Most companies already have a brand voice document somewhere. It describes the tone as “approachable yet authoritative” and lists values like “clear,” “honest,” and “human.”
That document is useful for a copywriter who can interpret what “approachable yet authoritative” means in context.
It is not useful for an AI model; which needs to be told:
- Use contractions in conversational emails; avoid them in formal proposals
- Open with the reader’s situation before the company’s response
- Never use words like “leverage” or “synergy”
- End every client email with a specific question or clear next step
An AI voice guide is a translation of the brand voice into the specific, operational format an AI model can follow reliably.
The difference is between a principle (“be direct”) and a rule (“when making a recommendation, state the recommendation in the first sentence before the rationale”). This article builds the AI voice guide section by section.
Why a brand voice guide is not an AI voice guide
A traditional brand voice guide is built to be interpreted by a skilled writer who understands the context behind the principles.
“Conversational but credible” is useful guidance for a copywriter who has read the company’s existing communications and can translate the principle into appropriate output.
An AI model has no such body of knowledge. It applies a generic interpretation; which produces outputs that are generically conversational and generically credible, indistinguishable from every other company that described itself as “conversational but credible.”
What an AI model can follow reliably:
- A list of specific words and phrases to use and avoid
- A structural rule for a specific output type (“client emails open with the client’s situation, not a greeting”)
- A negative example with an annotation (“this version sounds too formal: note the passive voice and the absence of any direct statement”)
- A model example with an annotation (“this version works: direct opening, single recommendation, specific next step”)
The translation from principle to rule:
| Brand guide principle | AI voice guide rule |
|---|---|
| ”Approachable but credible” | Use contractions in all email communications. Avoid academic or bureaucratic phrasing. Open with the reader’s context, not the company’s credentials. |
| ”Direct and decisive” | State the recommendation in the first sentence. Use active voice. Do not present options when one option is clearly right. |
| ”Human, not corporate” | Never open an email with “I hope this email finds you well.” Avoid leverage, synergy, holistic, robust, and similar corporate vocabulary. |
| ”Concise” | Emails under 200 words unless the content requires more. One main point per email. No paragraph longer than three sentences. |
The brand principle is a one-line summary. The AI voice guide rule is the specific operational instruction the model can follow.
Section 1: Tone register: the foundation
The four register questions
Question 1: What is the relationship dynamic?
Is the company an advisor speaking to a client? A peer speaking to a fellow professional? An expert educating a student?
Example: “We write as peers. Our clients are founders and COOs who are operating at the same level of professional sophistication we are. We do not write down to them. We write the way a highly competent colleague would write to another.”
Question 2: What register varies by context?
| Context | Register adjustment |
|---|---|
| Client proposal | Formal structure; first-person and direct; no passive voice |
| Client email (routine) | Conversational; contractions permitted |
| Client email (sensitive) | Direct without being cold; acknowledge difficulty before resolution |
| LinkedIn post | Peer-to-peer; observation-first, then implication |
| Internal documents | Plain; no performance for an audience |
Question 3: What does the voice signal about the company?
Write this as a test: “After reading a piece of communications from this company, the reader should feel [specific feeling] and believe [specific thing about the company].”
Complete the sentence honestly. The AI uses this as a calibration signal.
Question 4: What the voice is not
“Our voice is not: formal to the point of being distant, jargon-heavy, self-congratulatory, defensive, hedged, or enthusiastic in a way that feels performative. If an output sounds like a corporate press release, a consultant trying to impress, or a sales pitch from someone who does not know the product; rewrite it.”
This is often the most useful section. The anti-examples define the edges more precisely than any positive description.
Section 2: Vocabulary standards: the most valuable section
Why vocabulary standards outperform tone descriptions
An abstract tone description requires interpretation. A vocabulary list requires none.
“Avoid corporate jargon” is an instruction the AI interprets generically. “Avoid: leverage, synergy, holistic, robust, cutting-edge, seamlessly, impactful, best-in-class, game-changing, and value-add” is an instruction the AI can follow precisely.
The “we use” list (15–25 items)
Words and phrases characteristic of the company’s voice that the AI would not use by default. For each entry: the word or phrase, and a context note where usage is specific.
Example entries for a professional services firm:
- “Compound”: referring to how well-built systems improve over time
- “Judgment layer”: the decisions that humans, not AI, must make
- “Desk work” and “room work”: to distinguish AI-appropriate tasks from judgment tasks
- “Specific”: the quality standard; used as an adjective to mean precise and concrete rather than vague
- “The foundation”: referring to the context and documentation layer
The “we avoid” list (15–30 items)
Words and phrases the company consistently edits out of AI outputs. For each entry: the word or phrase, and why avoided plus what to use instead.
Example entries:
- Transform / transformation: use “change how [X] operates” or name the specific change
- Leverage (as a verb): use “use” or name the specific application
- Seamlessly: use a specific description of the transition; or simply omit
- Robust: use a specific quality claim
- Game-changing: use a specific outcome claim
- AI-powered: name the specific capability
- Cutting-edge: omit or name the specific capability
- Holistic: name the specific scope
- Best-in-class: use a specific proof point
- Disrupt / disruption: name the specific change
- Solutions (as a vague noun): name what is actually being provided
The structural phrases list
Separate from the vocabulary list: specific phrases the company uses for common situations.
- Standard ways the company introduces a recommendation (“The approach we’d suggest…” vs “We recommend…” vs “Our read is…”)
- Standard ways the company closes a communication with a specific next step
- Standard ways the company opens a follow-up email after a discovery call
Section 3: Structural rules by output type
Structural rules govern the architecture of specific output types; not the vocabulary, but the format and sequence.
They directly address the format editing that consumes team time. The account manager who spends fifteen minutes restructuring every AI proposal is editing for structure; the structural rules eliminate that edit.
The format for each output type:
OUTPUT TYPE: [e.g., Client proposal]
OPENING:
- Length: [word count range]
- Content: [what goes here — NOT the credentials opening; specifically what the company's
proposals open with]
- Example opening line: [write one]
STRUCTURE:
- Section 1: [name] — [what it contains; how long]
- Section 2: [name] — [what it contains; how long]
- [Continue for all sections]
CLOSING:
- What it accomplishes: [secure a specific next step; summarise the recommendation]
- What it does not do: [restate everything; use a generic CTA]
- Example closing: [write one specific to this company's style]
LENGTH: [total word count range]
SPECIFIC RULES FOR THIS TYPE:
- [Rule 1]
- [Rule 2]
- [Rule 3]The four to six output types to cover:
For most $5M–$25M companies: client proposals, routine client emails, difficult client emails, sales or prospect emails, LinkedIn posts, and internal operational documents.
Cover only the output types the team produces regularly. The voice guide should not be an exhaustive style manual.
Section 4: Negative examples with annotations
Why negative examples work
An AI model shown a bad output and told specifically why it is bad will avoid that failure mode more reliably than one shown only good outputs and told what makes them good.
The negative example creates a clear exclusion. The annotation explains the principle behind the exclusion.
Format for each negative example:
OUTPUT TYPE: [e.g., Client follow-up email]
BAD VERSION:
[The full text of a generic or off-brand version]
WHY THIS DOES NOT WORK:
- [Annotation 1 — "Opens with a pleasantry rather than the client's situation"]
- [Annotation 2 — "Uses 'leverage' and 'synergy' — avoid both"]
- [Annotation 3 — "Ends with vague 'looking forward to hearing from you' rather
than a specific next step"]
WHAT TO DO INSTEAD:
[One sentence describing the specific correction for each annotation]How many to include: two to three for the most common output types (client emails, proposals). One for less frequent types.
How to generate them: run the most common prompts without the voice guide loaded and capture the outputs. These generic AI outputs are exactly the negative examples needed. They represent what the AI produces in the absence of guidance. Annotate them with the specific reasons they are off-brand; the negative examples are complete.
Section 5: Annotated model outputs
Format for annotated model outputs:
OUTPUT TYPE: [e.g., Client proposal — opening section]
MODEL OUTPUT:
[The full text of a strong, on-brand example]
WHY THIS WORKS:
- [Annotation 1 — "Opens with the client's specific operational challenge, not credentials"]
- [Annotation 2 — "First recommendation stated in sentence two; no preamble"]
- [Annotation 3 — "Uses characteristic vocabulary: 'compound' and 'judgment layer'"]
- [Annotation 4 — "Closes with a specific next step the client can confirm or adjust"]How many to include: two to three for the most important output types.
Use existing outputs the company has produced that represent the voice at its best. Real examples are more credible to the AI than examples written for the purpose.
The pairing rule: every negative example should be paired with a model output of the same type. The contrast between the two calibrates the AI more precisely than either alone.
Testing and updating the voice guide
The completion test (run before using in production)
Test 1: Load the voice guide and ask the AI to produce a follow-up email to a new prospect from a company in the company’s target industry. Compare the output to the model examples. If it is on-brand without additional prompting; the voice guide is working.
Test 2: Load the voice guide and deliberately ask for an output in a register the company avoids (formal, corporate, jargon-heavy). The AI should resist or produce something clearly outside the company’s voice. If it complies without friction; the negative examples and avoidance vocabulary need strengthening.
Test 3: Give the voice guide to a team member and ask them to describe the company’s voice in their own words based on reading it. If their description matches the intended voice; the guide communicates clearly. If it does not; the guide needs revision.
The update trigger
Every time a team member edits an AI output for a voice reason; adding vocabulary the company uses, removing a phrase it avoids, restructuring an opening; that edit is a signal.
If the same edit appears three or more times across different team members: it identifies a gap in the voice guide. The relevant section is updated.
One person owns the voice guide: the person who writes best in the company’s voice and who reviews client-facing communications regularly. Without a named owner, the voice guide degrades as the business evolves.
Common questions on building an AI voice guide
”Can I use an existing brand voice guide as the starting point?”
Yes; use it as raw material, not as the finished document.
The process: take the existing principles and translate each one into AI rules using the principle-to-rule format above. For every abstract principle in the brand guide, ask: “What specific instruction would I give an AI so it follows this principle correctly?” The answer is the AI voice guide entry.
”How long should the AI voice guide be?”
The complete AI voice guide (all five sections with two to three model examples and negative examples): 2,000–3,500 words.
Shorter than this usually means the vocabulary guide is thin, the structural rules are missing output types, or the examples are absent. A voice guide under 1,500 words is likely incomplete.
”Should different team members have access to different versions?”
The core voice guide (sections 1–3) is shared across the team. The model examples and negative examples may vary by role; the account manager’s model email differs from the project manager’s status update.
In practice: one shared voice guide covers 80% of team use. Add role-specific model examples to the relevant workflow documents rather than creating separate voice guides.
”How do I handle the fact that different people in the company write with different voices?”
Choose one voice; the best version of the company’s communication register, not the average. The voice guide is aspirational in the sense that it describes the company’s best work, not its typical work.
The person who writes it: the founder, the COO, or whoever produces the client communications that best represent how the company wants to be perceived. That person’s instincts and habits are the ones to systematise.
”Does the voice guide work with both Claude and ChatGPT?”
Yes. The voice guide is a plain-text document; it loads into any AI tool. The specific vocabulary and structural rules are instructions that any capable AI model follows.
Minor calibration differences: Claude and ChatGPT apply the same instructions somewhat differently. The first time the voice guide is loaded into a new tool, run Test 1 above and make minor adjustments to the specific instruction phrasing if needed.
”What is the fastest way to build a voice guide if I’m under time pressure?”
Minimum viable version in 90 minutes:
- Write the “we avoid” vocabulary list (30 minutes; the easiest section to produce and highest immediate impact)
- Write structural rules for the two highest-frequency output types (30 minutes)
- Find and annotate one negative example and one model example for client emails (30 minutes)
Load these three sections into the workspace immediately. The outputs will improve. Build the remaining sections over the following two weeks.
Want the voice guide built to the standard that produces on-brand AI outputs from day one?
An AI voice guide is the operational translation of the company’s brand voice into the specific format an AI model can follow reliably.
Six sections. Three to four hours to build. The output is a document that; loaded into the shared workspace; produces on-brand outputs from the first session for every team member who uses it.
The editing for voice that currently consumes twenty minutes of every AI draft review drops to five minutes or less.
Path one: start with the vocabulary standards section today. Write the “we avoid” list first; it takes 30 minutes and produces immediate improvement in every AI output. Add the “we use” list. Load both into your Claude Project. The before/after on any client email will tell you immediately whether the vocabulary guide is working.
Path two: bring in a partner. If you want the voice guide built to the standard that produces on-brand outputs from day one; through a structured interview and writing process, tested against real outputs before delivery; that is part of the Phos AI Labs Phase 1 Foundations engagement. We have run 400+ AI engagements. Clients include Zapier, Coca-Cola, Medtronic, Dataiku, and American Express. Thirty minutes, no deck. Start here.