AI-drafted support replies, done honestly

Two weeks of AI-drafted support replies: 14 hours to 12 minutes, 62% sent as written, and the rules that kept it honest. Why it matters for AI visibility.

By Dasha. Updated November 1 2026. https://heardofyou.io/guides/ai-drafted-support-replies

AI can draft your support replies well. Two conditions: a person reads every one, and the model is not allowed to guess about your product.

For two weeks, Claude pre-drafted every support email and I reviewed each draft before it went out.
- Average response time fell from 14 hours to 12 minutes.
- 62 percent of drafts went as written, 23 percent lightly edited, 15 percent rewritten from scratch.
- Thank-you replies went from about 2 a week to about 5.

Below: the setup, the rules, and the three emails that taught me the rules. Then why it matters if you want an assistant to recommend you.

Why support is an AI visibility question

Assistants learn what a product is like (/guides/how-ai-decides-what-to-recommend) from pages other people write: reviews, community threads, comparison posts. Customers write those pages after an experience with you. For a small product, that usually means after a support email.

A fast, honest, specific reply becomes a four-star review on the site the assistant cites. A confident wrong answer becomes a thread titled "does X work with Y? (it does not)". The assistant finds that thread every time someone asks.

Buyers check, too. In Product.ai's survey, 89.9 percent of AI shoppers verify the assistant's advice with another source before buying. TrustRadius found 94 percent of B2B buyers fact-check AI research.

Where do they check? Reviews, and your replies to those reviews. Support sits upstream of everything an assistant will later say about you.

The setup

Every incoming support email fires a webhook. A small function bundles the email, my notes and a stored system prompt, and sends them to Claude. The result lands as a draft in my inbox. I read it, then send, edit or rewrite. Nothing goes out without me.

What the model worked from:
- the full product list, with prices and what each includes
- a FAQ of about 3,000 words
- the refund policy, one paragraph
- about 80 past replies to similar questions
- any notes on that customer

What it did not have: permission to send, and permission to invent.

The rules, in order

1. Voice. Lowercase, warm, brief: a friendly person who built the thing herself. No corporate phrasing. No em dashes.
2. Emotional emails. If the customer is angry, frustrated or grieving, write one line of acknowledgement, stop, and tag it for me.
3. Refunds. Quote the policy and offer the refund. Nobody should have to ask twice.
4. Unknowns. Say so. Never invent a feature.
5. Sign-off. Always my name.
6. Length. Four short paragraphs at most, ideally two.
7. Status. Every reply is a draft for review.
8. Added in week two. Unsure whether the product supports a specific tool or version? Write "I think yes but let me double-check" and tag it.

Rule 2 was the most important, and it came late. The model had answered a bereaved customer in a chirpy tone. Rule 8 came from the third email below.

Three emails

The refund it talked someone out of, honestly. A customer bought a product, then realised they used a different editor than the one it was built for. They asked about a return. The draft explained that the product worked with their editor too, and how. It offered the refund anyway, if they still wanted it. I sent it unchanged. They kept the product. That is the model at its best: correct, kind, not pushy.

The one I rewrote. A customer had bought a product for their dad's business. Their dad had died, and they asked for a refund past the window. The draft was correct and generic. I did not send it.

I wrote my own: do not worry about the refund, it is already done. And if it helps, I can record a ten-minute walkthrough of winding the account down. That offer was a judgment call I would not hand to a model. They wrote back to say thank you.

The near miss. A customer asked whether a technical feature worked with a particular variant of a tool. The draft said yes, confidently. It did not: that variant worked differently under the hood. I only caught it because I recognised the name.

That is the email that would have become a public thread. It is why rule 8 exists. To see what a wrong answer like that looks like once an assistant has learned it, read AI says something wrong about my product (/ai-says-something-wrong).

The numbers

| Metric | Before | After two weeks |
|---|---|---|
| Average response time | 14 hours | 12 minutes |
| Emails handled a day | about 6 | about 6 |
| Sent as drafted | | 62 percent |
| Lightly edited | | 23 percent |
| Rewritten from scratch | | 15 percent |
| Thank-you replies | about 2 a week | about 5 a week |

The thank-yous did not go up because the model is charming. They went up because I stopped answering at 11 pm, tired. I had time to read the draft and make it specific.

What to keep for yourself
- Routine replies: the model is a better version of me than I am at the end of a day.
- Emotional replies: still mine.
- Technical claims: checked, every time. A wrong "yes" is the most expensive sentence in support.
- Apologies: mine. An apology has to be specific to mean anything.

That split matches how we run this site. AI helps with first drafts. A person checks every page and every report. We say so openly on our clean methods (/clean-methods) page.

Saying where the AI stops is not a disclaimer. It is what keeps the reviews honest, and the reviews are what the assistants read.

A first version of this appeared on my studio blog, dashbuilds.dev.

What to do next

Want to know what assistants say about your product now, including any wrong technical facts? The free check (/check) records what each one said about you across 75 answers.

Replies are good and the assistants still name someone else? The cause is usually upstream of support. Read why AI doesn't recommend your product (/guides/why-ai-doesnt-recommend-you).

Questions

Should I tell customers a reply was drafted with AI?

I review and sign every reply, so what goes out is mine. If a customer asks, I say yes, AI drafts and I edit. On this site we state it openly: Dasha writes the pages, with AI help on first drafts. Hiding it is how you end up in a thread an assistant reads later.

What is the one rule you would keep if you could keep only one?

The emotional branch. If an email is angry, frustrated or grieving, the model writes one line of acknowledgement, stops, and tags it for me. Before that rule, it wrote a cheerful reply to someone whose father had died.

What does this have to do with being recommended by AI?

Assistants learn what your product is like from reviews and threads. Reviews come from support. A fast, honest reply becomes a good review. A confident wrong answer becomes a thread an assistant quotes to the next buyer.


Sources
- Product.ai survey of AI shoppers (n=550) (https://finance.yahoo.com/technology/ai/articles/product-ai-survey-finds-four-120000400.html)
- TrustRadius, 94% of B2B buyers fact-check AI research, via MarketScale (https://www.marketscale.com/industries/software-and-technology/94-of-b2b-buyers-fact-check-ai-research-and-peer-reviews-still-close-the-deal)