How to build a brand voice system that AI can follow

A brand voice system for AI is a compact, operable set of voice principles, vocabulary rules, worked examples and model-ready instructions. Unlike a traditional brand book, it is written to be pasted into prompts and used by editors daily. The strongest systems are grounded in audience evidence, so the voice reflects how real customers talk.

Why do brand books fail with AI?

Most brand books were written for designers and agencies. They describe a voice with adjectives, show a few hero lines and stop. A human copywriter can absorb that and extrapolate. A model cannot. It reads "bold yet approachable" and produces the same bold yet approachable copy it gives every brand that uses those words.

When we started working with GPT-3 in 2020, we learnt this quickly. The only thing that reliably moved the output was concrete material in the prompt. That lesson shaped how we build voice systems now, as we described in what few-shot prompting still teaches.

What goes into a voice system?

1. Three to five principles, written as behaviours

Not "we are confident" but "we state the answer first, then the reasoning". Not "we are warm" but "we address the reader's situation before our product". Behaviours can be checked; adjectives cannot.

2. Vocabulary: owned, preferred and banned

• Owned words: terms the brand uses distinctively.
• Preferred words: plain choices that match how customers talk.
• Banned words: clichés, category jargon and model favourites you never want to see.

3. An example bank

Pairs of brief and finished piece across your main formats: social post, product page, email, support reply. Include a few "before and after" edits showing a generic draft and the on-voice version. These teach both writers and models what correction looks like.

4. Audience language notes

A short, regularly refreshed list of phrases, tensions and objections from real audience conversations. We source these with SOMIN, an AI audience-research platform and our technology partner. Its social media manager and listening tools help surface what people actually say about a category.

5. Model-ready instructions

A condensed version of the above, formatted for a system prompt or custom instructions, tested against the models your team uses.

6. Context dials

Most voices flex by situation. Document how the voice shifts for a complaint versus a launch, or for senior buyers versus students, so neither people nor models overcorrect.

How do you build one in practice?

1. Audit. Gather recent content and score it for distinctiveness. Our voice audit guide covers the method.
2. Listen. Map the audience's vocabulary and tensions with SOMIN.
3. Draft principles. Turn the best of the brand and the clearest audience needs into behavioural rules.
4. Build the example bank. Select or rewrite examples that embody the rules.
5. Compress for models. Write the model-ready version and test it on real briefs.
6. Train and ship. Walk editors through it and connect it to your prompt library.

How do you test whether a model follows it?

Run the same ten briefs through the model twice: once with only a generic instruction, once with the voice system. Ask editors who do not know which is which to identify the on-brand versions and mark which rules were broken. You learn which rules are clear and which need examples. Repeat after major model updates, because defaults shift.

A voice system is finished when a new freelancer and a model can both produce something your editor would recognise as yours.

Checklist: is your voice system operable?

• Principles written as observable behaviours
• Owned, preferred and banned word lists
• At least two worked examples per main format
• Before-and-after edits
• Audience phrases refreshed in the last quarter
• A tested model-ready version
• A version number and an owner

Definitions

• System prompt: standing instructions given to a model before the task.
• Example bank: curated brief-and-output pairs used for training and prompting.
• Context dial: a documented shift in voice for a specific situation or audience.

Where agentic workflows fit

As more content is produced by multi-step AI workflows, the voice system becomes infrastructure: every agent that writes needs it. Our sister brand AgentC runs agentic marketing workflows and treats voice and judgment as human-owned inputs, which is the right instinct. The SM Universe case study is useful further reading on aligning content with what audiences care about.

The aim is simple: whoever or whatever writes for you, the reader should hear one voice, and it should be one shaped by them.

What are the common mistakes?

Writing for the brand team, not the user

Voice systems often read like manifestos. Writers and models need instructions they can apply in the middle of a task, so keep the language plain and practical.

Too many principles

Ten principles compete with each other and none gets followed. Three to five, ranked, give writers and models a clear order of priority when rules conflict.

No negative examples

Showing what the voice is not is often more instructive than showing what it is. A short set of rejected drafts, with notes on why, saves editors repeating the same correction.

Set and forget

Audiences shift their language, models change their defaults, and products evolve. A voice system without an owner and a review date will quietly fall out of use within a year.

Avoid these four and the system becomes something people reach for, rather than a document they were once sent.

Where should you start?

If you have nothing today, begin with a single page: three behavioural principles, a banned word list of twenty items, and four strong examples. Test it on a week of real briefs, note which rules editors still have to enforce by hand, and expand from there. A small system that people use beats a complete one that nobody opens.

Frequently asked questions

What is the difference between a brand book and a voice system?

A brand book describes the voice. A voice system makes it operable: rules, worked examples, vocabulary lists and model-ready instructions that writers and AI tools apply daily.

How long should a voice system be?

Short enough to fit comfortably in a prompt alongside a brief. A core of one or two pages, plus an example bank, is usually more useful than a forty-page document.

How often should a voice system be updated?

Review it quarterly and whenever audience research reveals shifts in how customers talk. Version it, so changes are visible and outputs can be traced to a version.