Customer service replies: templates and when to automate them
Four levels of reply automation, a three-question rule for choosing one, and five templates for the replies a person should still write.
By the Paperlink AI teamPublished 7 min read
Short answer
Automate a reply fully only when the answer is written down, does not depend on the customer, and a wrong answer costs little. Use saved replies or AI drafts that an agent edits for everything that depends on the customer, and keep judgment calls and angry complaints with a person.
Most support teams answer the same questions every day, and most teams have already automated some replies without calling it automation: saved replies, auto-acknowledgments, an out of office message. The real question is not whether to automate, but which replies. Automate the wrong ones and customers get confident answers that do not fit their case. Automate too few and your team types the same paragraph forty times a week. This guide gives a rule for deciding, and templates for the replies that a person should still write.
Four levels of reply automation
Reply automation is not one thing. It comes in four levels, and each level suits a different kind of reply.
| Level | What happens | Suits | Main risk |
|---|---|---|---|
| 1. Saved replies | An agent picks a template and fills in the details | Common situations with a fixed answer | Sounds copied when it does not fit the question |
| 2. AI drafts | AI writes a draft, an agent edits and sends it | Long emails, reviews, replies that need a careful tone | Invented details, if the draft is sent unread |
| 3. Automatic notices | A rule sends a reply with no agent | Acknowledgments, away messages, status updates | Promises a reply time that the team does not keep |
| 4. Automatic answers | A bot answers the question itself | Questions whose answer is written in your documents | Wrong answers, if the bot is allowed to guess |
Free tools for each level:
- Level 1: the canned responses library has 100 templates in 10 groups, from first replies to closing a ticket.
- Level 2: the AI reply generator for chat and social replies, the AI email response generator for email, and the review response generator for public reviews.
- Level 3: the out of office message generator for away messages.
- Level 4: the AI answer generator and the FAQ generator, to test whether your content can answer a question before a bot does it.
Which replies to automate: three questions
For each type of reply, ask these questions in order:
- Is the answer written down? If the answer is in a help article or a policy, a bot can answer from it. If it lives only in an agent's head, write it down first.
- Does the answer depend on this customer? An answer that needs their order, their account or their history needs a person or a system with access to that data.
- What does a wrong answer cost? A wrong opening hour costs a phone call. A wrong refund promise costs money and trust. The higher the cost, the lower the level.
| Reply type | Written down? | Depends on the customer? | Suggested level |
|---|---|---|---|
| How do I reset my password? | Yes | No | 4: automatic answer |
| Do you ship to Canada? | Yes | No | 4: automatic answer |
| Where is my order? | Partly | Yes | 1: saved reply with the order details |
| Can I get a refund after the deadline? | The rule, yes | Yes, it is a judgment | 2: draft, then a person decides |
| A long, angry complaint | No | Yes | 2: draft at most; a person writes it |
| We received your message | Yes | No | 3: automatic notice |
| A legal or data deletion request | The process, yes | Yes | A person, following a written procedure |
Templates for the replies a person should write
The replies at levels 1 and 2 are where a good template saves the most time, because the structure is the hard part. These five cover the situations where agents most often stall. Replace everything in [brackets].
Saying no to an exception
Hi [name], I checked your request against our [policy name]. Your order was placed on [date], which is [number] days past the [number]-day window, so I can't refund it. What I can do is [alternative, such as store credit or a repair]. If you'd like that, reply to this message and I'll set it up today.Admitting a mistake
Hi [name], you're right, and I'm sorry. We [what went wrong] on [date]. I have [what you did to fix it], and you'll see [result] by [date]. To stop this happening again, we [change, if there is one].Asking for what you need, once
Hi [name], I can fix this, and I need two things from you first: [item 1] and [item 2]. You'll find [item 1] in [where to find it]. When I have both, I'll [next step] within [time frame].Following up after a chatbot handoff
Hi [name], I'm [agent name] from the support team. I've read your conversation with our assistant about [topic], so there's no need to explain again. The answer is [answer]. [One sentence on the next step, if there is one.]Closing the loop after a fix
Hi [name], a quick update on the problem you reported on [date]: [what was wrong] is now fixed. You don't need to do anything. Thanks for telling us; your report helped us find it.Each template puts the answer or the decision in the first two sentences, names the date or number, and ends with one clear next step. That structure matters more than the exact words.
Check the tone before you save a template
A template is sent hundreds of times, so a phrase that sounds cold once sounds cold hundreds of times. Watch for phrases that put the policy before the person, such as "as per our policy", "unfortunately we are unable to" or "please be advised". Paste each template into the tone checker. It names how the text sounds and quotes the phrases that cause it, so you know what to change.
Write down the process behind the reply
Many replies are hard to template because the decision behind them is not written down. "Can I get a refund after the deadline?" has a different answer for each agent until someone writes the rule. Write the rule as a short procedure with the SOP generator: who may approve an exception, up to what amount, and what to offer instead. Then the template only carries the words, and the procedure carries the decision.
Turn repeat replies into automatic answers
Your most used saved replies are a list of questions that customers could answer for themselves, if the answer were easy to find. Move them up a level in three steps:
- Publish the answer. Put the answer to each top question on a help page or FAQ. The FAQ generator drafts questions and answers from an existing page, so you can see which questions the page already answers.
- Test that the page holds the answer. Paste the page into the AI answer generator as the source and ask the question the way a customer asks it. If it says the text does not answer the question, the page is missing a fact that your agents know.
- Let a bot answer from the page. Only after the page passes the test.
A support chatbot at level 4 needs a safe way out for questions that the documents do not answer. In Paperlink AI, the bot answers only from your documents and website pages. When they do not hold the answer, it says so and shows a contact card with your support email or support page. You can also let visitors send the conversation to your support email in one click, which is where the handoff template above comes in. The support contact guide shows the settings.
Mistakes to avoid
- Promising a reply time in an automatic notice that the team cannot keep. Use the time you actually meet on a busy day.
- Sending AI drafts unread. A draft can contain a detail that sounds right and is not in your policy.
- Letting the template library grow without owners. Delete templates that nobody used for a quarter, and update the rest when a policy changes.
- Automating the reply but not the fix. If the same question arrives every day, the product or the help page may be the problem.