Checking AI copy against your brand voice

AI can draft copy fast, but speed only helps if the result still sounds like you. A short checklist keeps AI drafts on-brand without slowing the team down.

Most marketing teams already use AI for first drafts. McKinsey’s 2026 global survey finds that revenue gains from AI are most often reported in marketing and sales.¹ The risk is not that AI writes badly. It is that it writes smoothly in a voice that belongs to no one, and occasionally states things that aren’t true.

Why brand voice needs a check

A brand voice is a promise of consistency. Readers learn to recognize it, and that recognition builds trust. AI drafts tend to drift toward the average of everything the model has read: polite, upbeat, and generic.

The costs of skipping the check can be public. In 2023, Futurism reported that Sports Illustrated had published product reviews under author names and profile photos that appeared to be AI-generated. The publisher said the content came from a third-party partner and ended the partnership.² One “author,” Drew Ortiz, had no social media presence and no publishing history, and his profile photo was for sale on a website that sells AI-generated headshots.² The same report recalled that when the publisher first announced its move into AI, its CEO, Ross Levinsohn, told The Wall Street Journal: “It’s not about ‘crank out AI content and do as much as you can,'” adding, “Google will penalize you for that and more isn’t better; better is better.”² The lesson for any brand: the promise is easy to make, and the checking is what keeps it.

Facts are part of voice too. OpenAI describes hallucinations as “plausible but false statements generated by language models.”³ Well-written copy with a wrong claim is still wrong, and your brand takes the blame.

The five checks

1. Words. Does it use your vocabulary, and avoid the words you never use?
2. Tone. Is it as formal, warm, or direct as your guidelines say?
3. Claims. Is every fact, number, and promise true and approved?
4. Rhythm. Does it read naturally out loud?
5. Ending. Does it close with the action you want the reader to take?

How to run each check

Words. Keep a short “say this, not that” list: your product names exactly as written, the terms you prefer, and the ones you avoid. Common AI habits to watch for include stock phrases and empty intensifiers that sound enthusiastic without saying anything.

A short example of what the list can look like, for an imaginary accounting software brand:

• Say “bookkeeping,” not “financial operations management.”
• Say “saves about an hour a week” (if you have the data), not “revolutionizes your workflow.”
• Say “you,” not “users” or “customers,” when talking to the reader.
• Never use “seamless,” “cutting-edge,” or “game-changing.”

Ten to twenty lines is enough. Put the list in the prompt and on the checklist, so the AI and the editor work from the same rules.

Tone. Compare the draft with two pieces your team considers perfectly on-brand. If the AI draft is noticeably more formal, more excited, or more cautious, adjust it.

Claims. This is the check that protects you most. Every number needs a source, every promise needs approval, and every comparison needs evidence. Be careful with claims about AI itself: a Washington State University study found that mentioning artificial intelligence in product descriptions lowered purchase intentions, and regulators have acted against overstated AI claims.⁴

Rhythm. Read the draft out loud, or ask a colleague to. AI copy often has sentences of the same length and shape. Vary them, and cut anything you wouldn’t say to a customer in person.

Ending. Make sure the piece closes with one clear next step, written your way. AI drafts often end with a summary of what was just said, which wastes the reader’s last moment of attention.

Build the voice into the prompt

The checks work better when the draft starts closer to your voice. Turn your brand guidelines into a reusable prompt of a few hundred words: who you are, who you write for, five words that describe your tone, your “say this, not that” list, and two short examples of on-brand copy.

Then ask the AI to check its own work before you do: “Review this draft against the guidelines above. List any words, tone or claims that don’t match, and suggest fixes.” It won’t catch everything, but it makes the human pass faster.

Train the people, not just the prompt. Microsoft and LinkedIn’s 2024 survey found that only 39% of AI users had received AI training from their company.⁵ A one-hour session that walks the team through the five checks with real examples will do more for your brand voice than any single prompt.

Build the voice by reverse engineering your brand guidelines

If you already have brand guidelines, you don’t have to write that prompt from scratch. Give the AI your guidelines document, or at least the parts on tone, vocabulary, and messaging, and let it do the translation:

“Read these brand guidelines. Write a summary that another AI could follow to write in this voice: who we are, who we write for, our tone in five words, the words we use and avoid, and the rules we never break. Keep it under 400 words. Use only what is in the guidelines, and list anything that is unclear or missing.”

Then check the summary the way you would check any draft. Does it match how your best writers actually write? Fix what is off, add your two on-brand examples, and you have the reusable prompt. The list of gaps is a bonus: it often shows where the guidelines themselves need an update.

No guidelines yet? Reverse engineer the voice from your best work instead. Paste five pieces the team considers perfectly on-brand and ask the AI to describe what they have in common.

Build the voice as a shareable skill

A prompt tends to live in one person’s chat history. A skill packages the same voice as a file that the whole team, and the AI tools it uses, can load. The open Agent Skills format, first developed by Anthropic and since adopted by a growing number of AI products, is a simple example: a folder with a SKILL.md file holding a name, a short description of when to use it, and the instructions, plus any reference files you want to add.⁶

For brand voice, that file holds your summary, the “say this, not that” list, the five checks, and a few on-brand examples. The description tells the AI to use it whenever it writes or reviews copy for your brand. Because it is plain text, the same file works across tools, can be versioned like any other document, and is easy to hand to an agency or a new team member. Keep one owner for it, and update it whenever the list changes.

Who owns the check

Make one person responsible for brand voice in each channel. That person doesn’t have to edit every post, but they review a sample each week, update the “say this, not that” list, and add new examples to the prompt.

For high-stakes content such as press releases, pricing pages, legal claims, and anything about health, money, or safety, every piece gets a full human review. For low-stakes content, such as internal summaries or first drafts of social posts, a quick check against the list is often enough. Write down which is which, so nobody has to guess.

Finally, keep track of what goes wrong. If the same problem keeps appearing, such as a banned word, a tone that is too formal, or an invented statistic, fix it in the prompt and the checklist. Over a few months, the AI drafts get closer to your voice and the checks get faster.

Coming soon: Prompts that work: a brief template for marketers →

When to say AI helped

Most marketing copy drafted with AI and edited by people needs no label. Some content does. In the EU, the AI Act’s transparency rules, in effect since August 2026, require deepfakes and certain AI-generated text that informs the public on matters of public interest to be clearly labeled.⁷ For images and video, add AI rules to your brand guidelines now.

So what counts as “public interest”? The Act itself doesn’t define it. It does exempt text that “has undergone a process of human review or editorial control” when a person or company “holds editorial responsibility for the publication of the content.”⁸ The European Commission’s July 2026 guidelines fill in the rest. They describe matters “relevant to society at large, whether at a local, national, Union or international level, and meriting public debate or scrutiny,” such as politics, public administration, justice, public health, environmental protection, consumer safety, and “any economic, financial, political, scientific, or cultural development that may be relevant subject of public debate.”⁹

The Commission’s examples help. In scope: an AI-generated summary of a news article about a town council decision, AI-edited investor information in a listed company’s reports, and an AI-written storm warning on a weather service’s social media. Out of scope: AI-edited ads and product descriptions, unless they make claims about, for example, health, consumer safety, or sustainability.⁹ For most brands, the copy to watch is investor news, public statements on current issues, and any health, safety, or sustainability claims. There, a real human edit, or a label, matters.

Coming soon: Brand-safe AI visuals: a guideline checklist →