Give the AI 10-15 of your own sent replies plus a one-page voice sheet (greetings, sign-off, sentence length, words you use and never use), store both where the tool reuses them automatically, then for each customer email add the facts and the decision you've made. Edit the draft, and fix the sheet whenever you keep changing the same thing.
Asking an AI to "reply in my tone" without examples gets you its tone: "I hope this email finds you well", three apologies and a closing offer to help further. Your voice is mostly what you leave out, and the AI can only learn that from real emails. The other half is facts. A reply that sounds exactly like you but promises the wrong date is still a bad reply.
Pick the twelve emails that teach your voice
Go to your sent folder and choose about a dozen replies you'd be happy to send again. Spread them across situations, because people write differently when delivering good and bad news:
- Three replies to new enquiries (one you wanted, one you were lukewarm about, one you turned down).
- Two replies giving bad news: a delay, a price rise, a problem found on site.
- Two replies to a complaint or a frustrated customer.
- Two routine updates to existing clients.
- One chaser (an unpaid invoice or an overdue decision).
- Two short replies: a quick yes, a thank-you.
Remove names, addresses, amounts and anything else that identifies a client before pasting them anywhere. Replace them with brackets such as [client] and [site], which also teaches the AI where specifics belong. Skip replies written in a rush or when you were annoyed; the AI copies habits, good and bad.
Turn the samples into a voice sheet
You could paste twelve emails into every request, but a written sheet is shorter, easier to adjust and works in tools that only take instructions. Let the AI draft it from your samples, then correct it:
Below are 12 emails I've written to clients. Describe my writing voice as a
one-page style sheet someone else could follow. Cover:
- how I open and close emails (quote my actual greetings and sign-offs)
- typical sentence and paragraph length
- how formal I am, with two example phrases
- how I deliver bad news and how I handle complaints
- words and phrases I use often
- things I never do (for example: exclamation marks, emojis, "I hope this finds you well")
Only describe patterns that appear in at least three of the emails.
[paste emails]Here's an illustrative result for a partner at a five-person architect practice:
VOICE SHEET - practice partner
Openings: "Hi [first name]," for clients; "Dear [name]," only for first contact.
Never "I hope you're well."
Closings: "Best, [name]" or "Speak soon, [name]". No "Kind regards".
Length: 3-6 short paragraphs. Sentences mostly under 20 words.
Formality: plain and direct. "We'd suggest" rather than "we would recommend".
Explains the reason behind advice in one sentence.
Bad news: states it in the first paragraph, gives the cause, then the options.
Apologises once, specifically ("Sorry, that delay is on us"), never repeatedly.
Complaints: acknowledges the specific point, says what happens next and when.
Doesn't argue in writing; offers a call for anything contentious.
Often uses: "in practice", "the next step is", "happy to talk it through".
Never: exclamation marks, emojis, "please do not hesitate", "going forward".Read it the way a colleague would. In this case the partner changed two things. The AI had written "warm and friendly" as a headline trait, which would have pushed drafts towards gushing; she replaced it with "courteous, brief". And it missed a habit she thought mattered: she always ends a bad-news email with a single, specific next step and a date, never a general offer to help. She added that line herself. Expect to make changes like these; the extraction is a starting point.
Where to keep your voice so you don't paste it every time
Once the sheet exists, put it somewhere that applies it automatically. The options depend on where you already write email. Check the plan details on your own account, because these features roll out unevenly.
| Where you write | Where the voice lives | What to know |
|---|---|---|
| Gmail | Help me write, which since May 2026 can personalise drafts to match the tone and style of your previously written emails | On Workspace Business Starter, Standard and Plus, and on Google AI Plus, Pro and Ultra. It learns from your mail, not from your sheet, so give it your sheet's key rules in the prompt |
| Outlook (new Outlook) | Copilot settings, Draft Instructions, with the toggle "Use custom instructions when drafting email" | Paste a condensed version of your sheet there; it applies to every Copilot draft. Needs a licence that includes Copilot in Outlook |
| ChatGPT | A Project with your sheet in its instructions and the 12 samples as files | Paste each customer email into a chat inside the Project. Business plans don't train on your content by default |
| Claude | A Project with the sheet in its instructions, or a skill | Claude's older "styles" feature was folded into skills in 2026, so don't follow older guides that tell you to create a style |
| A help desk | The help desk's AI drafting, fed with saved replies in your voice | Help Scout, for example, includes AI drafts, with unlimited drafts on its Plus and Pro plans |
If several of you answer from a shared inbox, one house sheet plus short personal additions works better than everyone improvising. The approach for a whole company's written voice, rather than one person's replies, is in building a brand voice guide that AI can follow.
What to hand over with each customer email
The voice sheet handles how you sound. Each request still needs what to say. The AI doesn't know you've decided to accept the job, that the planning department called yesterday, or that the fee can't move. Give it three things every time: the customer's email, your decision in a few words, and the facts it must include. A request template:
Draft my reply to the email below, following my voice sheet.
My decision: [e.g. yes we can do it / no / need a site visit first]
Facts to include: [dates, prices, names of documents, next step]
Don't mention: [anything off-limits, e.g. the other quote, the staffing issue]
Length: [short / normal]
Customer's email:
[paste]Writing "my decision" in six words is the step people skip, and it's the step that stops the AI inventing a position. Asked to reply to a vague enquiry without it, an AI will usually agree to everything, politely.
Reply blocks for the explanations you give every week
Most small firms explain the same few things again and again: how fees are structured, what happens between enquiry and start date, what the customer needs to do before a visit. Writing these once, in your voice, and storing them with the voice sheet gives the AI something better than a style to copy: your actual wording for facts that shouldn't vary.
The architect practice keeps four blocks in its Project files. One, filled in:
REPLY BLOCK: how our fees work for extensions. We charge in stages rather than one lump sum: a fixed fee for the measured survey and concept design, then a fee for the planning drawings, then the detailed construction drawings if you want us to carry on. You can stop after any stage. We'll confirm each stage's fee in writing before starting it, so there are no surprises.
The instruction to the AI is simple: when a reply needs this explanation, use the block's wording, adjusting only the grammar to fit. That keeps fee explanations consistent across partners and stops the AI paraphrasing a commercial term into something slightly different. A realistic slip to guard against: the practice once left an old block in the Project files describing a stage it no longer offered, and a draft to a new client included it. Date each block and review them when your services or prices change, the same way you'd review a price list.
One reply, three drafts: the architect practice
A client emails the partner, annoyed that the planning application hasn't been submitted two weeks after the target date. The actual reason: a structural engineer's report arrived late and showed a beam that needed redesigning. The partner's decision: apologise once, explain, give a new date, offer a call.
Without examples or a voice sheet, asked only to "reply professionally and apologise for the delay", a typical AI draft opened like this (illustrative):
Dear [client], I hope this email finds you well. Thank you so much for reaching out, and please accept our sincerest apologies for any inconvenience the delay may have caused. We truly value your patience and understanding during this process. Please rest assured that we are working diligently to submit your application as soon as possible...
It apologises three times, explains nothing, gives no date, and sounds like no one in particular. The same request with the voice sheet and the three-part brief produced:
Hi [first name], you're right, the application is two weeks late, and I'm sorry. The structural engineer's report arrived later than promised and showed the steel beam over the kitchen opening needs to be deeper than we'd drawn. That changes the rear elevation slightly, so we need to redraw it before submitting. The next step is revised drawings to you by Thursday 9th; once you've approved them, we'll submit within two working days. Happy to talk it through on a call if that's easier. Best, [name]
The partner changed two things before sending. The AI wrote "you're right" at the start; she thought that conceded more than she meant, and changed it to "I'm sorry the application is two weeks late". And it gave the reason in more detail than the client needed, so she cut the half-sentence about the elevation. Two minutes of editing instead of ten minutes of writing, and the email reads as hers. That pattern, small cuts rather than rewrites, is the sign the setup is working.
The replies AI gets wrong even with your voice
Some email types go wrong in predictable ways, and the voice sheet doesn't fix them. Watch these most closely:
- Complaints with a possible claim behind them. AI drafts tend to concede ("we should have caught this earlier") because apologising sounds helpful. In your voice, that concession reads as an admission from you. For anything that could become a dispute, write the key sentence yourself, and take advice if money or liability is involved.
- Declining work. Drafts either over-explain the reasons or soften the no so much that the customer writes back to book. Add a line to your sheet: "When declining, say no in the first sentence, give one reason, and suggest where else to try if I know."
- Prices. An AI will happily restate a price from the customer's own email, including their misremembered figure. Always supply the price in your brief; never let it take one from the thread.
- Long threads. In a 15-message thread the AI can attribute a point to the wrong person, or reply to an early message rather than the latest. Paste the latest message and a one-line summary of where things stand, not the whole thread.
- Mirroring. Faced with an angry email, some drafts become defensive or stiff; faced with a chatty one, they become overfamiliar. Your sheet's rule for complaints is what holds the line.
When three people reply from one inbox
Shared inboxes complicate "sound like you", because customers write to the firm but often know the person who answers. Take an illustrative HVAC installer where three office staff answer service and quote emails. Before AI, each wrote differently: one formal, one chatty, one terse. Customers noticed only when a thread passed between them.
The fix that worked was two layers. A house sheet, agreed in a 30-minute meeting, set the things that should never vary: how the firm signs off, that engineers' arrival windows are always given as a two-hour range, that no email promises a same-day visit, and the phrases nobody uses. Each person then kept a short personal addition of three or four lines covering their own greeting and sign-off habits. The AI drafts to the house sheet first and the personal lines second.
The first week showed a conflict worth knowing about. One person's personal sheet said "keep it friendly and informal", and drafts to a customer complaining about a failed repair came out breezy. The house sheet gained a rule that overrides every personal one: for complaints, use the formal complaint pattern regardless of who is replying. If two rules can collide, decide which wins and write it down; the AI won't decide sensibly on its own.
Short-form voices: a painter and decorator
Not every voice is paragraphs. A two-person painter and decorator writes to customers the way he texts: two or three lines, first names, no sign-off beyond his name. Generic AI drafts are hopeless for him because they're all too long. His voice sheet is five lines:
Max 3 sentences unless giving a quote.
Start with the customer's first name, no "Hi" or "Dear".
Plain words. Say "Tuesday" not "early next week".
No apologies for normal things like needing a site visit.
Sign off with just my first name.With that, a reply to "can you give me a rough price for doing the hall, stairs and landing?" comes out as: "[first name], hard to price stairwells without seeing them, as it depends on the height and the state of the walls. I can pop round Tuesday or Wednesday after 4 to measure up. Which suits? [his name]". He changed nothing. The lesson: for short voices, the length rule does most of the work.
Check that it really sounds like you
Two checks tell you whether the setup is working, and both take minutes.
The edit test. For two weeks, note roughly how much of each draft you change: light touch, moderate, or rewrote it. If more than one in four is a rewrite, look at what you keep changing. It's usually one or two things (an opening phrase, too much apologising, too long), and each becomes a new line on the sheet. Most people need three or four rounds of this before drafts settle.
The blind test. Take five real replies you wrote and five AI drafts you edited lightly, remove the dates, and ask a colleague, or a long-standing client you know well, to pick which are which. If they can't reliably tell, you're done. If they can, ask what gave it away. The answer is usually a specific phrase, and it goes on the "never" list.
Keep client details out of the wrong places
Customer emails contain personal information, and sometimes more than that: addresses, family circumstances, money troubles. Draft replies in a business plan where the vendor doesn't train on your content by default (the business tiers of ChatGPT, Claude, Microsoft 365 Copilot and Gemini in Workspace all work that way), or in the drafting tools built into the email system you already use. On a personal plan, switch off the model-training setting in privacy settings. Strip details the AI doesn't need; it can write a reply about "[the extension]" as well as one naming the street.
If you're working out which emails to hand to AI at all, and which to keep for yourself, whether AI can answer your business emails for you sorts them by risk, and triaging and drafting in Outlook with Copilot covers the Outlook side in more depth. For a small practice juggling client correspondence and project email together, AI email triage for professional firms shows how to sort before you draft.
Customer email drafting with AI: more questions
Is it dishonest to send customers emails drafted by AI?
Not if you read, correct and stand behind every reply, the same as with a draft written by a colleague. The customer is dealing with you, and the content, decisions and promises are yours. Where it gets risky is sending unread drafts, which is when invented facts and tone mistakes reach customers. For fully automated replies with no human check, say so, and keep them to simple acknowledgements.
Should everyone in the team use the same voice sheet?
Use one house sheet for the shared inbox, covering greetings, sign-offs, banned phrases and how you handle bad news, and let named individuals keep a short personal sheet on top for their own clients. Customers who deal with one person for years notice when that person suddenly writes differently, so personal sheets matter most for long client relationships.
Can I let the AI read my whole sent folder?
Some tools now do this themselves. Gmail's Help me write can match the tone of your previously written emails on eligible plans, which saves building samples by hand. The trade-off is less control: it learns from everything, including rushed replies. A chosen set of 12 good emails plus a written sheet gives a more deliberate voice, and you can use both together.
Further reads
- How Agencies Keep Brand Voice Consistent in AI-Written Content — How agencies keep one voice across many writers.
- AI Email Triage for Shared Inboxes: Sales, Support, and Invoices — Sort a shared inbox so the right person drafts each reply.
- How Photographers Can Handle Client Emails With AI — Client-email drafting in another small creative business.
- How to Set Up Human Review for AI Work Without Slowing Down — Keep a light review step without slowing replies down.
- How to Chase Late Payments With AI Reminders That Sound Human — The same voice approach applied to payment chasers.
- Can AI Help a Small Architecture Practice? Where It Saves Time — Other time-savers for a small architecture practice.
- AI Quick Wins: 12 Things a Small Business Can Set Up This Week — Twelve small AI jobs sorted by setup time, each with a copyable prompt, a real business example and the catch to watch for.
- AI Prompts for Bakers: Pricing, Captions and Customer Replies — Twelve copy-ready prompts for bakers covering costing and prices, social captions, and customer replies, each with what to check before you use the answer.
- Your First 30 Days of AI Customer Support for an Online Shop — Week by week: collect real questions, let AI draft while people send, automate tracking, delivery and returns, then review 30 conversations.
- How to Answer Etsy Customer Messages Faster With AI — Quick replies for the repeat questions, AI drafts for the rest, and a tone check for the hard ones, so Etsy messages get answered well inside 24 hours.
- AI for Interior Design Studio Admin: Sourcing, Quotes, Updates — Where AI saves a design studio real admin hours: supplier PDFs into schedule rows, quote wording, chasers and Friday client updates.
- How Funeral Directors Can Use AI for Admin With Care — Where AI can take funeral admin off your desk, where it must never speak for you, and the read-back routine that stops a misspelt name reaching print.
- How Business Coaches Use AI Without Losing the Personal Touch — Where AI belongs in a coaching practice and where it doesn't, with a client memory file, recap before-and-afters and voice rules.
- Writing Pet Owner Emails and Treatment Explanations With AI — Turn a vet's shorthand into discharge emails, estimate explanations and results letters owners understand, with the vet still checking every clinical line.
- AI Service Reminders That Bring Garage Customers Back — Predict when each car is really due, write reminders that mention the actual vehicle, and let AI sort the replies, with worked garage figures.
- How to Reply to Every Enquiry in Under Five Minutes With AI — A three-lane set-up that gets a useful first reply out in minutes on every channel, without letting AI answer the questions only a person should.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Google Workspace Updates on Help me write personalisation (May 2026); Microsoft Support on Draft Instructions for Copilot in Outlook; Anthropic help on styles moving to skills; Help Scout pricing page.