Automating Customer Support Replies with Claude
The Goal Isn't Full Automation, It's Faster Humans
The best support automation setups don't remove people, they remove the blank-page problem. An agent staring at "where's my order" for the 40th time today isn't adding value by typing the same sentence from scratch, they're adding value by catching the 5% of cases that need real judgment.
- Ticket comes in
- Routine or sensitive?
- Claude drafts a reply via MCP
- Confidence check
- Human approval
- Send the reply
Sort Tickets Into Two Buckets First
Before automating anything, separate your ticket volume:
- Routine, low-risk: order status, shipping timelines, return policy questions, sizing questions with a clear chart.
- Sensitive or judgment-heavy: damaged goods, billing disputes, angry customers, anything involving a refund exception.
Only the first bucket should get AI-drafted replies with light review. The second bucket needs a human from the start, Claude can still help by summarizing the ticket and pulling relevant order history, but it shouldn't be drafting the emotional parts of that reply.
A Working Draft-and-Review Setup
- Ticket comes in through your help desk.
- Claude (connected via MCP to your order data) drafts a reply using the actual order status, not a guess.
- The draft is flagged by confidence: routine tickets get a "looks good, send" queue; anything ambiguous gets flagged for a full rewrite.
- An agent reviews, edits if needed, and sends.
- Every edited reply is worth reviewing weekly, patterns in what gets changed tell you what to fix in your instructions.
What to Give Claude So Replies Don't Sound Robotic
You're drafting a support reply for [store name]. Tone: warm but concise, no corporate
filler like "we sincerely apologize for any inconvenience this may have caused."
Here's the customer's message: [paste]
Here's their order info: [paste or pull via connector]
Here's our return policy: [paste relevant section only, not the whole policy page]
Draft a reply. If you're not confident about any detail, say so instead of guessing.
Where This Goes Wrong
- Auto-sending without review on anything involving money, refunds, or a clearly upset customer.
- Feeding it your entire policy document instead of the relevant section, longer context isn't always better, it's easier to bury the actual answer.
- Not tracking edit rates. If agents are rewriting 80% of drafts, the prompt needs work, not more patience.
The Metric That Actually Matters
Don't just track response time. Track edit distance, how much agents change before sending. A shrinking edit rate over time means your setup is actually learning your store's voice and policies, not just generating plausible-sounding text.
