Build a saved reply library your team trusts: structure, naming, ownership and review habits for website chat, WhatsApp and Messenger support.
Every support team ends up with the same problem: the same twenty questions arrive all day, and agents answer them from memory, from old chats, or from a document nobody updates. The result is inconsistent tone, outdated policy and slow replies. This guide shows how to build a saved reply and macro library that people actually use, and how to keep it accurate as your business changes.
Saved replies and macros: what they are and where they differ
A saved reply is a prepared piece of text an agent inserts into a conversation and then adjusts. A macro usually goes further: it bundles a reply with actions such as tagging, assigning, changing status or sending a follow-up. Teams use the words loosely, and that is fine, as long as you decide internally what you mean.
For this article, the working distinction is:
- Saved reply: text only. The human decides everything else.
- Macro (as a team practice): a saved reply plus a written routine, such as "insert this, flag the chat, and resolve it once the customer confirms."
Even if your tooling only offers text snippets, you can still run macros as a habit by writing the routine in the reply's internal description.
Why it matters more with an AI agent in front
When an AI agent handles first-line questions, humans receive the harder remainder: refunds, exceptions, upset customers, account problems. Those conversations still contain repeatable parts, such as asking for an order number, explaining a verification step or closing politely. Saved replies cover those parts so people spend their effort on judgement, not typing.
Start with an audit, not a blank page
Resist the urge to write fifty replies on day one. Pull a sample of recent human-handled conversations from website chat, WhatsApp and Messenger, and look for repetition.
- Export or filter a recent set of conversations that a human answered.
- Group them by intent: shipping status, returns, billing, appointment changes, login trouble.
- Count how often each group appears, even roughly.
- Pick the top five to ten groups. Those become your first library.
- Note which groups the AI agent should be answering itself. If a question is purely factual, it probably belongs in your knowledge base, not in a human's saved reply.
- 01Audit recent chats
- 02Group by intent
- 03Draft and review replies
- 04Retire or update monthly
Write replies that sound like a person
The common failure of saved replies is that customers can smell them. A few habits fix most of it.
Structure each reply the same way
- Acknowledge the specific situation in one short line.
- Answer or act in the next one or two lines.
- State what happens next, including who does it, without inventing deadlines you cannot keep.
- Close with an easy way to continue.
Leave deliberate gaps
Include obvious spots where the agent must personalise: the customer's name, the order reference, the date of the issue. Write these gaps so they are impossible to miss, for example by starting them with a recognisable marker your team agrees on. A reply that goes out with a gap unfilled is worse than no reply.
Write for the channel
The same answer needs different shapes on different channels:
| Channel | What works | What to avoid |
|---|---|---|
| Website chat | Short messages, one idea each, quick back-and-forth | Long formal paragraphs |
| Friendly, concise, one clear question at the end | Heavy formatting, long link lists | |
| Messenger | Conversational tone, brief steps | Copy-pasted email language |
Keep a base version and, where it matters, a short channel variant. Do not maintain five copies of everything; two variants for your busiest topics is usually enough.
Name, organise and own the library
A library fails when nobody can find the right reply in the middle of a live chat. Naming is the cheapest fix.
A naming convention that scales
Use a consistent pattern: topic – situation – language. For example, "Returns – outside window – EN" or "Shipping – delayed – ES". Agents can then type the first word and land in the right area.
Keep it small enough to browse
- Aim for a library people can scan, not a warehouse. Merge near-duplicates.
- Group by topic, not by author.
- Mark internal-only notes clearly, such as "needs manager approval before sending a refund promise."
Give every reply an owner
Each reply should have one named owner who is responsible for its accuracy. Policy replies (refunds, warranty, data requests) should be owned by whoever owns the policy, not by whoever happened to write the text. Without owners, libraries quietly rot.
Handle the handoff moment well
The most valuable saved replies are often the ones used right after an AI agent passes a chat to a human, because the customer has just changed "who" they are talking to.
A good handoff reply:
- Greets the customer by name and confirms the person has read the conversation so far.
- Does not make them repeat themselves.
- Asks only for what is genuinely missing.
- Avoids promising a response time unless your rota really supports it.
Draft a small set for the common cases: billing question, technical fault, complaint, and "I want a person." Keep the tone calm in all of them; the customer's patience is already thinner than at the start of the chat.
Keep the library honest: review and governance
Accuracy matters more than volume. One outdated refund reply can cost more than the time the library saves.
A lightweight review rhythm
- Monthly: the support lead scans for replies that were never used and for replies agents routinely edit heavily. Both are signals. Unused replies get retired; heavily edited ones need rewriting.
- On every policy change: the policy owner updates affected replies the same day.
- Quarterly: read ten replies aloud. Anything stiff, apologetic to excess or full of jargon gets reworked.
Learn from edits
When an agent changes a saved reply before sending, they are telling you the reply is wrong for some situation. Collect those edits. If three people fix the same sentence, change the source.
Track who changed what
For regulated or high-stakes topics, you want to know who edited a reply and when. Whatever tool you use, make sure changes are attributable to a person, not a shared login.
Doing this in AI SNS CHAT
AI SNS CHAT supports saved replies in both places where humans work with conversations: the email Inbox and Activity → Chat logs, which holds conversations from every other channel, including the chat widget, WhatsApp and Messenger. In both, you can take over a chat (the bot pauses while you reply), use live translation, and insert saved replies.
A practical setup looks like this:
- Audit first. In Activity → Chat logs, use the filters to look at conversations a human handled. Bulk select, flag and export help you collect examples for grouping.
- Write the library. Create saved replies using the topic – situation – language naming pattern described above, so they are quick to find during a live takeover.
- Separate facts from judgement. Facts the AI agent should state itself belong in its knowledge, under Data sources (Files, Text snippets, Website, Q&A and others). Saved replies are for the human side.
- Turn good edits into training. When you edit a reply into the ideal answer, you can use that to train the agent, so a correction made once can improve later automated answers.
- Set up the handoff. Under Helpdesk → Scheduling (Standard plan or higher), set shifts per person and timezone so handoffs go to whoever is on shift. Alerts can reach Slack, Teams, LINE WORKS, WhatsApp, email or a webhook. When a chat is handed over, the server reads the rota and adds the "what happens next" line itself; it does not invent a wait time, so your saved handoff replies should not either.
- Review who changed what. Owners and admins can open Workspace settings → Activity to see every create, edit, delete or upload with the name, email, feature and time. That gives your monthly review a factual trail.
Roles (Owner, Co-owner, Member, Support, Viewer, with per-member permission overrides) let you decide who may change shared content and who may only use it. You can compare plans on the pricing page and browse the full feature list.
Synthia, the built-in AI fleet manager, can also guide you to the right screen if you cannot find a setting; she navigates there and highlights the control.
Conclusion
A good saved reply library is small, named consistently, owned by specific people and reviewed on a schedule. Start from real conversations, write in a human voice for each channel, and let the AI agent handle the purely factual questions. The library then supports your team's judgement instead of replacing it.
Frequently asked questions
Should I use saved replies or train the AI agent?
Use the knowledge base for factual answers the AI agent can give on its own, and saved replies for situations where a person is involved and the wording repeats. If you notice a human typing the same factual answer daily, move it into the agent's knowledge.
How many saved replies should a small team start with?
Start with the five to ten most common human-handled topics from your own audit. A small, accurate library beats a large one nobody trusts, and you can grow it as real conversations show gaps.
Can saved replies be used across WhatsApp and Messenger?
In AI SNS CHAT, saved replies are available in the email Inbox and in Chat logs, which is where non-email channel conversations live. Write a short channel variant for your busiest topics so tone fits each channel.
How do I stop outdated replies from reaching customers?
Assign an owner to every reply, update policy replies the same day a policy changes, and review the library monthly. The workspace Activity log helps you see who last edited what.


