A specific example. One founder, one workflow, one change.
What this is
Not a case study with logos. Not a testimonial. A demonstration. Names are disguised; the work is real. If you have been inside a workflow like this, you will recognize the shape.
The situation
A twelve-person professional services firm. The founder — call her M — ran the proposal pipeline herself. She wanted to stop.
Proposals were being drafted inside the email thread with each prospect. M would copy-paste pricing language and scope boilerplate from the last three proposals she had written, edit inline, and send. Each proposal took roughly 40 minutes of her focused time, at least once a week.
AI was already in the workflow, but in the wrong place. Staff were using a chat tool to "improve the draft," which mostly meant re-generating sentences that then needed to be re-edited back toward something that sounded like M. Net effect on time: slightly slower. Net effect on voice consistency: worse.
The observable pattern: M was the bottleneck, because the proposal needed her judgment and her voice, and because the draft-generation step and the context-gathering step were happening in the same place, at the same time, by the same person, inside the same email thread.
Before
Draft produced inside the email thread. Context, tone, pricing, scope, and prior-conversation reference all mixed together in M's head and in inline edits.
First-draft usability: ~30%. Every proposal needed substantive rewrite before sending.
Founder rewrite time: ~40 minutes per proposal. At roughly 4 proposals a week.
Staff AI use: generative-rewrite inside the draft. Helpful on surface phrasing, harmful on voice consistency.
The move
The unlock was not a new tool. It was splitting the workflow into two separate steps and giving each one its own place, its own artifact, and its own AI role.
Step one became context bundle: a short structured document assembled by a staff member at the start of every new proposal, pulling together (a) the prior conversation, (b) the specific scope that is being proposed, (c) the three most relevant prior proposals we would use as voice reference, and (d) two or three sentences from M about what is actually at stake for this prospect. The AI's job here is compression, not generation.
Step two became draft generation: the AI produces a first draft directly from the context bundle, using M's own prior proposals as voice reference. The draft lands as a separate artifact, not as text inside an email. M reads it, makes one pass, sends.
Three supporting changes made the split hold. The context bundle lives in a predictable template with named sections, which makes it cheap to assemble and cheap for M to trust. The prior-proposal voice-reference set was curated and stays stable — not re-chosen per proposal. The staff member doing the context bundle is the one who was already in the email thread, so no handoff cost was added.
What got thrown away: generative-rewrite-inside-the-draft. That role, which the staff had been using AI for, was actively bad for the voice, and the new structure makes it unnecessary.
What changed
The volume change is the secondary effect, not the primary one. The primary effect was that M stopped being the bottleneck. She reviews; she does not draft. Proposal throughput rose because she was no longer the gating step; it was not the goal.
Voice consistency across proposals improved. The prior-proposal voice reference was more stable than M's in-the-moment phrasing, which had drifted over time under rewrite pressure. The drafts are now slightly more like M's voice than M's own rushed edits had been.
Hidden effect, named by M in the follow-up: the time she had been using for rewrite was getting pulled out of the part of her week that had been meant for sales conversations. The change made her week legible to herself again.
Why this generalizes, and where it doesn't
What travels: the pattern of splitting context-gather from draft-generation, and giving each its own artifact. This is a workflow shape, not a tool choice. It works in proposal flow, in research synthesis, in client reports, in follow-up emails that are too personal to templatize and too repetitive to handwrite from scratch.
What does not travel: the exact structure of the context bundle. That is shaped by M's business, her voice, her clients, and the specific decisions buyers in her category are making. A different business needs a different bundle. That is the part of the work that cannot be pre-baked.
The deeper point: AI did not fix the workflow. The workflow redesign made AI useful in a place it had not been useful before. The redesign was the leverage. AI was the medium.
This is one example. The shape of the move is consistent across engagements; the particulars never are.