TL;DR
Paste your top 10 recurring messages, a 3-5 word brand voice, your policy guardrails, and 5 real replies from your best agent into ChatGPT (GPT-5.5) or Claude (Sonnet 4.6). Ask for one reply per message type built on a four-line skeleton — acknowledge, address, action, close — capped at 90 words with [bracketed] placeholders. A human reviews and personalizes before sending; AI never auto-sends. Teams that template their high-volume queries this way routinely cut first response time by 40-74% in the first year (per 2026 industry data below), without the robotic feel of a generic macro.
Why templates, not a black-box bot
There are two ways to apply AI to a support inbox, and they solve different problems:
- Drafting templates (this article): you generate a reusable reply library in a chat tool, then agents paste, edit, and send. You keep full control of tone and policy. Cost is your existing ChatGPT or Claude subscription.
- A live AI agent (Intercom Fin, Zendesk AI): a bot reads each ticket and answers autonomously. As of June 2026, Intercom Fin charges about $0.99 per resolution with a 50-resolution monthly minimum, and resolves roughly 67% of conversations on average across its customer base. Zendesk’s Advanced AI add-on runs about $50/agent/month on top of a seat license.
Templates are the right first step for most stores under a few thousand tickets a month: near-zero added cost, no integration project, and a human always in the loop. You can graduate to a live agent later for the queries the templates handle cleanly.
When this workflow fits
- You handle 30+ messages a day and see 5-15 recurring topics.
- You want consistent tone across multiple agents.
- You can supply 5-10 real past replies from your best agent as a voice reference.
- You review every template before it reaches a customer.
Skip AI templates — and route to a human — for complaints involving safety, refunds above your escalation threshold, legal claims, high-value accounts where personalization beats speed, and sensitive topics (deaths, accidents, medical issues), where a template reads as cold.
Pick a tool
Any of the three big assistants can draft these. Pricing and context are as of June 2026.
| Tool | Plan to use | Price | Strength for support replies |
|---|---|---|---|
| ChatGPT (GPT-5.5) | Plus | $20/mo | Fast iteration; strong at matching a pasted tone; ~320-page in-app context on Plus |
| Claude (Sonnet 4.6) | Pro | $20/mo ($17 annual) | Most natural first drafts; holds a long brand-voice brief without drifting; 1M-token context |
| Gemini 3.1 Pro | Google AI Pro | $19.99/mo | Best when replies must cite live policy or shipping-carrier facts; 1M context + Workspace |
For pure tone-matching and a clean first draft, Claude Sonnet 4.6 is the safest default. If your replies need current external facts (carrier cutoffs, live promo codes), Gemini’s search grounding helps. See our ChatGPT vs Claude vs Gemini comparison if you’re choosing a single subscription.
What to feed the model
Quality of output tracks quality of input. Give it all five:
- Top message types, verbatim. Paste 10 real customer messages, not paraphrases.
- Brand voice in 3-5 adjectives. For example: “warm, direct, slightly playful, no exclamation marks.”
- Policy guardrails. Refund window, replacement rules, shipping policy, and an explicit “do not promise” list.
- 5-10 sample replies from your strongest agent. This is the tone anchor that keeps output human.
- Constraints. Max length, required closer, and the placeholder format you want.
The prompt
Write customer service reply templates.
Brand voice: [3-5 adjectives]
Policy guardrails:
- Refund window: [e.g. 30 days, unworn]
- Replacement rules: [e.g. defects only, photo required]
- Do NOT promise: [e.g. expedited shipping, price matching]
Reference replies from our best agent (tone anchor):
"""
[paste 3-5 real replies]
"""
Recurring messages to template (write one reply per type):
1. [message type 1, with example wording]
2. [message type 2]
3. ...
Reply structure for each:
- Acknowledge: name what they feel, in one line.
- Address: the specific issue plus the relevant policy fact.
- Action: what we will do or what they do next, with a timeline.
- Close: warm, brief, signature placeholder.
Constraints:
- Max 90 words per reply.
- No "we apologize for the inconvenience" — use specific language.
- Insert placeholder fields like [order number], [customer name].
- After each template, add one line: "Do NOT use when: ..."What good output looks like
Each template should be plug-and-play: clearly bracketed placeholders, the four-line skeleton, under 90 words, and a one-line “do not use when” note (for example, “do not use for refund requests over $200 — escalate”). Group templates by topic so an agent can find the right one in seconds.
A “where is my order” reply built this way reads like:
Hi [customer name] — totally fair to want eyes on this. Your order [order number] shipped on [date] and is currently with [carrier], expected by [date]. If it hasn’t moved in 48 hours, reply here and I’ll open a trace and send a replacement. Tracking: [link]. Anything else I can check?
Do NOT use when: package is marked delivered but customer says not received — escalate.
That is one acknowledge line, two address-and-action sentences with a real timeline, and a brief close — no filler apology.
How to check the output before you ship it
- Run a few templates past your angriest customer scenario. Do they hold up?
- Compare against your top agent’s actual replies. Does the AI version still sound human?
- Are placeholders clearly bracketed so agents know exactly what to fill?
- Does every template name a specific action and timeline, not “we’ll look into it”?
- Does each carry its “do not use when” escalation note?
Common mistakes
- Generic apology templates. “Sorry for the inconvenience” is the phrase customers most associate with being brushed off.
- No clear next action. A reply without a timeline reads as “we ignored you.”
- Tone flattened to zero. Keep one line per template where the brand shows up — a small joke, a warm closer — or every reply sounds like the same robot.
- Letting AI translate for other markets. Generate the source-language template, then have a native speaker localize; don’t ship raw machine translation to customers.
Keep improving the library
After two weeks, pull the replies that scored highest on satisfaction and the ones that got escalated, then feed both back: “Our order-delay template gets escalated 18% of the time — rewrite it to set a firmer timeline.” Any template your agents edit more than 50% of the time isn’t a template; rewrite it. Refresh the whole set quarterly, or immediately when a policy changes.
The numbers behind this
Support teams that template and assist their high-volume queries report real gains. Across 2026 industry surveys, AI-assisted teams cut first response time by 40-74% in year one, and ecommerce deflection (questions resolved without a human) typically lands in the 50-70% range, climbing higher for narrow, well-templated topics. Klarna publicly reported cutting average resolution time from about 11 minutes to 2 minutes with AI assistance. The point of templates is to capture that speed on your top 10 questions while a human still owns tone and judgment.
FAQ
Should AI auto-send these replies? No. Templates draft; a human reviews, personalizes the bracketed fields, and sends. Auto-send belongs to a configured live agent (Fin, Zendesk AI) with guardrails, not a chat-tool template.
ChatGPT or Claude for this? Either works on the $20 tier. Claude Sonnet 4.6 tends to produce the most natural first draft and holds a long brand brief without drifting; GPT-5.5 iterates fast and matches a pasted tone well. Try the same prompt in both and keep the voice you prefer.
How do I handle multiple languages? Generate the source-language template, then have a native speaker localize it. Machine translation alone misses idiom and politeness norms that customers notice.
How often should I refresh templates? Quarterly as a baseline, and immediately whenever a policy (refund window, shipping, returns) changes. Also rewrite any template that gets edited or escalated more than half the time.
Can I just use a live AI agent instead? You can, once the templates prove which questions resolve cleanly. As of June 2026 a live agent like Intercom Fin costs about $0.99 per resolution at ~67% average resolution; templates cost only your existing chat subscription and keep a human in every loop.