The short version: ChatGPT and Claude default to a bland, slightly American, faintly corporate voice because that is the average of everything they were trained on, and your business is not average. Build a one page "voice document" from your own past writing and feed it in as custom instructions, and your AI drafts stop sounding like a template within a week. Most small businesses skip this step entirely, which is why so much AI-written content on LinkedIn sounds like the same person wrote all of it.
The email that made me stop and rewrite everything
Eighteen months ago I asked ChatGPT to draft a follow up email to a client who'd gone quiet after a proposal. It came back polished, grammatically flawless, and completely unlike me. It opened with "I hope this email finds you well" (I have never once said that to a client in twenty years) and closed with "I look forward to hearing from you soon" in a way that read like it was written by someone who had never met the client, because it hadn't.
I almost sent it. That's the bit that worries me, not that the AI got it wrong, but that I was tired enough and busy enough to nearly forward on words that weren't mine to someone who knows my voice well enough to spot the difference. That near miss is the whole reason this post exists.
What "voice" means (it's not three adjectives on a brand slide)
Most brand guidelines I've seen over the years describe voice as three words: "warm, professional, confident" or something equally useless. That tells an AI model nothing. Coco Chanel didn't build one of the most recognisable brand identities in history with three adjectives, she built it through hundreds of specific, repeated choices, the black, the removal of the corset, the deliberate refusal of decoration. You can read more about how that discipline played out in these business lessons from Coco Chanel, and the pattern holds for any brand voice: it's built from specific, repeatable choices, not adjectives.
For your writing, voice is things like:
- How you open an email to a warm lead versus a cold one
- Whether you use contractions (I do, constantly)
- Sentence length when you're explaining something complicated
- The specific words you'd never use (I've banned "use" and "delve" from my own writing for years, long before it became an AI tell)
- How you deliver bad news, do you soften it or say it straight
- Whether you use one liners for emphasis or build up to a point slowly
None of that fits on a brand slide. All of it fits in a document an AI model can read and use.
Building the voice document, step by step
Here's the exact process I use now, and I'll be honest that my first attempt was rubbish.
Step 1: Pull 15 to 20 pieces you've already written. Not marketing copy, your actual voice, so pull real emails, LinkedIn posts, client proposals, even a few WhatsApp messages to team members. I used newsletters going back three years, roughly 40,000 words in total.
Step 2: Paste them into one document and ask an AI model to find the patterns. I use a prompt like: "Read this collection of my writing. Tell me my average sentence length, the words and phrases I repeat, how I open pieces, how I close them, and what makes this recognisably mine rather than generic business writing." The output is a rough first draft of your voice profile.
Step 3: Edit it yourself and add the things the AI missed. This is the step that matters and it's the one I got wrong the first time. My initial voice document was about 200 words of vague description ("direct, warm, British") and the outputs it produced were still generic, just slightly less formal. It took two more attempts over six weeks before I had a document that worked, and the version that finally clicked was 900 words long with specific examples: three sample opening lines, a banned words list, and a note that says "when delivering criticism, be direct in the first sentence, don't build up to it."
Step 4: Load it as custom instructions. In ChatGPT this goes under Settings, Custom Instructions. In Claude it's a Project with a system prompt. Either way, the document sits there permanently so you're not re-explaining your voice every single time you open a new chat.
Step 5: Test it against a real task and compare drafts. I picked ten client emails I'd sent over the previous month, ran the same briefs through the AI with the voice document loaded, and compared. Once the document was right, editing time on AI drafts dropped from roughly 15 minutes per piece to about 4, which is the difference between AI saving me time and just moving the work from writing to fixing.
Step 6: Update it monthly. Your voice shifts as your business does. Add new banned words when a phrase starts feeling overused (right now I'd add "" and "" to most people's lists, they're everywhere in AI output and it shows).
The part agencies charging for "humanised AI content" don't tell you
Here's the uncomfortable bit. There's a whole cottage industry now, agencies and freelancers charging £1,500 to £3,000 a month promising "AI content that sounds human," and a lot of what they're doing is running your draft through a paraphrasing tool that varies sentence length and swaps a few words. That doesn't fix voice, it just adds noise. It's polish on top of a document that was never yours to begin with, the same problem I nearly created with that client email, just wrapped in an invoice.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
Sentence variance isn't voice. Voice is the specific, stubborn choices a real person makes over years, which is exactly why brands people remember, from Disney's storytelling discipline to GoDaddy's decision to be deliberately provocative rather than safe, are built on repetition of specific choices, not on smoothing everything into inoffensive sameness. The humaniser tools do the opposite of that on purpose, because sounding distinctive is riskier and harder to sell as a monthly retainer than sounding "polished."
Where behavioural psychology helps here
Richard Thaler's work on nudges is mostly applied to pricing pages and checkout flows, but the same idea explains why generic AI voice spreads so fast: it's the path of least resistance. Left alone, both the model and the person using it will default to the smoothest, most average option unless something specifically pushes back against it. The voice document is your nudge in the other direction, it's the friction that stops the default from winning. There's more on how small, deliberate choices shape customer behaviour in these business lessons from Richard Thaler, and the same logic applies to how you shape your own AI outputs.
Where this shows up in customer facing tools, not just emails
This isn't only about drafts you edit before sending. If you've got an AI chatbot on your site, a booking assistant, or anything customer facing that generates text on the fly, the same generic voice problem shows up there in real time and customers notice it faster than you'd think. A calm, well built tone can be a genuine brand asset, as Headspace has shown for years, but a chatbot that sounds like every other chatbot undoes that instantly. If you're using any interactive tool on your site, a chatbot, an interactive calculator, or a quiz, the same voice document should feed into whatever copy that tool generates, not just your email drafts.
When it's worth bringing someone in
If you've got the time, this is a one afternoon job, most of the cost is your attention, not money. But if you're running a team of five or more and want everyone using AI to draft in one consistent voice rather than five different interpretations of "friendly but professional," that's a system worth setting up rather than doing it piecemeal per person. That's the point where working with an AI implementation coach tends to pay for itself, someone who sets up the shared voice document, the custom instructions, and the testing process once, rather than five people each spending a Saturday figuring it out badly on their own.
What I'd tell you to do this week
Don't start with the AI tool. Start with the folder of your own writing. Pull twenty pieces, paste them into one document, and ask an AI model to find the patterns before you ask it to write anything new. Then spend thirty minutes fixing what it gets wrong, because it will get things wrong, mine took three attempts. Load the finished document as custom instructions and test it against something you've already written by hand so you have a real comparison, not a guess.
The businesses that get real time back from AI aren't the ones with the fanciest prompts, they're the ones who stopped letting the model guess who they are.
Frequently asked questions
How long should a voice document be?
Between 500 and 1,000 words works best in my experience. Shorter than that and it's too vague to change the output, longer than that and most AI models start ignoring parts of it. Mine settled at around 900 words after two rewrites.
Can I use the same voice document for every AI tool?
Yes, the content transfers, though where you paste it differs. ChatGPT calls it Custom Instructions, Claude calls it a Project system prompt, and other tools may call it something else entirely, but the document itself doesn't need rewriting for each one.
Will this stop AI detection tools from flagging my content?
That's not really the point of it and it's not guaranteed either way. The goal is content that sounds like you and reads well to a human, which matters far more for a small business than whether a detector flags it, since your actual customers aren't running your emails through a detector.
How often should I update the voice document?
Check it monthly and update it whenever you notice a phrase creeping in that feels overused, either in your own writing or across AI content generally. Words go stale fast once everyone starts using them, so treat the banned words list as a living thing, not a one time job.
Related reading: The AI Subscription Stack: What I Pay For (and What I Cancelled) in 2026 and I Fed 14 Months of Invoices Into AI. Here's What It Found That My Accountant Never Mentioned.
