Give every client a written voice card: three or four traits, each defined as "this, not that", vocabulary rules, banned phrases and three to five approved sample paragraphs. Keep it in one shared AI project per client, score every draft against a short rubric, and make one named person responsible for each client's voice. Samples do more work than adjectives.
Most agencies start with a list of adjectives, and that's why the copy drifts. "Warm, witty and authentic" describes half the brands in any town. A model reading those words falls back on its default register, a polished, upbeat marketing voice that sounds like everyone. What pins a voice down is contrast ("confident, not boastful"), concrete word choices and real examples the client has already approved.
Why AI copy slides away from a client's voice
Four causes account for nearly all of it, and each has a different fix:
- The model's default voice. Without strong examples, output regresses to the average of marketing copy it has seen. Fix: samples, not adjectives.
- Different writers, different prompts. Three writers describe the same client three ways in their prompts. Fix: one shared project per client, so everyone starts from the same card.
- Long conversations. In a chat that has run for weeks, early instructions carry less weight than recent messages. Fix: start a fresh chat inside the project for each piece.
- Silent client edits. The client rewrites a paragraph, the writer accepts it, and the card never learns. Fix: a change log feeding back into the card.
The first cause is easiest to see in a before and after. An illustrative new client, a small garden-design studio, sent this as its brand voice: "Friendly, professional, approachable, passionate about outdoor spaces." Every draft from that brief came back sounding like a hotel brochure. After a twenty-minute call, the traits were rewritten as contrasts: "Practical, not dreamy: talk about drainage and sunlight before mood boards"; "Expert, not intimidating: use the plant name, then say what it does"; "Calm, not salesy: no countdowns, no 'transform your garden'". The underlying character barely changed. What changed is that each trait now tells the model what to leave out, and that's where most of a voice lives.
A voice card, filled in for a craft brewery
Here is a complete illustrative card for a small craft brewery with a taproom. It fits on one page, which is the point: writers actually read a page.
CLIENT VOICE CARD: craft brewery (v4, approved by client 2 Sept)
TRAITS
- Playful, not silly: one joke per piece at most, never at the
customer's expense
- Knowledgeable, not nerdy: name hops and styles, explain them in
one clause
- Local-pub friendly, not corporate: "we" and "you", short sentences
- Proud of the beer, never boastful: describe, don't rank ("our best
ever" is banned)
VOCABULARY
- Say: beer, pint, taproom, brew day, "on tap", "in the fridge"
- Avoid: craft beverage, artisanal, curated, elevate, experience
(as a noun)
- Always: "responsibly" line on any promotion; no suggestion that
drinking improves mood or performance
RHYTHM
- Sentences mostly under 15 words; one longer one per paragraph is fine
- Paragraphs of two or three sentences online
SAMPLES (approved)
1. Product: "Hazy, soft and full of mango. This one's for sunny
evenings and long chats. 5.2%, and dangerously easy."
2. Email opener: "Brew day ran late, the new pale's in the tank, and
Friday's quiz is back. Here's the week."
3. Social: "Taproom's open from 4. Bring a mate, bring a dog, leave
the laptop."Notice the vocabulary section doing work the adjectives can't. "Artisanal" and "curated" are the words a model reaches for when it hears "craft", so banning them by name stops the most common slip. For building a card like this from scratch, including interviewing the client, see building a brand voice guide AI can follow, and for a starter banned list, AI words to avoid.
New brands are the awkward case, because a client launching next month has no approved copy to use as samples. Don't let the card run on adjectives alone. Write three versions of the same short paragraph in deliberately different registers (one chatty, one plain, one more formal) and ask the client to pick the closest and mark what they'd change. For an illustrative mobile dog groomer, the client chose the plain version, crossed out "pamper" and "furry friend", and added a line of their own: "we never rush a nervous dog". That edited paragraph becomes sample 1, the crossed-out words go on the banned list, and the card is marked "provisional" until three published pieces have been approved.
Two clients, one brief, two voices
The test of a voice card is whether the same brief comes out differently for different clients. Give both the brewery and a traditional family butcher the same task, "announce new Sunday opening hours, 10 to 2", and compare illustrative outputs from their projects:
| Craft brewery | Family butcher | |
|---|---|---|
| Card traits | Playful, knowledgeable, pub-friendly | Plain, reliable, quietly expert, three generations in the trade |
| Output | "Sundays just got better. The taproom's now open 10 till 2, so brunch has a new best friend. Dogs welcome, as ever." | "From this Sunday we're open 10am to 2pm. Your Sunday joint can be collected fresh on the day, and orders placed by Friday will be ready and waiting." |
| What an editor would fix | "New best friend" is fine once; check it isn't the third friend joke this month | Good. Add the phone number; butcher's customers order by phone |
If both outputs sound alike, the cards are too vague. Rewrite the traits with sharper contrasts and swap in more distinctive samples before blaming the model.
A useful onboarding test builds on the same idea. Once a new client's card is drafted, generate three short pieces from it and mix them with two pieces the client wrote themselves, without labels. Ask the client which ones sound like them and what feels off in the rest. If they can't tell the AI pieces from their own, the card is ready. If they pick out the AI pieces instantly, their reasons ("too many exclamation marks", "we'd never say 'delighted'") go straight onto the card as rules. It takes one short meeting and saves weeks of vague feedback.
When one client needs more than one voice
Many clients speak differently to different audiences, and a single card flattens that. The brewery above writes to drinkers in the taproom voice, but its wholesale emails to pubs and bottle shops need to be brisk and practical: lead times, keg sizes, prices per case, delivery days. Pasting the consumer card into a wholesale email produces a sales sheet with jokes in it, which trade buyers find tiresome.
The fix is a core card plus short channel variations, each a few lines long:
- Core (always applies): vocabulary rules, banned words, responsible-drinking line, "describe, don't rank".
- Consumer channels: playful traits, one joke allowed, samples 1 to 3.
- Trade emails: no jokes, figures first, one sentence on what's new, a trade sample from an email the client's sales lead actually sent.
Tell the model which variation applies in the first line of every prompt. Without that line, it blends the two, and the blend is the voice that belongs to nobody.
The difference shows quickly. Asked for a trade email about a new stout, with "Trade emails" named in the first line of the prompt, an illustrative draft reads: "New: Winter Stout, 4.8%. 30-litre kegs and cases of 12 cans, available from 4 November. Orders in by Tuesday go out on Thursday. Price list attached." Without that line, the same request opened with "Dark as a winter night", which works on a taproom chalkboard and wastes the first line of an email a buyer is scanning for lead times.
Where the voice card lives so every writer uses it
The card only works if nobody has to remember to paste it. Put it where each piece of work starts:
- Shared projects on a business chat plan. ChatGPT Business and Claude Team both offer projects the whole team can share; put the card in the project instructions and the samples and approved past work in the files. Gemini's equivalent is a Gem (becoming a skill from November 2026). Claude also supports skills, and Anthropic publishes guidance on packaging brand guidelines as one, which suits agencies wanting the same voice rules applied across several projects.
- Specialist brand-voice tools. Jasper's Pro plan lists at $59 per seat per month billed annually ($69 monthly) and includes two Brand Voices and five knowledge assets; its Style Guide feature sits on the custom-priced Business plan. Writer is built around enforceable style guides and terminology lists and mainly sells to larger teams. A side-by-side view is in brand voice tools compared.
- Not custom GPTs. OpenAI is retiring them: they stop running on 11 December 2026, and OpenAI's migration turns a GPT into a plugin with its instructions as a skill. Agencies still running client voices as custom GPTs should move them into projects now rather than in December.
Most small agencies don't need a specialist tool if each client has a good card in a shared project. The specialist tools pay off when many writers, or many brands per client, need central control.
A quick sum helps with that choice. For an illustrative five-writer agency, ChatGPT Business Standard seats list at $25 a user a month on monthly billing, so $125 a month, or $100 a month billed annually; Claude Team Standard seats are priced the same. Jasper Pro for the same five people is 5 x $59, or $295 a month on annual billing. The chat plan is cheaper and doesn't charge per client project; before paying for Jasper, a twelve-client agency should check how many Brand Voices it would need beyond the two Pro lists and what those would add to the bill.
Prompt from the samples, then ask the model to check itself
Inside the project, the prompt can be short, but two lines make a difference: point the model at the samples explicitly, and ask it to review its own draft against the card before you see it.
Write a 120-word product description for our new stout, using the
tasting notes in this chat. Match the rhythm and word choices of the
approved samples in the voice card, especially sample 1.
After the draft, list any words or sentences that might break the
voice card, and why.An illustrative reply for the brewery:
"Dark as a winter night and smooth as you like. Roasted coffee up front, a little chocolate, then a dry finish that keeps you coming back. 4.8%, and happiest by the fire with a plate of cheese. On tap now, in the fridge from Friday."
Possible voice issues: "keeps you coming back" could read as suggesting you drink more; consider "keeps it moreish" or cut. "Dark as a winter night" is a familiar phrase; fine for this client, but check it hasn't been used on the porter.
The self-check is not a replacement for an editor, but it catches the obvious slips and, usefully, it flagged the line that sits closest to the responsible-drinking rule. The editor makes the call: cut "keeps you coming back", keep the rest.
Samples teach facts as well as voice, which catches agencies out. Suppose the brewery's email sample originally ended "and it's two-for-one on Thursdays". Months after the offer finished, a draft newsletter promised two-for-one Thursdays again, because the model treated the sample as a source of news as well as a model of tone. The editor caught it only because they had worked on the original campaign. Strip prices, dates, offers and opening times out of samples before they go into the project, replace them with a placeholder such as [offer], and keep live facts in a separate, dated file that the account lead updates each week.
Score every draft with a five-line rubric
"Does this sound like them?" is too vague to check consistently across a team. Score each draft from 1 to 3 on five lines, and don't send anything below 12 out of 15 without rework. A filled-in score for a brewery email draft:
| Line | What 3 looks like | Score | Note |
|---|---|---|---|
| Traits | All four traits visible, none overdone | 2 | Two jokes; card allows one |
| Vocabulary | No banned words; uses the client's own terms | 3 | |
| Rhythm | Sentence length and paragraphing match samples | 3 | |
| Freshness | No phrase repeated from the last five pieces | 1 | "Brunch has a new best friend" used on 12 Sept |
| Rules | Legal and compliance lines present | 3 | Responsibly line included |
| Total | 12 | Send after fixing the repeat and one joke | |
The freshness line matters more than agencies expect. AI drafts for the same client tend to reuse their own favourite phrases, and readers who see every email notice long before the agency does. The broader habits behind generic-sounding output are covered in why AI marketing copy sounds generic.
Keeping the voice stable across writers and months
- Name a voice owner per client. Usually the account's lead writer. They approve changes to the card and settle "is this on-voice?" questions.
- Log client edits. When a client rewrites a line, record the before and after in a "client changes" section of the card. Three similar edits mean the card needs a new rule.
- Refresh samples quarterly. Swap in recent approved pieces so the samples reflect where the brand is now, and note the card version.
- Run a monthly drift check. Put the last five published pieces next to the samples and ask the project: "Where do these pieces differ in voice from the samples? Quote examples." Read the answer, then decide.
The change log from step 2 needs no special tool; a short table at the bottom of the card will do. Three entries in, the brewery's might read:
CLIENT CHANGES LOG: craft brewery
12 Aug Newsletter AI: "a real flavour sensation"
Client: "tastes of toast and treacle"
26 Aug Instagram AI: "Don't miss out!"
Client: line deleted
9 Sept Product page AI: "a pint you'll never forget"
Client: "a proper pint for cold nights"
Pattern (entries 1 and 3): say what it tastes like, not how
memorable it is. Added to card v5 as a vocabulary rule.The voice owner spotted that entries one and three were the same edit in different words, and one new rule on the card replaced a correction the client would otherwise have kept making by hand.
The drift check in step 4 needs reading with care. Put to the brewery project, an illustrative reply might be: "1. Recent posts use longer sentences than the samples, often over 20 words. 2. 'Beer lovers' appears in four of the five pieces; the samples speak to 'you'. 3. The 18 Sept post contains no humour, which the card allows but which reads flatter than sample 3." Points one and two are real drift and go straight to the writer. Point three is the model over-reading the card, and the voice owner can ignore it. Check any count it quotes against the pieces themselves, because models estimate sentence lengths rather than measuring them.
A realistic slip shows why step 2 matters. For the butcher, the card said "plain and quietly expert", but over two months AI drafts started calling cuts "hand-selected" and the counter "curated". Each draft passed a quick read. The owner eventually rewrote three posts himself, replacing both words with "chosen by us" and "our counter". Once those edits were logged and the two words added to the banned list, the drift stopped. Without the log, the next writer on the account would have reintroduced them within weeks. This is also the kind of detail that belongs in the editing gate of your brief-to-approved-draft workflow.
Brand voice questions from agency teams
How many sample paragraphs does a voice card need?
Three to five is the working range. Fewer than three and the model copies one sample's quirks; more than about eight and the samples start to contradict each other. Choose pieces the client has approved recently, from different formats, such as a product description, an email opener and a social caption, so the model learns the voice rather than one format.
Should the client sign off the voice card?
Yes. Show it to them as a one-page document, ask them to mark anything that doesn't sound like them, and record their approval with a date. It turns vague feedback such as "it doesn't feel like us" into a conversation about a specific line on the card, and it gives new writers something the client has agreed to.
Does using a specialist brand-voice tool make drafts more consistent than a chat assistant?
It makes the voice easier to apply consistently across a team, because the voice is attached centrally rather than pasted by each writer. The quality still depends on the samples and rules you put in. An agency with a strong voice card in a shared project on a business chat plan often gets similar results to a specialist tool at lower cost.
Further reads
- Tone of Voice Examples: AI Copy Before and After Editing — More before-and-after examples of voice editing.
- How to Write a House Style Guide Your Team and AI Both Follow — A house style guide that people and AI both follow.
- How to Write Instagram Captions With AI That Sound Like You — Applying the voice card to short social captions.
- How Freelance Writers Use AI Without Losing Their Voice — How individual writers protect their own voice.
- How to Write AI Prompts That Sound Like Your Business — Prompt patterns that make AI sound like a business.
- How to Choose a Marketing Agency That Uses AI Responsibly — What clients look for in an agency's AI practice.
- Common AI Mistakes Small Businesses Make and How to Avoid Them — Twelve everyday AI mistakes in small firms, grouped by data, output, money and ownership, each with how to spot it and fix it in under an hour.
- AI Rollout Plan for a Ten-Person Marketing Agency — A ten-person agency's 12-week AI rollout, week by week: platform choice, client rules, two pilots, the numbers after a quarter and what went wrong.
- AI Clauses for Agency Contracts: Disclosure, Ownership, Approvals — Clause-by-clause wording for agencies using AI, with a worked campaign example and a prompt for spotting AI conflicts in a client's own contract.
- Which Agency Tasks AI Handles Well, and Which It Doesn't — Sort agency work by evidence, review effort and consequences, then test a reporting workflow with a clear human sign-off.
- How to Draft Customer Email Replies With AI That Sound Like You — AI drafts sound generic because nobody showed it how you write. Build a voice sheet from 12 of your own emails and give it the facts every time.
- How to Build a Product Glossary AI Must Use in Every Draft — How to build a product and brand glossary AI uses in every draft: eight term types, a filled-in table, a compact AI block, loading, checking and upkeep.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Jasper pricing page (Pro plan Brand Voices and knowledge assets; Style Guide on Business); Claude help page on personalisation (project instructions and skills) and Claude's brand-guidelines skill guidance; OpenAI's custom GPT retirement FAQ; Writer brand and style guide pages.