Copilot and Your Brand Voice: Why It Keeps Producing Copy That Does Not Sound Like You

I have lost count of the number of times a marketing team has said the same sentence to me in a training room. The copy is fine, it just does not sound like us.

Somebody usually adds that they pasted the tone of voice document in, and it made no difference. That bit is the clue.

Copilot is not reading your tone of voice document, and it does not remember your brand between sessions. Unless you deliberately put your own published writing in front of it, you get the average of everybody else’s.

Why does Copilot write in a voice that is not yours?

Nothing has told it what yours is.

A large language model gives you the most statistically unremarkable version of whatever you asked for. Ask for a product announcement and back comes the average of every product announcement ever written, which is why it reads as competent, bland and faintly American, which I don’t want as a Brit. That is the default behaviour of any model working without something specific to imitate.

Your voice, though, lives in the places you break from that average. Maybe you never use exclamation marks. Perhaps you always name the person instead of the department. We have one rule at Prohibition that we have never written down properly, which is that nobody gets to write “solutions”, and every new starter learns it within a fortnight. Those small refusals are your brand voice. A tool that has never read a word you published cannot see any of them.

Tom Roach made a version of this point when he came on the podcast. His team at the time let KFC drop Finger Lickin’ Good and watched Weetabix leave “Have you had your Weetabix?” in a drawer for seven years, and when they finally measured it, the things they had stopped using were the strongest things they had. The whole episode with Tom is worth an hour. The lesson transfers directly. The bits of your writing you think are too small to write down are usually the bits doing the work.

Is Copilot actually reading our tone of voice document?

Usually it is not, and this is where most teams get caught.

Microsoft now sells this as three tiers with confusingly similar names, and only the top one, Premium, can see your files at all. The current overview has the detail, but the part that matters for your voice is this.

Read that bottom row twice. Neither version remembers your voice from yesterday, which surprises people who spent twenty careful minutes explaining it last Tuesday and assumed the lesson had landed. Every new chat starts from the same blank average as the first one.

That single misunderstanding explains most of the “it was brilliant last week and rubbish this week” reports we hear.

Why does pasting the brand guidelines not fix it?

Because brand guidelines are written for humans who already have taste. They describe a feeling.

Here is the sort of line that appears in almost every tone of voice document in the country.

Our voice is confident but never arrogant. Warm, human and approachable. We speak to our audience as equals.

A person reads that and roughly knows what to do, because they arrive carrying twenty years of context. A model reads it and stays exactly where it was, because every brand on earth claims to be warm and human. The instruction carries no information. It cannot move the output, since it fails to separate you from anybody else.

What does work is behavioural and slightly rude. Use contractions. Never open with a rhetorical question. Always name the client rather than saying “the client”. State the number first and the caveat second. Rules a subeditor could enforce on a Friday afternoon.

What actually makes Copilot sound like you?

Three things, and the last one does nearly all the work.

Start with a rules list rather than an adjectives list. Ten to fifteen enforceable instructions, written the way you would brief a freelancer you did not entirely trust. Include the negatives, which are far more useful than the positives. Never use “leverage”. Never open with “In today’s fast paced world”. Never write a sentence longer than twenty five words.

Then add your banned words and what you say instead. Every organisation has this list and it usually lives in one person’s head, unwritten, enforced by sighing. Getting it out of that head is often the most valuable hour of the whole training day, and it quietly improves the human writing as well.

Last, give it three to five examples of your own published work. This is the part that changes everything. A model imitating real examples of your writing beats any amount of description of your writing. Hand it a press release you were proud of, a client email, a campaign summary, a social post, and tell it to match the register rather than describe it.

Description tells the model what you believe your voice is. Examples show it what your voice actually does on the page. The second is more accurate, partly because most brands write nothing like their own guidelines say they do.

What does a brief that holds your voice look like?

Closer to a subediting instruction than a prompt.

The structure that survives contact with a real marketing team runs like this. Role, so who is writing. Audience, so who is reading and what they already know. The enforceable rules. The banned words. Then your examples pasted in full, with an instruction to match their register and sentence length. The actual task goes last.

Put the task last. Models weight the end of a long instruction heavily, so a brief opening with “write a press release about our new office” and burying the voice rules underneath will hand you back a press release that ignores every one of them.

Save that brief where the whole team can reach it and keep editing it. It is the nearest thing to brand voice memory that either version of Copilot gives you.

Where it should differ is by channel. A brief tuned for press releases produces dreadful social copy, because the rules that make a release credible make a post unreadable. Two or three channel briefs is usually right. More than that and nobody uses any of them, which I have watched happen more than once.

When should you keep Copilot away from the writing altogether?

When the writing is the product rather than the packaging.

Anything carrying a named person’s opinion belongs to that person. Comment quoted in a release. A chief executive’s LinkedIn post. A statement during a crisis. The output would often be perfectly serviceable, and that is rather the problem, because putting invented phrasing in a real human’s mouth is a reputational bill that arrives later. The moment an audience can tell, everything around it loses credit too.

Facts are the other place to stop. Copilot will hand you a fluent, confident, beautifully toned paragraph containing a figure it made up. Fluency is not accuracy, and a voice that sounds exactly like yours makes an invented claim harder to catch, which is a genuinely uncomfortable thought.

So write the rules list this week. Fifteen minutes with whoever edits everything, because they already know all fifteen rules and have been enforcing them silently for years. Add three examples of published work. Test it against something a human wrote and compare the two honestly.

If you want the whole team working from one brief, with the licence question settled and the checking habits built in, that is what a day of Microsoft Copilot training for marketing teams is for. Teams leave with the brief written, tested and in use, rather than added to a list of things somebody will do in September.

Nobody is going to teach the tool your voice while you wait. Somebody has to sit down and write it out.