The demo that never survives contact with real work

Every generic AI tool sells the same demo. You paste a brief, you press one button, and a complete document scrolls down the screen in under a minute. It is genuinely impressive the first time. It is also the reason so many firms tried AI drafting in 2024, felt the magic, and quietly stopped using it by March.

The gap is between the demo and the deliverable. In the demo, the output is the finish line. In real professional services work, the output is the start of the job: now someone senior has to read all 22 pages, find the three claims that are subtly wrong, notice the two sections that contradict each other, catch the pricing paragraph that quietly invented a discount you never offered, and rewrite the executive summary because it summarised a document the AI imagined rather than the one it produced. The thirty-second draft cost you four hours of forensic reading.

This is the choice hiding underneath “how fast is your AI tool”: does it draft the whole document in one shot, or one section at a time with a review gate between each? It sounds like an implementation detail. It is actually the single biggest determinant of whether the output is usable. So we ran both approaches on the same real engagement and watched what happened.

The test: one proposal, two workflows

The document was a fixed-price transformation proposal for a mid-market logistics client. Twenty-two pages: executive summary, understanding of requirements, proposed approach across four workstreams, team and governance, assumptions and dependencies, commercial section, and terms. The source material was a 70-minute scoping call transcript, the client's original brief, and two of the firm's prior proposals for similar work.

We gave both workflows identical inputs. The only variable was structure: one drafted the whole thing from a single prompt, the other drafted section by section with a human review gate between each. Same underlying model. Same source documents. Same firm voice and glossary. We measured time to first draft, time to a version we would actually send, and the kinds of errors each produced.

One-shot: fast to a draft, slow to a deliverable

The one-shot run produced 22 pages in about forty seconds. It read fluently. It had the right headings in the right order. If you skimmed it, you would think the job was basically done.

Then we read it properly, and the problems were exactly the problems one-shot drafting always produces:

Invented specifics. The approach section confidently named a “six-week discovery sprint” that appeared nowhere in the scoping call. The client had actually asked for a four-week window. The AI had no way to flag the gap, so it filled it with something plausible.
Internal contradictions. The team section assumed a dedicated solution architect. The commercial section priced the engagement without one. Neither paragraph knew what the other had committed to, because they were generated in the same pass with no checkpoint between them.
Voiceless boilerplate in the sections that matter most. The executive summary was competent and generic, the kind of paragraph that could open any proposal from any firm. That is exactly the section a prospect reads first and uses to decide whether you understand their problem.
An unreviewable review job. Because every section arrived at once, there was no natural place to stop and correct. You either accepted the whole thing and edited it afterwards, or you rejected it and re-prompted the entire document, losing the parts that were fine.

Time to first draft: forty seconds. Time to a version we would send a client: a little over three hours of reading, cross- checking, and rewriting. Most of that time went to finding the errors, not fixing them. A wrong sentence takes ten seconds to correct once you have spotted it. Spotting it in fluent, confident prose is the expensive part.

Section-by-section: slower to start, done sooner

The section-by-section run drafted the executive summary first, then stopped. Before we advanced, it surfaced a clarification: the scoping call mentioned a four-week window in one place and “roughly a month, maybe six weeks with the integration piece” in another, and it asked which we wanted to commit to. We answered “four weeks for discovery, integration is a separate later workstream.” That answer became context for every section that followed.

That single question is the whole ballgame. The one-shot draft hit the same ambiguity and silently guessed wrong. The section-by-section draft hit it, recognised it as genuinely ambiguous, and asked. Read more on why that behaviour has to be designed in rather than hoped for in our piece on the clarifications loop.

From there the pattern repeated. Each section was drafted, shown, and reviewed before the next began. Because each section saw the approved versions of the sections before it (including our edits), corrections propagated forward instead of getting lost:

We tightened the executive summary. The understanding-of- requirements section that came next echoed the tighter framing automatically.
We confirmed a two-person delivery team with a fractional architect. The commercial section priced exactly that team, because it read the approved team section as context. No contradiction to catch later, because the contradiction never got written.
When the commercial section reached the discount question, it did not invent one. It asked whether the standard pilot discount applied. It did not, and we said so.

Time to first section: about a minute. Time to a full draft we would send: roughly fifty minutes, most of it our own judgement calls at each review gate rather than error hunting. The draft that arrived at the end was one we had already read and approved in pieces, so the “final review” was a formality, not a forensic audit.

Why the slower-looking option wins

The one-shot approach front-loads the speed and back-loads the work. You get the dopamine hit of a finished document in seconds, then pay for it with hours of reading you cannot delegate, because only someone senior can tell which confident sentences are wrong. The section-by-section approach spreads small decisions across the drafting itself, so there is no cliff of accumulated error to climb at the end.

There is a deeper reason it wins, and it has nothing to do with speed. A document generated in one pass has no memory of your corrections. If you fix the delivery model in the team section, nothing downstream knows. In a section-by-section workflow, every edit you make becomes part of the context the next section reads. The document converges on your intent as you go, instead of drifting further from it with every page. This is the same reason we argue against a “draft everything” button at all in our how-it-works walkthrough: the review gate is not friction, it is the mechanism that makes the output trustworthy.

Where one-shot is genuinely fine

None of this makes one-shot drafting useless. For short, low-stakes, single-purpose text, the overhead of a review gate per section is not worth it. A one-paragraph project update, a first-pass meeting summary, an internal note nobody outside the team will read: draft it in one shot, skim it, ship it. The cost of an undetected error is a shrug.

The calculus flips the moment the document is (a) long enough that sections can contradict each other, (b) going to someone who will make a decision based on it, or (c) committing your firm to something (scope, price, a delivery date, a compliance claim). Proposals, SOWs, engagement reports, tender responses, and audit evidence are all three at once. That is exactly the territory where one-shot's hidden editing tax is largest, and where section-by-section's early questions save you from the errors that actually cost money.

What to ask a vendor

If you are evaluating AI document tools, the “whole document in one click” demo is a warning sign, not a feature. The questions that actually separate a usable workflow from an impressive toy:

Can I review and edit between sections? Or do I get the whole document and an editing job?
Do my edits carry forward? If I fix section two, does section five know?
What does it do when it is uncertain? Does it ask, or does it guess plausibly and leave me to find the guess?
Can I stop partway and come back? Long documents are not written in one sitting, and neither should they be drafted that way.

We built SkyDraft around the section-by-section answer to every one of those, because it is the version that survives contact with a real deadline. Even in Generate all mode, each section completes and can be reviewed before the next begins, and you can stop at any step. There is no button that drafts 22 pages you have not looked at. See how the whole workflow fits together on our use-cases page, from proposals to compliance evidence.

The honest summary

One-shot drafting optimises the demo. Section-by-section optimises the deliverable. On our 22-page proposal, one-shot was forty seconds to a draft and three-plus hours to something sendable; section-by-section was a minute to the first section and under an hour to a draft we had already approved as we went. The difference was not the model. It was whether the workflow let a human stay in the loop while the document was still cheap to correct. That is the entire game.