Skip to content
A handwritten note on a wooden counter beside a pen

Keeping Your Own Voice in AI-Drafted Replies

Up to date as of 17 November 2026

Photo: Pexels on Pixabay

Your replies sound like a call-centre script. "We appreciate your feedback and strive to deliver excellent service." No real owner talks like that, and customers can tell.

Building a voice the tool can followCollect replies you likeDescribe the voice in plain wordsGive examples with the briefEdit the result by hand
Collect, describe, test, then edit.

Cause one: the tool has no idea how you talk

A model starts from a general style. Give it nothing about yours and you get the average of everyone's. The fix is to show it your voice in two ways.

First, describe it. Wikipedia lists role assignment among prompt-engineering techniques: you give the model a part to play.1 Be specific: "a plain-spoken owner of a two-person plumbing business who uses first names and short sentences" works better than "professional and friendly".

Second, show it. Examples inside the prompt let a model learn from them in-context.1 Paste in a few replies you wrote yourself and liked. If you don't have any, write three short ones from scratch. They're your sample of how you'd talk.

Cause two: you told it to sound "professional"

That word nudges a tool towards stiff, formal wording. Try other instructions:

  • "Write the way I'd talk across the counter."
  • "Use contractions. No corporate phrases."
  • "Short sentences. No exclamation marks."
  • "Don't say 'valued customer' or 'we strive to'."

Tell it your own banned list. Most owners have phrases they'd never say. Write yours down once and keep it in your template.

Cause three: the phrases came back

Even with good instructions, stock lines sneak in: "thank you for taking the time", "we take feedback seriously", "we hope to see you again soon". None is wrong. Together, in every reply, they're a fingerprint.

When you edit, hunt for these and replace them with something you'd actually say. Better still, cut them. A reply with fewer words often has more voice.

Cause four: you're signing off as nobody

Google posts the reply as the business and doesn't show your personal name.2 That means the words themselves have to carry the personality. A small detail helps: the shop's name for a regular product, a local reference, a phrase you use on the phone. Only use details that are true and not private.

Where a platform allows a sign-off, a first name and role ("Sam, owner") can make a reply feel human. Decide a policy once and tell the tool.

Cause five: the voice drifts over time

A tool that handled your voice well last week may drift this week, especially if you start a new chat without the brief. Keep a saved voice note and paste it in every time. Wikipedia describes prompt engineering as designing and refining instructions to get more accurate, relevant or useful output.1 Refining is a regular habit, not a one-off job.

Edit by hand, always

A model can write smooth text that is simply wrong.3 Voice and accuracy are linked: the more a draft sounds like a template, the less likely you read it closely. Google tells the reviewer about your reply.2 That customer is reading closely, so you should too.

A quick test: read the reply out loud. Does it sound like something you'd say to the customer's face? If not, change the parts that stumble.

A voice note you can keep

We're a small, family-run business. We talk plainly. We use first names, short sentences and contractions. We never use "valued customer", "strive to" or "we apologise for any inconvenience". We keep replies to a few sentences. We don't offer anything I haven't approved.

Paste something like that at the top of every request. The rest of the routine is in the notes on building the brief and the final edit.

Reference list

  1. Prompt engineering, Wikipedia
  2. Manage customer reviews, Google Business Profile Help
  3. Large language model, Wikipedia