Letting AI draft customer messages: what stays with the owner's approval
Split customer messages by where the deciding material lives, and only per-client promises need your approval.
Key points
- Sort lines by checker: text, record, you
- For tone, template your rewrites for AI
- For prices, pick citable projects first
- Approve promises; write a one-line check
Contents
Not style: sort lines by what checks them
A&A perspective
Whether AI can draft a customer-facing message is decided not by your writing style but by whether the material that settles pass or fail is the text, a record or the owner's own call. Drawing the line here means reading one message sentence by sentence and splitting it into the parts a staff member, contractor or AI can draft once given a template, and the parts that stay with the owner's approval.
A&A perspective
By where that material lives, the sentences of a message fall into three groups. The first is settled inside the text. How you introduce yourself, the order of points and the tone can be judged just by reading.
A&A perspective
The second rests on facts outside the text, such as the basis for a price or which past work you may cite. The third is the promise you make to this particular client. Drafting can be delegated for the first two; the third stays with approval.
A&A perspective
Automate & Augment (A&A), the company writing this article, thinks owners end up rewriting whole messages because a single paragraph mixes all three. It is not that their style is unusual.
A&A perspective
Nothing marks which sentences are risky, so the only option left is to fix every one of them yourself.
Hypothetical example
Say a five-person company that sells corporate training replies to an inquiry like this. The message is a fictional one written by A&A.
Hypothetical example
“Thank you for reaching out. For what you describe, we can take this on as two half-day sessions. We expect the fee to be around 300,000 yen. We will build the program on what we previously delivered for a company in the same industry.”
Hypothetical example
The first sentence can be judged by reading it. The fee in the third sentence and the past work in the fourth are settled by facts outside the text. Past invoices and delivery records can confirm them.
Hypothetical example
Only the second sentence, “we can take this on as two half-day sessions,” is a promise to this client. Of the four sentences, this is the one that stays with approval.
For tone and structure, template your rewrites
A&A perspective
Tone and structure can become a template once you collect the drafts you rewrote, with each sentence as it was before and after your fix. You do not need to start by defining your style in words.
From the sources
Anthropic described Rebel Cheese, an Austin plant-based cheesemaker, in a post on September 15, 2026. The company reportedly grew from one restaurant into a national e-commerce business after appearing on Shark Tank.
From the sources
According to the post, the brand voice lived in the co-founder's head and could not scale past her own hours.
From the sources
After vetting Claude on data security, the co-founder trained a voice skill (a bundle of instructions the AI reads) on her own writing. She connected it to her email and social tools, and a teammate now drafts on-brand campaigns without routing every line through the founder, the post says.
From the sources
The official Claude Code documentation says to create a skill when you keep pasting the same instructions or checklist into chat. Once the file of instructions is in place, Claude loads it when it is needed.
A&A perspective
What A&A notes is that the material for the skill was the co-founder's own writing.
A&A perspective
From this, A&A reads that you can start by handing over past messages as material, before putting the rules of your style into words from scratch.
A&A perspective
Start with the last ten drafts you rewrote. Mark each place you changed and sort the marks into two columns.
A&A perspective
The left column holds changes you made because something read wrong. The right column holds changes you made because you knew something the drafter did not. Whoever made the fix remembers why.
A&A perspective
The left column becomes template material as it stands. Pair each sentence before and after the fix and keep the pairs in one file. Attach it to the skill's instructions and have the drafting AI read it. The right column splits into two more: facts outside the text (the second group) and promises to the client (the third group).
A&A perspective
Rebel Cheese's writing, though, is consumer marketing. A reply to a quote request differs in both its reader and the weight of the decision. The post does not say the same approach works for quotes.
A&A perspective
Leave only what landed in the left column, tone and structure, to the template.
| Group | What decides pass or fail | Drafting and approval |
|---|---|---|
| Inside the text (how you introduce yourself, order of points, tone) | The text itself; reading it settles it | Template the before-and-after sentences of rewritten drafts and let AI draft |
| Facts outside the text (basis for a price, past work you may cite) | Past invoices and delivery records; checking them settles it | The owner picks citable projects first, then drafting is delegated |
| Promises to each client (whether to take the work, discount, how far to admit fault) | The owner decides each time | Keep in approval, with one sentence saying what to look at |
For prices and credentials, pick citable projects first
A&A perspective
The basis for a price and the past work you may cite can be left to the draft once the owner has picked which past projects may be referenced.
From the sources
The same Anthropic post also describes KANE, a creative and growth agency in Princeton, New Jersey. KANE built a pipeline that scores fit from meeting notes, prices from its own past projects and drafts the proposal, according to the post.
A&A perspective
What A&A notes here is that KANE is handing over facts outside the text, not a writing style.
A&A perspective
Handing over past projects as they are causes trouble, though. The rates from a project you discounted as a favor, or one that ended at a loss, become the basis for the next quote. When you price it yourself, your memory corrects for them.
A&A perspective
So choose in advance the projects that may serve as a pricing basis and the work you may name in proposals, and list them. Once chosen, the same list works whoever the client is.
A&A perspective
How to choose is covered in our separate article on which past projects to leave out before handing quotes to AI.
A&A perspective
Facts in the right column that are already written down somewhere belong to this group. Facts not yet written down but that could be also move here, such as a client's circumstances you heard last time or terms agreed by phone.
A&A perspective
Hand over the draft before choosing, and the drafter has no option but to fill the gaps. A filled-in number reads naturally as a sentence even when it is wrong.
A&A perspective
Even the owner may miss it on reading. That is why the order stays fixed: choose first, then hand over.
Approve promises and write a one-line check
A&A perspective
Promises to each client stay with the owner's approval, and you write in one sentence what you look at to decide pass or fail. Whether to take the work, whether to discount, and how far to admit fault in an apology belong to this group.
From the sources
The same Anthropic post says that by default Claude does the work and waits for the owner's approval before anything sends, posts or pays.
From the sources
The official Claude Code best-practices guide says Claude stops when the work looks done. Without a check it can run, “looks done” is the only signal, and every mistake waits for a person to notice it.
From the sources
Give Claude something that produces a pass or fail, and the loop closes on its own, the guide adds.
A&A perspective
The guide is about coding work. A machine cannot return pass or fail on a written message. So A&A writes one sentence saying what to look at for each item kept in approval, and treats it as the check for that message.
A&A perspective
With that sentence, the drafter can check the draft before passing it to the owner. The owner, too, only needs to look at that one point rather than the whole text.
A&A perspective
For example: “Do the terms in this quote conflict with the terms we last gave this client?” For an apology: “Is the fault this message admits the same as what we agreed internally?”
A&A perspective
If you cannot write that sentence, there are two cases. If the standard exists in your head but has not been put into words, writing it down is the next job. If there is no standard at all, the item stays in approval until you set one.
A&A perspective
Something like a discount ceiling, decided once regardless of the client, moves into the same “facts outside the text” group as prices and past work as soon as you decide it.
A&A perspective
For apologies and notices of changed terms, you can also let the drafting go and keep only the sending for approval. What you split then is not groups of sentences but two permissions: drafting and sending.
A&A perspective
Whether the line is working is not measured by how many drafts you send back. Measure it by how many places the owner fixed per message.
Summary
A&A perspective
Not style: sort lines by whether the text, a record or your own call decides pass or fail. For tone and structure, turn the before-and-after sentences of your rewritten drafts into a template before handing them to AI. Hand over prices and past work only after choosing which projects may be referenced. Keep taking work and discounting for approval, and write in one sentence what to look at.
Split customer messages by where the material that decides pass or fail lives, not by style, and drafting of tone, structure, prices and past work can be delegated while only per-client promises stay with the owner's approval. What to decide next is the one sentence for each approval item that says what you look at.
FAQ
If AI reads all my past messages, won't it learn my pricing too?
It will, but it learns discounted and loss-making projects with the same weight. Choose which projects may set prices before handing them over, and those exceptions stay out of the next quote.
Should I build a template if I send only a few customer messages a month?
At a few a month, writing them yourself is often faster than building a template. The work of building it and then correcting it over time outweighs what it saves.
What if the material needed to answer isn't there at all?
Then no drafter can fix it. What needs fixing is not the line in the message but what you ask the client at the inquiry stage.
Sources & editorial note
Primary pages read for this article. Publication dates below belong to the sources; access dates record our research.
- Claude for Small Business launches new workflows, integrations, and training programs
Anthropic · 2026-09-15
Accessed 2026-10-24 - Best practices for Claude Code
Anthropic · no publication date shown; page dateModified 2026-10-20
Accessed 2026-10-24 - Extend Claude with skills
Anthropic · no publication date shown
Accessed 2026-10-24