AI Brand Voice Guidelines

Why most voice guides fail with AI

A typical voice guide says the brand is "confident, friendly and expert". A person who has worked at the company for a year can turn that into sentences, because they have absorbed hundreds of examples. A model has absorbed none of yours; it interprets "confident, friendly and expert" the way the average of the internet does, which is how your copy ends up sounding like everyone else's.

So an AI-ready voice guide replaces adjectives with behaviours (what the writing does) and examples (what it looks like). Anthropic's prompting documentation describes examples as one of the most reliable ways to steer a model's tone and structure, and recommends several varied examples rather than one. That principle drives the template below.

Voice versus tone

Keep the distinction Mailchimp's content style guide popularised: voice stays the same, tone changes with the situation and the reader's state of mind. Your voice is who you are; your tone is how you speak to a worried customer compared with a delighted one. A guide that does not separate them produces either a rigid voice that jars in a complaint reply, or a tone so flexible that there is no voice left.

The template

Copy this structure into a document. Every section is written so it can be pasted into an AI tool as-is.

AI-ready brand voice guide template

Section	What to write	How much
1. Who we write for	The reader, what they already know, what they are worried about, the words they use for their problem	One short paragraph per audience
2. Voice traits	Three traits. For each: what it means in behaviour, a 'this, not that' pair of sentences, and the most common way it goes wrong	Three traits, never more than four
3. Tone by context	How each trait flexes for sales pages, social posts, support replies, bad news and so on, scored on the four tone dimensions	A table, one row per context
4. Words we use	Preferred terms, product names spelled correctly, customer vocabulary	A list
5. Words we avoid	Jargon, clichés, competitor terms, phrases that sound like AI filler	A list, with the replacement for each
6. Sentence rules	British or US English, sentence length, contractions, headings style, how numbers and dates are written	Five to ten rules
7. Proof rules	Which claims we can make, where the evidence lives, what we never claim	A short list with links to the evidence
8. Exemplars	Three to five approved pieces in different formats, each with one line on why it works	Full text, not links
9. Anti-examples	One or two pieces that are wrong for us, with what is wrong	Short extracts

Filling in the voice traits

Each trait needs a behaviour and a contrast, or a model will default to the generic reading. For example:

• • Plain-spoken. We say what we mean in the first sentence. This: "Your invoices are late because approval sits with one person." Not that: "There are a number of factors that may contribute to delays in the invoicing process." Goes wrong when: plain becomes blunt in a support reply.
• • Specific. We give the number, the name or the example. This: "Three of the five steps can be automated." Not that: "Much of the process can be streamlined." Goes wrong when: we cite a number we cannot source.
• • Warm without gushing. We acknowledge the reader's situation once, then help. This: "That is a frustrating week to have. Here is what to do first." Not that: "We completely understand how incredibly stressful this must be for you!" Goes wrong when: every message opens with sympathy.

Tone by context

Use the four tone dimensions from Nielsen Norman Group (funny to serious, casual to formal, irreverent to respectful, enthusiastic to matter-of-fact) and give each context a position. Numbers are more useful to a model than adjectives.

Tone by context (illustrative scores, 1 = first word, 5 = second word)

Context	Funny to serious	Casual to formal	Irreverent to respectful	Enthusiastic to matter-of-fact
Social post	3	2	3	2
Sales page	4	3	4	3
Support reply	5	3	5	4
Bad news or apology	5	4	5	5

Before and after: what the guide changes

Illustrative example, not a client. A small commercial cleaning company asks an AI tool for a LinkedIn post about switching to evening cleaning.

Before (prompt: "Write a LinkedIn post about our evening cleaning service. Tone: professional and friendly.")

In today's busy world, a clean workplace is more important than ever. At CleanCo, we're thrilled to announce our new evening cleaning service, designed to seamlessly fit around your business needs. Our dedicated team is committed to delivering exceptional results, so you can focus on what matters most. Get in touch today!

After (same request, with the voice guide, two exemplar posts and the proof rules pasted in)

Most offices we clean are empty by 6pm, and most cleaning still happens at 7am while people are arriving. From next month we are offering evening slots instead: your team walks into a clean office and nobody has to work around a hoover. If your building has an out-of-hours access system, it takes one call to set up. Worth a conversation?

The second version follows the traits (plain first sentence, one specific detail, warm without gushing), uses the social-post tone scores, avoids three phrases from the "words we avoid" list, and makes no claim outside the proof rules. Nothing about the model changed; the input did.

Testing the guide

Do not assume the guide works; test it.

1. • Pick a brief you have already published by hand.
2. • Generate the same piece with the guide and without it.
3. • Show both, plus your published version, to two colleagues without labels and ask which sound like you.
4. • Where the guided version loses, find the missing rule or example and add it.

Two or three rounds usually expose the gaps: a missing anti-example, an exemplar that is too long, a trait nobody can define in behaviour.

Keeping the guide alive

Voice guides decay when nobody owns them. Name an owner, add every piece a reviewer has to heavily rewrite as a new anti-example with a note on why, and refresh the exemplars each quarter with your best recent work.

Human approval still matters once the guide is good. When AfM built a content agent for HEAVENSAKE, a premium sake brand, it was built from the brand's existing materials, tone guidelines and approved terminology, outputs were reviewed against the brand's standards before handover, and it was designed so every output is a draft for the brand team to approve. HEAVENSAKE chose not to deploy it into their live workflow, so it shows the design, not results. A guide improves first drafts; it does not remove the need for someone who owns the brand to sign off.

To check whether your current channels already agree with each other before writing a guide, run the five-minute brand audit first.

FAQ

How long should an AI brand voice guide be?

Long enough to include three to five full exemplars, which is usually the longest part. The rules themselves should fit on two or three pages; beyond that, contradictions creep in and models follow them less reliably.

Do I need different guides for different AI tools?

No. The same document works as pasted context in a chat assistant, as a system prompt in a custom assistant, or as the reference file in a managed workflow. Keep one master version so the tools cannot drift apart.

Should the voice guide include our competitors' language?

Include it in the words-we-avoid list if prospects confuse you with a competitor or if their terms carry a positioning you disagree with. Otherwise leave competitors out; the guide should describe you, not them.

What if our founder writes very differently from the brand?

Write two voices deliberately: the company voice and the founder's personal voice, each with its own exemplars. Mixing them in one guide produces copy that is neither.