Your voice, learned once, applied forever

The problem with most AI social copy is not that it is badly written. It is that it is written by nobody. It hits the register of competent marketing prose and stops there: a rhetorical question, three tidy benefits, a closing line asking what you think. Readers have learned to scroll past that shape, and they learned it fast.

Voice is what stops that. Not tone, which is a dial you set — voice, which is the accumulation of specific choices. You open with a number instead of a question. You never use the word 'journey'. You write in fragments when you are annoyed. You name the client's industry but never the client. Those choices are legible in your existing writing, and they are exactly what a tone dropdown cannot reach.

So the brand voice layer starts from evidence rather than description. Instead of asking you to characterise yourself in an adjective or two, it reads what you have already published and works backwards to the rules you were following without naming them.

A draft can fail in two independent ways. It can sound wrong, or it can be wrong. Those need different fixes, so Bolta keeps them as separate inputs to the same generation step.

The voice profile handles sound: rhythm, vocabulary, structure, the things you refuse to say. Business DNA handles substance: what you sell, who you sell it to, the pricing shape, the proof points you are allowed to cite, the internal words you use for your own categories. Both are workspace-level, so a new agent job inherits them instead of starting cold.

This separation is what makes delegation practical. When the agentic execution layer picks up a job — repurpose last week's webinar, draft this month's competitor commentary, write the carousel from the blog post — it is not being told who you are from scratch each time. It reads both records, writes through them, and hands the result to your approval inbox.

Every decision you make in the approval inbox is a labelled example. Approving a draft confirms the current profile is close. Editing one shows precisely where it drifted and what the corrected version looks like. Rejecting one marks a pattern to stop producing. None of that requires you to write a style guide or fill in a form.

Those signals compile back into the voice card the writer reads: a working set of exemplars and explicit do-and-don't rules. Because it is in-context rather than baked into weights, a correction you make this week shows up in next week's drafts rather than after some retraining cycle. It also means you can read the rules, which matters more than it sounds — an opaque model that has 'learned your voice' is impossible to argue with.

The practical measure of whether this is working is not a quality score. It is how much you edit. In the first fortnight you will rewrite a lot of openers. If the layer is doing its job, that shrinks, and the review pass turns from rewriting into approving.

Founders and operators who are the brand, and whose credibility depends on posts sounding like them rather than like a marketing department. If someone who knows you can tell an assistant wrote a post, the post cost you something.

Consultants and small professional-services teams whose entire pipeline runs on being recognisably specific in public. And agencies running social for several clients at once, where the failure mode is not silence but sameness: five accounts that all sound like the same junior wrote them on the same Tuesday. Separate voice profiles, one per client, is the direct fix for that.