Voice is the product, not the packaging

When a client picks your firm over a competitor, they're not usually picking on methodology (the methodologies are similar). They're not picking on pricing (the pricing is comparable). They're picking on how it feels to read your work. The proposal that sounded like it understood the problem. The report that argued a case rather than listed observations. The status update that respected their time and told them the truth.

All of that is voice. It's the accumulated stylistic choices of the firm: what you emphasise, what you skip, how you frame recommendations, which hedges you use and which you refuse, the register you write in when you're confident and the register you shift to when you're not. Ten years of consulting practice has quietly shaped it. It's often the single hardest thing to explain to a junior in an onboarding, because most of it lives in the partners' heads and gets absorbed by reading a lot of the firm's past work.

This is precisely why generic AI tools break professional services drafting. They flatten voice. Not through malice, through statistics: the model is drawing from a giant corpus of business writing and averaging toward the middle of it. What comes out sounds competent, structurally reasonable, and indistinguishable from every other firm's AI-generated output. The clients notice. So do the partners. So does the market.

What "brand voice" actually encodes

Voice is fuzzy when you say it out loud, but it's concrete when you break it down. In every consulting or professional services firm we've worked with, voice reduces to five recurring elements. A useful AI workflow captures all five, once, as first-class workspace assets.

1. Register (formality band)

Where on the formal-informal spectrum does the firm write? Some firms are strictly formal (Big Four audit); some are studiedly informal (boutique strategy shops writing "we think" rather than "our analysis suggests"); most sit somewhere in between, often with different register bands for different document types (proposals warmer, SOWs cooler). This is the easiest voice element to describe and the one chat-style AI most often gets wrong.

2. Point of view (person and pronouns)

First-person plural ("we recommend"), passive ("it is recommended that"), third-person institutional ("the firm recommends"), or second-person ("you should consider"). Every firm has a house preference, often unwritten. Getting this wrong is what most clearly signals AI-generated text to a reader who knows the firm.

3. Sentence cadence

Do you write short and sharp, or long and cumulative? Do paragraphs open with a claim or build toward one? Are transitions explicit ("However,", "That said,") or implicit? A firm's cadence is one of the strongest identifiers of its voice, and it's rarely captured in any style guide. It has to be learned from actual documents.

4. Signature phrases and no-go words

Every firm has its preferred formulations ("in our experience", "the practical implication is", "we would flag") and its actively banned ones (nobody at the firm ever writes "leverage" as a verb, or "at the end of the day", or "circle back"). The signature phrases are the parts a client would recognise; the no-go words are the parts they'd recognise as coming from somewhere else. Both matter.

5. Argument shape

How does the firm build a case? Some open with the recommendation and defend it; some walk through the analysis and arrive at the recommendation. Some foreground caveats; some foreground confidence. Some use structured lists; some prefer prose. This is the deepest layer of voice and the hardest to capture, because it lives in how sections are structured, not in individual sentences.

Why generic AI can't solve this with prompting

The natural first attempt is to describe the voice in a prompt. "Write in a professional but warm tone, using first-person plural, with cadenced sentences and our signature phrases." This works for one paragraph. It stops working around paragraph three, when the model reverts to its training-data average because the prompt's influence on generation fades over distance. It fails entirely across a 20-page document because there's no persistent handle for the voice to hang on to. Every section starts drifting.

The other natural attempt is to paste a sample of the firm's past writing into the prompt. This works better, but it doesn't scale. Every new document needs the sample re-pasted. Every junior needs to remember to include it. The voice lives in a fragile copy-paste convention that decays as soon as someone's in a hurry. And the sample-in-prompt approach doesn't survive the model choosing which of your samples to imitate for which section.

The workflow-engineering answer is to make voice a first-class asset of the workspace, captured once, and anchored automatically to every generation. That's what a workspace-level brand-voice asset is. Set it up on day one, refine it as edge cases surface, and every document generated in that workspace inherits it without anyone having to remember.

What voice extraction looks like in practice

The first document a firm uploads to SkyDraft is usually one of their best proposals or reports. Fifteen minutes later, the workspace has a brand-voice asset extracted from that document. It captures the five elements above, in a structured way, ready to anchor every future generation.

The extraction reads the source document's register, its cadence, its signature phrases, and (harder) its argument shape. It ships that back to the user as an editable structured asset, not a wall of prose. You can see the extracted phrases; you can see the register description; you can see the banned words. You can add to it, edit it, and refine it. It lives in the workspace, next to the glossary and company info, and every future document draws on it.

This is also why voice extraction matters more than template extraction (though we do both). A junior can copy a template. A junior cannot easily copy a voice. Voice is the thing that has been quietly protecting the firm's premium pricing for a decade, and it's the thing that's hardest to hand off. Encoding it into workspace-level infrastructure is how the firm keeps that moat while junior staff use AI to produce senior-quality output.

Where voice sits in the workflow

Every section that SkyDraft generates has access to three workspace-level assets: brand voice, glossary, and company info. The voice asset shapes register, cadence, and phrase-level choices. The glossary enforces the firm-specific terms and bans the no-go words. Company info anchors "who SkyDraft is writing AS" (the firm entity, its positioning, the identity signals it needs to maintain).

Together, these three assets do the heavy lifting of voice. They're not per-prompt inputs; they're constant context across every generation, every regeneration, every section, every document. Which means the voice stays coherent whether you're generating section 1 of a proposal or section 12 of an assessment report drafted three weeks later. The voice is the firm's, not the model's default.

Human edits still count. If you refine a phrase in section 4, subsequent sections in that document see your edit as authorial signal and pick it up. If the pattern of edits shows a shift in the firm's voice over time, the workspace-level asset gets updated (with review) so every future document inherits the change. Voice is a living asset, not a frozen snapshot.

Signals to check when you're evaluating tools

If you're shortlisting AI drafting tools for a professional services firm, voice handling is the most important thing to probe. A few pointed questions cut through the marketing.

Where does voice live? If the answer is "in the prompt", the tool doesn't have a real voice asset. If the answer is "as a workspace-level asset attached to every generation", that's the shape you want.
Can I upload one of my documents and have voice extracted? If setup requires me to write a "voice guide" from scratch, the tool assumes I already know how to articulate what I do. I don't; almost nobody does. Extraction is the difference between a workflow you'll actually adopt and one that dies in setup.
Does voice survive across sections? Generate a 10-section document and read section 8. If section 8 sounds like section 1, the anchoring works. If section 8 has drifted into generic consultant-speak, the voice is being lost across distance.
Are my edits treated as voice signal? If I rewrite a phrase in section 3, does section 4 pick up the pattern? If not, the tool is treating voice as static template rather than a living asset shaped by the humans using it.

The compounding win

A firm's voice, well captured, is an asset that grows in value with every document. Each edit refines it. Each junior who uses it produces senior-quality output faster than they would have on their own. Each document that ships strengthens the pattern for the next one. The moat isn't the AI model (everyone has access to those). The moat is the accumulated, structured, workspace-level capture of how your firm writes when it's at its best.

For the mechanical detail of how templates and workspace assets combine to produce section-by-section drafts, see how it works. For the related concept of a clarifications loop (the other pattern that keeps AI output honest to the firm's actual expertise), read that one. For specific use cases where voice matters most, the use cases page walks through consulting reports, proposals, SOWs, compliance evidence, and RFP responses.

Where to start

Upload one of the documents you're proudest of. We'll extract the voice with you in a working session, review the structured asset together, and generate a first draft on a real engagement using it. You'll know within the first section whether the voice is landing. If it isn't, we edit the asset and try again. That loop is fast.

SkyDraft pilot workspaces are open. Free during pilot, founder-led setup, no credit card. Pilot pricing locks in when standard pricing publishes.