AI suggested replies are answers that an AI drafts for a support agent during a conversation. The agent reads the draft, edits it if needed, and decides whether to send it, so nothing reaches the customer without a person approving it. Used well, they save typing on repeat questions and help newer teammates answer with confidence. Used carelessly, they let a wrong answer go out with a human's name on it.
This guide covers how suggested replies work, where they help, a review routine your team can follow on every draft, and the house rules that keep them accurate.
What AI suggested replies are (and how they differ from other options)
Most support teams already use some kind of shortcut. Suggested replies sit between canned text and a fully automated AI assistant:
| Saved (canned) replies | AI suggested replies | AI answering visitors directly | |
|---|---|---|---|
| Who writes the text | You, in advance | The AI, for this conversation | The AI, for this conversation |
| Who sends it | The agent | The agent, after review | The AI, with no review |
| Fits the exact question | Only if you wrote one for it | Usually, if your content covers it | Usually, if your content covers it |
| Main risk | Sounds generic or slightly off-topic | Agent sends a wrong draft without reading it | Wrong answer reaches the visitor directly |
| Best for | Greetings, closings, fixed policies | Repeat questions with small variations | After hours and high volume |
The drafting step usually works like an AI assistant trained on your content: the tool looks at the conversation so far, finds the parts of your knowledge base that match, and writes an answer from them. Our guide to the AI chatbot trained on your data explains that retrieval step in plain English.
Why start with suggested replies
If you're curious about AI but nervous about letting it talk to customers on its own, suggested replies are a sensible first step. Your team keeps full control, and you get to see the quality of AI answers on your own content before anything is automated.
There's also research on how this kind of help affects support work. A study of 5,179 customer support agents found that access to a generative AI assistant raised the number of issues resolved per hour by 14% on average, with the largest gain, 34%, for novice and less-skilled agents, and minimal impact on the most experienced ones (Brynjolfsson, Li and Raymond, "Generative AI at Work", NBER). That's one study, not a promise for your team, but the pattern makes sense: a good draft helps most when you don't know the answer by heart yet.
The flip side is that language models can sound sure of themselves while being wrong. The US National Institute of Standards and Technology describes this as the production of confidently stated but erroneous or false content. A suggested reply moves that risk from the customer to your agent, which is only useful if the agent actually checks the draft.
Where suggested replies help, and where to write your own
| Use a suggested reply | Write it yourself |
|---|---|
| Questions your FAQ or policies answer, phrased in a new way | Complaints, frustration or anything emotional |
| Explaining how something works, step by step | Refunds, exceptions, discounts and custom quotes |
| Answering in a language you read but write slowly | Questions about a specific order or account |
| Restating a policy in a friendlier, shorter way | Legal, medical, financial or safety questions |
| A first draft when you're new to the business | Anything your knowledge base doesn't cover |
A simple test: if you'd need to look something up in your own help pages to answer, a draft can save time. If the answer depends on judgment, a relationship or a promise, start from a blank reply.
A review routine for every draft
Give your team one short routine and use it on every suggestion. It takes seconds once it's a habit.
- Read the visitor's last message again. Does the draft answer what they actually asked, or a nearby question?
- Check every fact. Prices, times, dates, steps, names of plans or services. If you aren't sure, open the source page before sending.
- Look for promises. Delete or soften anything that commits you to something you can't guarantee ("we'll have it to you by Friday").
- Add what only you know. The visitor's name, what they told the chatbot earlier, the page they're on, the next step you'll take.
- Cut it down. Chat replies work best short. Two to four sentences, then a question or a clear next step.
- Match your voice. Swap stiff phrases for the way your team really talks.
- Send, or start over. If the draft needs more than a light edit, write your own reply instead of fixing a weak one.
Example: editing a draft
Say you run a small landscaping company. A visitor writes: "Do you do weekend visits? Our yard is a mess after the storm and we need someone fast."
The AI suggests:
Yes, we offer weekend appointments and can usually send someone within 48 hours. Would you like to book a visit?
Your knowledge base says you visit on Saturdays only. Your team also knows that demand spikes after storms, so you never promise a response time then. The agent edits:
Sorry about the storm damage, that sounds stressful. We do visits on Saturdays (not Sundays). After big storms we're busier than usual, so I can't promise a time yet, but if you share your address and email, I'll check this Saturday's schedule and get back to you today.
The edit fixed a fact (Sundays), removed a promise (48 hours), added empathy, and gave a concrete next step. It also showed a gap: the storm rule lived only in your team's heads, so the AI guessed. Adding one line about it to the knowledge base improves future drafts.
Set up the knowledge behind the drafts
Suggested replies are only as good as what the AI can draw on. Before turning them on:
- Write down what agents answer every week: services, prices, hours, delivery or visit times, returns, cancellations.
- Say what you don't do. "We don't work Sundays" prevents a confident wrong yes.
- Keep one source of truth per topic, and remove outdated versions.
- Add house rules for risky topics, such as "never promise delivery dates" or "don't agree to refunds; offer to pass the request to the owner."
- Choose a tone that matches how your team writes, so drafts need less editing.
Our guide to preparing a knowledge base for an AI chatbot walks through this content work step by step, and the same preparation serves both suggested replies and AI answers to visitors.
Team rules for using AI drafts
A few written rules prevent most problems:
- Never send a draft you haven't read in full. The agent who sends a reply owns it.
- Sensitive topics are always handwritten. Use the "write it yourself" column above as your list.
- Flag gaps as you find them. When a draft is wrong because the knowledge base is missing something, leave a private note or tell whoever maintains the content.
- One person per conversation. Suggested replies don't fix double replies; clear ownership does. Our guide to the shared inbox for live chat covers assignment and notes.
- Keep your saved replies. Greetings and closings you've already polished don't need AI. Our live chat scripts are a good starting set.
How to tell whether it's working
You don't need a data team. Over the first month, watch:
- How often drafts are sent with light edits versus rewritten. If agents rewrite most of them, your content has gaps or the tone setting is off.
- The gaps your team flags. This list should shrink as you add content.
- First response time, if your tool reports it. Drafts should help agents answer repeat questions faster.
- Mistakes caught in review. Read a sample of sent replies each week, especially from newer teammates.
If drafts keep coming out right for a topic, that's a sign the topic may be ready for an AI assistant to answer visitors directly, inside a flow that hands off when unsure. If they keep going wrong, our guide on how to stop AI chatbot hallucinations helps you find the cause.
How to do this in RetroChat
On the Autopilot plan, agents can click Suggest reply while chatting with a visitor in the shared inbox. The AI drafts an answer from your knowledge base and the conversation so far, and you review it and send it, or edit it first. Nothing goes to the visitor until you send it. See AI assistant for the setup.
The drafts use the same knowledge as RetroChat's AI assistant: typed knowledge (up to 15,000 characters), uploaded PDF, Word, text, Markdown or CSV files, and imported web pages you can refresh when they change. Your tone setting and house rules apply too. Each suggested reply counts as one AI reply, from the 20,000 included every month on Autopilot.
Next to the conversation, agents see everything your chatbot collected, so they can add the personal details a draft can't know. Private notes and assignment help the team flag gaps and keep one owner per chat, and saved replies are included on every plan. See Inbox and team for how the inbox works.
Try it on your own content
Start with your ten most common questions, write clear answers for them, and let your team use drafts with the review routine above. The free 14-day RetroChat trial includes 10 AI replies, so you can test suggested replies on your own content with no credit card. Start a free trial, or book a demo to see it with your questions.
Frequently asked questions
What are AI suggested replies?
They're replies an AI drafts for a support agent during a conversation, based on your knowledge base and the messages so far. The agent reviews the draft and decides whether to send it as is, edit it or write their own. The customer only sees what the agent sends.
Are AI suggested replies accurate enough to send without editing?
Sometimes, but you shouldn't send any draft without reading it. Accuracy depends on whether your knowledge base covers the question clearly. Check facts and promises every time, and write sensitive replies yourself.
How are AI suggested replies different from canned responses?
Canned (saved) replies are fixed text you write in advance and reuse. AI suggested replies are written fresh for each conversation, so they fit the exact question better, but they need checking because the AI can get details wrong. Most teams use both.
Do AI suggested replies replace support agents?
No. They draft text; a person still reads the conversation, judges what the customer needs and decides what to send. They're most useful for speeding up repeat answers and helping newer teammates.
What should I check before sending an AI-drafted reply?
That it answers the visitor's actual question, that every fact matches your own content, that it doesn't promise anything you can't guarantee, and that it sounds like your team. If it needs heavy changes, write your own reply instead.