The first reaction to raw GPT drafting is usually disappointment. The reply is grammatically fine, on-topic, polite — and visibly nobody. The principal sees it, decides it does not sound like them, and re-types it from scratch. The four-hour day did not move; the only thing that moved was the tool.
Voice-matched drafting is a different category. The reply is grounded in how the principal already speaks — the construction of sentences, the use of a colon vs a comma, the way an investor note ends with a soft yes, the way a customer escalation opens with a hard "got it." The reply lands already in the principal's register, which means the re-read tax the principal would have paid is removed before they open the thread.
Three differences matter. First, voice-matched drafting is built on the principal's prior correspondence — not a generic model. "Draft with GPT" assumes the principal's voice can be inferred from a prompt; "draft in your voice" assumes the voice has to be modeled, and starts from the model. Second, voice-matched drafting keeps the tone stable across stakeholders — investor, board, customer, vendor — without the principal toggling a register dropdown, because the model already knows which is which. Third, voice-matched drafting is auditable: every reply carries a small audit trail back to the prior thread, so the principal can see why the draft says what it says, and edit one line instead of rewriting the whole.
The practical effect on the four-hour day is not "the model writes the reply." It is that the reply lands ready to approve, and approval is a 90-second decision, not a 12-minute rewrite. Over a week, that is the gap between a four-hour executive inbox and a 45-minute morning review with a queue of drafts that were already in the principal's register.
If you measure, do not measure tokens drafted. Measure the four-hour day. The model that trims the four-hour day is the one that already speaks with the principal's voice — not the one that drafts fast.