Send a customer an AI-written email and there's a decent chance they know. Not because they caught a hallucination. Because of the rhythm.
Something like: "I wanted to reach out and thank you for the opportunity to serve your heating and cooling needs. Our team is committed to delivering reliable, efficient, and professional service."
Nobody talks like that. You definitely don't. But that's what ChatGPT hands you when you ask it to write a follow-up email, and if you paste it and hit send, your customer reads it in the same voice they read spam.
The AI accent
Every model has a default register. It's the average of everything it read on the internet, which means it lands somewhere between a press release and a LinkedIn post. Hedged. Padded. Fond of words like "leverage," "robust," and "seamless." Allergic to a short sentence.
You've been trained to spot this without realizing it. So has your customer. The tell is accumulation. Three adjectives where one would do. A closing line about how much the company values the relationship. Zero specifics about the actual job. If you want the full list, I've written up the five patterns that give unedited AI writing away.
A better prompt won't fix this, and I say that as someone who wrote a whole framework about prompt structure. People try. They write elaborate instructions about being "conversational and friendly and authentic," and the model responds by producing the corporate voice with contractions bolted on. Slightly worse, actually, because now it sounds like a corporation pretending to be your buddy.
The model has no idea what you sound like. You gave it an adjective when it needed a sample.
Voice training, minus the jargon
The setup takes one sitting and no technical skill.
Go into your sent folder. Pull 5 to 10 emails you actually wrote: real ones, to real customers, that got the result you wanted.
Before any of it gets pasted, scrub it. Strip names, addresses, phone numbers, account numbers, and anything identifying a specific case or patient. Voice matching runs on rhythm and word choice, so a redacted sample teaches the model exactly as much as the original does.
If you run a medical practice or a law firm, go further than that. A consumer ChatGPT or Claude account comes with no BAA and no arrangement that protects privilege, which makes pasting real patient or client correspondence into one a genuine exposure rather than a theoretical one. Use a business or enterprise tier with the agreement your compliance rules require. Most state bars have published an opinion on generative AI by now. Read yours before client material goes anywhere near a chat window.
While you're in settings, look at the data controls. On the consumer tiers of both products, your conversations may be used to improve future models unless you turn that off. It's one toggle, and it's worth finding before you paste a year of customer email. If you've never checked what the free tiers do with what you paste, start with the rundown on AI data privacy for small businesses.
Then paste the scrubbed samples into a single chat with this instruction:
"Here are examples of how I write to customers. Study the sentence length and the word choices. When I ask you to draft something, match this voice exactly. Do not make it more formal or more polished than these samples."
That's the whole technique. You've replaced an abstract adjective with evidence.
Now save it so you're not redoing this on Monday. Both products have a workspace that holds standing instructions across every conversation inside it: Claude calls these Projects, and ChatGPT has Projects as well, plus Custom GPTs, which does something similar and currently requires a paid plan. Feature names and plan tiers on these things change every few months, so look for whichever of the two your account offers today. Paste the samples into the instructions once and every future conversation opens with your voice already loaded.
You're not limited to a paragraph, either. Paste until you run out of samples you're genuinely proud of. If the tool complains about length, cut the weakest ones and keep the rest. More real writing beats more careful instructions, every time.
What to feed it
Not everything you've written deserves to be a training sample. Be picky.
The follow-up email you sent after winning a job is a good one, because it's warm and it's specific and you weren't performing. So is the reply where a customer was furious about a delay and you talked them down: that message probably contains the clearest example of your actual judgment on the page. The proposal that closed at a rate above your usual number is worth including too, since whatever you did in that document is the thing you want repeated.
What to skip: anything you copied from a template. That estimate boilerplate you inherited from your old boss, the "thank you for your business" line you found in a Facebook group, the LinkedIn-flavored paragraph you pasted from somewhere. It's not your voice. Feeding it in teaches the model to imitate a stranger, and you'll end up right back where you started.
Skip your internal messages to staff, too. How you write to your lead tech at 6am is not how you write to a homeowner. Keep the samples pointed at one audience.
What to actually expect
The first few drafts will be close but not right. Maybe it's too eager. Maybe it uses "reach out" and you never say that. Fix it and tell the model what you fixed:
"Good, but I'd never say 'at your earliest convenience.' I'd say 'whenever works.'"
Here's the part nobody mentions: that correction lasts until you close the conversation. A plain chat doesn't learn. Open a fresh one tomorrow and "at your earliest convenience" is back, and you can spend a week fixing the same habit before concluding the whole approach is nonsense. So when a correction sticks, write it into the saved instructions on your Project. A running list of "never say this, say this instead" is the thing that carries forward.
You'll still read every draft before it goes out. That's how this works: AI writing is a first draft engine, not an autopilot. Anyone selling you "set it and forget it" email is selling you a way to embarrass yourself at scale. One contractor lost a client over an AI-drafted contract that contradicted what he'd already agreed to on the phone, because nobody read it before it went out.
A few honest limits worth knowing. Voice matching works well for email and proposals. It works less well for anything long, where the model drifts back toward its default around the third or fourth paragraph. And it will not invent facts about a job it doesn't know about, so if you don't give it the details, it'll write around them with vague filler. Give it the specifics.
The team angle is where this gets genuinely useful. Once you've built a voice profile that works, share it. Your office manager, whoever covers the inbox on weekends, anyone else who answers customers: they all draft from the same profile. The company sounds like one business instead of five people with different email habits. For a small shop where the owner's personality is the brand, that consistency beats any of the individual time savings.
One warning. Voice training makes it easy to produce a lot of writing quickly, and volume is not the goal. A customer who gets three well-written AI emails a week from you is more annoyed than one who gets a short human text. Use it to write the things you were already going to write, faster and better. Don't use it to start writing more things.
Try it this afternoon
Run the comparison before you commit to any of this. Take whatever's sitting in your to-do pile and ask for the same draft twice: once cold, once with your samples loaded. Put the two side by side and decide which one you'd actually put your name on.
If you'd rather not sort through it alone, book an AI Concierge session. We'll go through your real messages, build the voice profile, and show your team how to use it without turning your inbox into a robot. If the drafts themselves are the problem and not just the voice, this piece on why ChatGPT didn't work for your business covers the rest of it.