AI Agent Brand Voice: Guardrails That Hold Past Week One

The first week of drafts sounds like your company. A month later the newsletter and the blog could have come from any other office that pasted a style guide into a prompt and hoped. You are about to let a marketing agent draft both, and you can already picture the moment a customer says the writing does not sound like you. A marketing agent can write in your voice, but only inside guardrails built for the two ways models slip: they miss rules they have not seen before, and they drift toward the average of what everyone else writes. The rules have to live where the agent reads them, a check has to stop a draft that breaks one, and you still read the draft before a customer does.

What does keeping an AI agent on brand voice actually mean?

A brand voice is three lists, and the agent needs all three written down. Words: the terms you always use, the ones you never use, and the spelling of your product names. Rhythm: how long your sentences run, how you open, and how you close. Claims: what you will state as fact, and what needs a named source in the same sentence. Leave any list in someone's head and the agent will guess, and the guess will sound like every other company that asked a model to sound friendly.

Tone is not the same thing as voice. Voice stays put across everything you publish. Tone moves with the moment. A refund email and a launch note should sound like the same company, and they should not be in the same mood. Nielsen Norman Group describes tone on four scales: funny or serious, formal or casual, respectful or irreverent, and enthusiastic or matter-of-fact. Place your company on those four scales, and write down how far each kind of piece may move, and the agent has a range to work inside instead of an adjective to guess at.

Why "friendly but professional" fails as an instruction.

Every company wants to sound friendly and professional, so a model given those two words writes the way every company writes. A rule the agent can follow names the behavior. Sentences under 20 words. No exclamation points. "Customers," and never "users." A question as the opening line no more than once a month. Each of those can be followed, checked, and broken on purpose when you decide the moment calls for it. An adjective cannot be checked. A behavior can.

Why does tone of voice matter enough to guard?

Because readers judge the company by it, and the difference shows up in a survey rather than in a feeling. In a Nielsen Norman Group study by Kate Moran, published August 7, 2016 and last reviewed January 30, 2024, researchers wrote four pairs of website samples, eight in all, that were nearly identical in content, layout, and detail. Only the tone and the company name changed. An online survey of 100 US adults then rated each sample on 5-point scales for friendliness, trustworthiness, and how likely they were to recommend the company.

Tone moved all three. The more casual of two bank pages was rated 0.7 points friendlier, 0.3 points more trustworthy, and 0.4 points higher on willingness to recommend than its serious twin. How trustworthy a company seemed was a strong predictor of whether people would recommend it. That is the argument for writing the tone down instead of leaving it to whoever opens the draft.

The wrong tone can cost trust while it wins warmth.

The same study found a pair that cut the other way. One insurance sample was rated significantly friendlier than its twin and significantly less trustworthy, and people were no more likely to recommend it. A humorous home security page scored well in the survey, yet three people in the in-person sessions disliked the humor, and one called the headlines corny. Tone has to fit the company and the moment.

Why isn't a style guide in the prompt enough?

A prompt is a request, and models follow familiar requests far better than new ones. IFBench, from Valentina Pyatkin and colleagues and presented at NeurIPS 2025, tested exactly that gap. The most popular test of checkable instructions, IFEval, has 25 rule types, and many leading models now score above 80% on it, some with as few as 2 billion parameters. IFBench wrote 58 new rules of the same checkable kind, covering counting, formatting, changes to words and sentences, and copying. Leading models such as GPT-4.1 and Claude 3.7 Sonnet scored below 50%. The researchers' reading was that most models had overfit to the small set of rules they were usually tested on, and did not carry the skill over to rules they had not seen.

Your style guide is a list of rules the model has never seen.

House rules are new rules by definition. No model was tuned on your banned words, your sentence limit, or the way you spell a product. Some rules will hold most of the time and some will slip. Keep the prompt as the first layer, and have a script check anything a script can check before you read the draft.

Why do AI drafts drift toward sounding like everyone else?

Two studies found the same pattern from different directions. In a 2024 study in Science Advances (volume 10, issue 28, article eadn5290; open copy on the UCL page), Anil Doshi and Oliver Hauser gave some writers story ideas from a large language model. Those stories were rated more creative, better written, and more enjoyable, especially among the less creative writers. They were also more similar to each other than stories written without that help. Each writer was better off, while the group produced a narrower range of work.

A controlled experiment by Vishakh Padmakumar and He He, presented at ICLR 2024, tested co-writing directly. People wrote argumentative essays alone, with the base GPT-3 model, or with InstructGPT, a version tuned on human feedback. Writing with InstructGPT, and not with GPT-3, produced a statistically significant drop in diversity. Essays by different people grew more alike, in their words and in what they said.

Each draft passes review. The set does not.

This is the drift a one-draft-at-a-time review misses. Every draft can read well on its own and still share its openings, its transitions, and its favorite phrases with the last twenty drafts, and with pages your competitors publish from the same models. You only see it when the drafts sit side by side.

The model's words converge. Yours do not.

The ICLR study also found where the sameness came from: the text the model contributed. The text the writers added themselves stayed as varied as before. Feed the agent phrases customers use on calls and in reviews, notes from the founder, and your best past pieces, and let the model arrange that material rather than supply the sentences.

What goes in a brand voice rulebook an agent can follow?

Keep it to one file the agent reads before every draft, in three parts. The same file should cover every format, from newsletter drafts to social posts cut from your blog, so the voice does not split by channel.

Part one: rules a script can check.

A banned-word list of 10 to 25 terms, including the filler your industry overuses. Punctuation rules, such as no exclamation points. A sentence-length ceiling. The exact spelling of your product and team names. And one rule for claims: no number without a named source in the same sentence. If a script can say yes or no, it belongs here.

Part two: examples the agent imitates.

Three to five before-and-after pairs taken from your own editing. Examples teach rhythm in a way adjectives cannot, because the agent can see the distance between the two versions.

How to write one before-and-after pair.

Take a sentence the agent wrote that you had to fix, and put your rewrite next to it. Under the pair, write the one rule the rewrite shows, in a single line, such as "Lead with the customer's problem, not our feature." One pair per rule. Retire a pair once the agent stops needing it.

Part three: calls a person makes.

List the moments the agent drafts and never decides: replies to a complaint, news about prices or policies, humor, anything that names a customer, and any claim about results. The agent writes the draft and flags it, and you pick the tone. A price, a promised time, a discount, or a new claim about the business waits until you approve it.

Which voice rules can a machine check before a person reads the draft?

Everything in part one, on every draft, not on a sample. A draft that breaks a rule fails, the check names the exact line, and the agent rewrites it before anyone on your team sees it. A warning a busy reviewer can wave through is not a gate.

What the gate checks on every draft.

• Banned words and phrases, matched as whole words
• Punctuation you have ruled out
• Sentences over your length ceiling
• Product and team names spelled any other way
• A number with no named source in the same sentence
• An opening line that repeats one from the last 20 approved drafts

This blog runs a gate like that on every post, this one included. An em dash, or any word on a 25-word banned list, fails the automated check, and the post cannot go live until the line is rewritten. The same gate works on landing page drafts and on email. It runs in seconds, and it does not get tired on the fortieth post of the quarter.

What stays with the reviewer.

With the mechanical rules handled, you can spend the read on what only a person can judge: whether the tone fits this reader today, whether the argument holds, and whether the piece says something your competitors are not already saying.

How do you catch voice drift after the first month?

Check the set once a month, not single drafts. It takes about 30 minutes and a spreadsheet. Single drafts will keep looking fine. The month is where the sameness shows up.

A 30-minute drift check you can run today.

Pull the last 20 drafts the agent wrote and a person approved, plus 20 of your best pieces written before the agent. Copy the first sentence of each into two columns. Mark every opening that shares a shape with another, such as a question to the reader or a sweeping line about the industry. Then list the three-word phrases that show up in 5 or more of the agent's drafts. A phrase that appears in a quarter of the agent's drafts and in none of yours goes on the banned list, and a repeated opening shape becomes a rule in part one.

An illustrative example: if 7 of 20 agent drafts open with a question and 2 of your 20 do, the agent has picked up a habit your team does not have, and one new rule fixes it.

The three numbers to write down each month.

Repeated openings out of 20. Phrases added to the banned list. Drafts where the reviewer rewrote more than half the text. All three should fall over the first quarter. If the third one rises, the rulebook is missing a rule, and the reviewer's edits are where to find it.

The first 30 days.

Week one, write the rulebook from your 20 best pieces. Week two, run the agent and the gate alongside your normal process, and publish nothing the agent writes. Week three, send drafts that pass the gate to a reviewer. Week four, run the first drift check. If the newsletter and the blog have already started to sound like everyone else's, Get your free Systems Plan and send us the three pieces you most want the drafts to resemble. The plan arrives by email a few minutes after you send it, with the first rules we would put in the file the agent reads.

Common questions.

How do you keep an AI agent on brand voice?

Write the voice down as rules the agent reads before every draft, run an automated check that blocks any draft breaking a rule a script can test, and keep a person on the final read for tone and judgment. Then check the set of drafts once a month for drift, because AI-assisted writing tends to converge. A prompt alone is not enough: on 58 new checkable rules, leading models scored below 50% (IFBench, NeurIPS 2025).

Is putting your style guide in the prompt enough?

No. A prompt is a request, and models follow familiar rules far better than new ones. Many leading models score above 80% on IFEval, the most popular instruction test, yet models such as GPT-4.1 and Claude 3.7 Sonnet scored below 50% on the 58 new rules in IFBench (NeurIPS 2025). Your house rules are new to every model, so a script should check the ones it can on every draft.

Why does AI-written content start to sound the same?

Because the model pulls every draft toward the same center. In a 2024 Science Advances study, stories written with AI ideas were rated more creative and better written, but were more similar to each other than stories by humans alone. At ICLR 2024, Padmakumar and He found that co-writing with InstructGPT made essays by different people more alike, and that the sameness came from the model's text, not the writers' own.

Which brand voice rules can be checked automatically?

Any rule with a yes-or-no answer: banned words and phrases, punctuation you have ruled out, a sentence-length ceiling, the exact spelling of product and team names, numbers without a named source in the same sentence, and openings that repeat recent drafts. Run the check on every draft as a gate, not a warning, so a person only reads drafts that already pass. Tone, humor, and anything about a named customer stay with the reviewer.

How often should you review an AI agent's brand voice?

Review each draft before it publishes, and review the set once a month. The monthly drift check takes about 30 minutes: compare the first sentences and the repeated three-word phrases of the last 20 approved agent drafts against 20 of your best pieces. A phrase that shows up in a quarter of the agent's drafts and in none of yours goes on the banned list.

Where these figures come from.

Kate Moran's Nielsen Norman Group study (August 7, 2016, last reviewed January 30, 2024) used four pairs of samples, eight in all, and a survey of 100 US adults on 5-point scales. It is the source for the 0.7, 0.3, and 0.4 point gaps, the insurance pair that gained friendliness and lost trustworthiness with no significant difference in recommendation, and the home security page where three people disliked the humor and one called the headlines corny.

IFBench (Pyatkin and colleagues, NeurIPS 2025) is the source for models above 80% on IFEval's 25 rule types, including some with as few as 2 billion parameters, and GPT-4.1 and Claude 3.7 Sonnet below 50% on 58 new rules. Doshi and Hauser, Science Advances, 2024 (open copy: UCL), and Padmakumar and He, ICLR 2024, are the sources for the sameness findings, including the InstructGPT result against base GPT-3. The 7-of-20 opening example is illustrative.