Who’s Guarding Your Brand Voice While AI Does the Talking?

You are renting your voice

When you hand your brand's tone to a generative artificial intelligence (AI) model, you are not building an identity. You are renting a temporary mathematical state.

OpenAI publishes deprecation notice periods for its models, so the model writing your voice today will one day be switched off. Before that, its behaviour drifts. Chen, Zaharia and Zou found GPT-4 following one formatting instruction 99.5% of the time in March 2023, and 0.5% of the time in June. Harvard Data Science Review

Meanwhile your brand now has two audiences: people, who respond to emotion, and AI systems, which look for evidence. Let the machine write without a protected human core and you fall into the "helpful assistant" trap: polite, interchangeable and forgettable.

IA voice is a list of refusals

Most brand voice documents are lists of adjectives. A real voice is a list of things you will not say, even when they would work.

Approachable. Innovative. Trusted. Bold. I have read near-identical versions of that list at a telecom operator, a software firm and a bank. My test takes four seconds: could your closest competitor sign your voice document without changing a word? If yes, you have a mood board.

Mailchimp's public style guide shows the difference: "You have the same voice all the time, but your tone changes." Its rules cost it something. "We are plainspoken" rules out the punchy metaphor. "Our humour is dry" rules out the joke that would get more engagement.

This is where AI comes in. A model's untended output is the statistical middle of your category. It does not flatten your voice. It returns the category's voice, which is what you get when nobody has decided anything. In twenty-five years I have watched desktop publishing, digital, social and marketing automation each promise to do the thinking. Each applied decisions brilliantly and could not make one. AI is the most capable version of that pattern. It has not changed it.

IIThree brands that held the line

None of these is a story of creative brilliance. Each is a story of restraint, held for decades.

Amul. The topical hoarding, created in 1966 by Sylvester da Cunha and Eustace Fernandes, holds the Guinness World Record for the longest-running advertising campaign. Sixty years of Indian news offered ten thousand chances to turn serious, sanctimonious or tearful. It stayed wry. Somebody, every decade, said no to the more dramatic version.

Mailchimp. Its voice is written down in rules a new writer can apply on day three. Most firms I work with keep theirs in the head of one eleven-year veteran. That is a single point of failure, and it is why their AI output wanders. You cannot hand a machine a voice that exists only as instinct.

Zerodha. Its voice is defined by what it does not do. "When we started @zerodhaonline, we didn't have the money to advertise, not that it would've worked," wrote Nithin Kamath. In a category built on urgency, it writes like a friend explaining a risk.

IIIWhere it can be done better

Changing a voice is allowed. Becoming less distinctive while claiming to be more is the failure.

Jaguar. On 19 November 2024, Jaguar relaunched under the line "Copy Nothing", and two measurement firms tracked the fallout (Marketing-Interactive). The defence is fair: Landor's Chris Moody said, "It's not for the existing customer base, that's the point." But Graham Hitchmough of The Bonsey Design called the film "entirely derivative" for a brand claiming to copy nothing. That is a voice problem, not a visual one.

Most of B2B technology. Open five enterprise software websites, cover the logos, and try to tell them apart. Transform. Unlock. Seamless. End-to-end. Powered by AI. AI did not create that sameness. It removed the last brake on it, which was that generic copy used to be slow to write.

IVWhere AI genuinely helps

Every real benefit depends on one condition: the voice already exists, in writing.

A caveat first. Most published evidence on AI and brand consistency comes from vendors selling brand-consistency tools. So treat this as my argument from experience, not as research.

It enforces better than people do. A junior writer drifts back to their own style within a fortnight. A model does not. It is a strong second reader. Asking it which paragraph breaks your "no unprovable claims" rule is easier than asking it to write one that doesn't. It makes localisation survivable. One personality across English, Hindi and Tamil is now a tractable problem.

Audiences are not simply rejecting it either. Of Coca-Cola's earlier AI ad, its global head of generative AI, Pratik Thakar, said "consumer engagement was very high" (Campaign).

VWhere it goes badly wrong

Three public failures. In none of them did the technology fail at its task.

DPD, January 2024. After a system update, its support chatbot swore and called DPD "the worst delivery firm in the world", even though it was told not to curse (Fortune). The rule was written down. After the update, it stopped binding.

Google, August 2024. The "Dear Sydney" Olympics ad showed Gemini writing a child's fan letter. Critics called it bleak, and Google pulled it while noting "the ad tested well" (CNN). Testing tells you whether people understand and like an ad. It does not tell you whether the ad contradicts what the brand believes.

Coca-Cola, November 2025. Its AI remake of "Holidays Are Coming" drew nearly 1,900 YouTube comments, many objecting to AI, under a tagline that reads "Real Magic" (Campaign). The extra wheels on the trucks will be fixed. The clash between medium and voice will not.

VIThe gap nobody owns

The voice is decided in one place, executed in another, and noticed nowhere.

Strategy decides the voice. A vendor's system executes it, and updates or retires that system on its own schedule. Spotting drift is on nobody's objectives. That is the Fragmentation Tax: the cost of the gap lands where no one is measured.

This goes well beyond marketing. A Boston Consulting Group (BCG) survey of 1,330 senior leaders found 42% of companies expect to give AI agents genuine autonomy by 2030. Only 5% have the full set of controls in place today, from oversight and rollback gates to security, audit and cost guardrails. I would add one more, the one only marketing will own: a check that we still sound like us.

VIIFour things to do before the next release

None of this is expensive. All of it is unglamorous, which is why it does not get done.

Each step closes part of the gap in Section VI. The first gives the machine a decision to apply. The other three make sure somebody notices when it stops applying it.