Training an AI on Your Brand Voice, Properly

Most brand voice documents say the same four things: friendly but professional, confident not arrogant, clear not corporate, human. They are useless to a model for the same reason they are useless to a new hire, which is that every company on earth would sign the same sentence. Nobody has ever written a guide saying unfriendly and unclear. What a model needs is not a description of your voice. It is your voice.

Adjectives fail for a specific reason. A model is already sitting at the average of all writing. Asking it to be professional moves it toward the average professional voice, which is where it already was. The instruction has no direction because everybody is pointing the same way.

Voice is what makes you different from the average, so it cannot be conveyed by words the average also uses.

The same document usually fails a human too, for what it is worth. A new hire asked to write in the brand voice reads the guide, learns nothing, and then goes and reads twenty things the company published, which is what they needed in the first place.

Two kinds of input work, and both are concrete.

Examples. Actual published or sent material. This does most of the work.

Checkable rules.Not adjectives, constraints somebody could verify compliance with in five seconds. Never say "solutions". Never open with a question. Contractions yes. Sentences under twenty-five words. We say customers, not clients. We do not use exclamation marks.

Rules of that kind are also the ones a human can enforce in review, which matters because someone has to.

Selection matters more than volume, and most teams pick badly by choosing their most polished material.

Fifteen to twenty is plenty for register. Coverage is what needs the attention: five situations with three examples each beats twenty examples of one situation.

This is the part almost everyone throws away, and it is worth more than the style guide.

Every time you take a generated draft and edit it before sending, the diff between those two versions is an exact, situated description of your voice. Not a description of what you like in general. A record of what you changed, in this context, for this recipient. It is produced free, it is unambiguous, and there is a lot of it.

The premise of most brand voice work is that a company has a voice. Companies have several, and pretending otherwise produces output that is wrong in at least one context.

A founder's outreach email, a support reply about a broken thing, a product announcement and an invoice reminder are four different registers. They share values, not sentences. Forcing one voice across all four makes the support reply sound like marketing, which is the single most alienating thing a company can do to someone with a problem.

The distinction worth encoding is whether the reader believes they are hearing from a person or from the company. Person-voiced messages should differ by person. Company-voiced ones should be uniform. Most teams get this backwards, standardising the personal messages and letting the company ones drift.

Not by reading a draft and deciding it seems fine. Two tests are worth actually running.

The mixed pile. Take five things you wrote and five the model wrote, in the same situation, and ask a colleague who knows your writing to sort them. If they can do it easily, ask what gave it away. The answer is almost always structural and almost always fixable, and it is more useful than any score.

The edit distance. Track how much you change before sending. This is the only metric that matters and it is free. If your edits are getting shorter month over month, the voice is being learned. If they are not, feeding it more adjectives will not help and you need better examples.

The specific tells to look for when sorting the pile are in why your AI emails sound like AI, and the limits of what any of this can achieve are in what AI still cannot do in marketing. Voice can be learned. Judgement about what to say cannot.