Brand tone of voice: how to write one that AI and suppliers actually follow

The website says "bespoke solutions", the newsletter is stiff and formal, the LinkedIn post sounds like another company and the ads were put together by AI in ten minutes. Nobody really got it wrong. It's just that nobody ever wrote down how the brand speaks.

The short answer: a tone of voice that works fits on one page, is written with examples and bans rather than adjectives, and is handed to everyone who writes, people and AI alike, before every text. If it sits in a PDF nobody opens, it doesn't exist.

Why the problem is bigger now

A few years ago a company's texts were written by a handful of people, and the tone was passed on by word of mouth. Today lots of people write: the salesperson using AI, whoever runs social media, the web supplier, the tools that generate ads and automatic replies. More texts, more hands, less control.

AI doesn't have a tone of its own that suits you. Without rules, it writes in the most average register possible: inflated sentences, strings of adjectives, "innovative" and "end-to-end". The result isn't wrong, it's anonymous. And an anonymous brand has a cost: readers don't recognise you, and a customer who doesn't recognise you compares you on price alone.

What doesn't work: the list of adjectives

Almost every tone-of-voice document we see says the same thing: "Our tone is professional, empathetic, authoritative and close to the customer." Give that sentence to two different people and ask for the same text: you'll get two different texts, and both will follow the rule.

Adjectives describe an intention, not a behaviour. Whoever writes, person or AI, needs decisions: formal or informal? How long is a sentence? Is irony allowed? Which words are never used?

How to write it: one page, six parts

1. Who you are in one sentence. What you do, for whom, and what sets you apart. Without this sentence, every text starts from scratch.
2. Three clear register choices. For example: "we write informally", "short sentences, one idea per sentence", "no exclamation marks". Each choice must be checkable by reading a text: it's followed or it isn't.
3. Words in, words out. A short list of your words and one of banned words. If you say "clients" and never "users", or "team" and never "staff", write it down. It's the part AI follows best, because it's the most precise.
4. Two right examples and two wrong ones. The same message written well and written badly, side by side. Examples teach more than any explanation, to people and to AI.
5. How it changes by channel. The tone stays the same, the register adapts: more direct on social media, more precise on the website, more restrained in emails to people who don't know you. One line per channel is enough.
6. What you never promise. Unverified numbers, guaranteed results, client names without permission. These limits protect the brand when the writer doesn't know the background.

One page, not a manual. If it takes twenty minutes of reading to work out how to write a post, nobody will do it.

Illustrative scenario. Weak rule: "Our tone is dynamic and close to the customer." Useful rule: "We write informally. Twenty words per sentence at most. Never 'solutions', never 'end-to-end'. Always start from the reader's problem."

How to make AI follow it

Writing the document is half the job. The other half is making sure it's used every time, not just in the first week.

• Give it first, not after. The tone document should be pasted or uploaded as a standing instruction in the tool you use, before you ask for the text. Fixing an anonymous text afterwards costs more than getting it written well from the start.
• One document for everyone. The same file for people writing in-house, for external suppliers and for AI. If everyone has their own version, you'll be back to three brands.
• A check before it goes out. Every text that reaches the public is reread by a person, with the document alongside. Three questions: does it follow the register choices? Does it use a banned word? Does it promise something we can't prove?
• Update it when you get it wrong. Every time you correct the same mistake twice, it becomes a new rule. The document improves with use.

It's how we work too. Before writing, our AI agents that prepare content read the CASTE Strategy tone rules, with the words to use and the ones to avoid. And before a text goes out, there's a check. It isn't a promise: it already happens here.

Where to start this week

1. Collect ten recent texts from different channels: website, emails, social media, ads, quotes.
2. Read them one after another. Do they sound like the same company? Mark the differences that bother you.
3. Pick the two texts that sound most like you. They become the right examples in the document.
4. Write the page with the six parts above. An imperfect draft today is worth more than a perfect manual in six months.
5. Test it. Give the document to the AI you use and ask for the same text as before. If the result sounds more like you, it works. If not, you're missing examples or bans.

In short

• The more people and tools write for you, the more you need a written tone of voice.
• Adjectives aren't enough: you need checkable choices, words in and words out, right and wrong examples.
• The document fits on one page and is the same for people, suppliers and AI.
• It's given before every text and checked before the text goes out.
• Every mistake corrected twice becomes a new rule.

If you want to know whether your brand speaks with one voice or three, write to us: on the first call we read a few of your texts together and tell you where to start.