafts with only small edits, and how often customers reply again because the answer did not solve the problem. A rising share of light-edit drafts with a steady reopen rate means the assistant is working. Faster replies with more reopened tickets mean it is just moving work around.
Add first response time and customer satisfaction ratings for the same topics, and compare them with the four weeks before you started. For example, if a five-agent team sees its first response time on how-to emails drop while reopen rates stay flat, that topic is a good candidate for send mode. This is an illustration, not a benchmark: your own numbers are the only ones that count.
What do customers think about AI in support?
Customers accept AI help, but they want a way out. In a Gartner survey of 3,566 B2B and B2C customers from early 2026, 50% said their interactions are easier when companies use GenAI, and 87% said it is essential to offer an option to reach a human agent (Gartner, August 2026 ).
For email, that means a reply should never trap the customer. Whether a person or an AI agent sends it, the customer must be able to answer back and reach someone who can make a decision. Draft mode handles this by design, because a person already owns the ticket.
Which AI model should draft your replies?
Pick the model by testing it on your own tickets, not by its name. A fast, low-cost model is often enough for short, repeated questions, while a stronger one may suit long threads. Compare drafts from two models on the same 20 emails and keep the one your agents edit less.
LiveAgent’s AI agents can use models from several providers. The October 2026 product update added OpenAI’s GPT 6 Astra, GPT 6 Sol and GPT 6 Luna and Google’s Gemini 3.7 Flash and Gemini 3.8 Flash, and it lists Claude Opus 5, Claude Sonnet 5.5 and Claude Opus 5.5 for Anthropic.
Anthropic released Claude Haiku 5.5 on October 7, 2026, after that update came out, and it is already available to everyone in LiveAgent. Anthropic calls it its fastest model to date and says it works especially well for speed-sensitive tasks like live customer support (Anthropic, October 2026 ). It lists $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens.
A simple test shows which model fits your inbox:
- Take 20 real tickets from the topics you chose in step 1.
- Generate a draft for each ticket with each model you want to compare.
- Ask an agent to edit the drafts and count the changes and the wrong facts.
- Keep the model with fewer edits, then check what it costs at your ticket volume in the AI credits guide .
Model lists change quickly, so check the list in your own account before you decide.
How LiveAgent helps with AI email replies
LiveAgent is our product; this guide covers it alongside alternatives. LiveAgent offers both modes from this guide: AI drafts for agents inside the ticket, and AI agents that can answer email tickets on their own.
Drafts for agents. The AI Answer Composer opens from the robot icon in the ticket reply editor. The agent picks casual, neutral or business formality, can add custom instructions, and clicks Generate answer to get a draft based on the ticket’s conversation history.
The draft then opens in the AI Answer Improver , where the agent can improve, extend or simplify it, or give their own instruction. The Improver keeps every version, and Use in ticket reply moves the chosen one into the reply editor.
Knowledge behind the drafts. The Answer Composer runs on a FlowHunt flow, and its answers come from knowledge sources you add in FlowHunt: scheduled checks of your website, question and answer pairs, and uploaded documents. Your existing predefined answers still work for the replies that should never change.
Replies sent by AI. For send mode, the Reply via email tool on the LiveAgent MCP server lets an AI agent write the answer and send it to the customer, as described in the October 2026 product update . The reply goes to the people on the thread, the AI agent can adjust the To and Cc recipients, and it does not reply on closed or deleted tickets. There is no draft step in this mode, so keep it for the narrow topics that passed your review.
Cost. LiveAgent charges no setup fee for the Answer Composer. Usage is billed through FlowHunt credits, and the AI credits guide helps you estimate the cost for your ticket volume.
If your team handles most of its work in email, the email management page shows how email tickets fit into the shared inbox alongside chat and calls.
Your next step
Pick the five email topics your agents answer most often and run AI drafts on them for two weeks. The edits your agents make will tell you which topics are ready for an AI email assistant to send, and which ones need a person.
