How to Use AI to Review and Edit a Business Book You Wrote

You wrote the book, perhaps over eighteen months of early mornings, perhaps by rewriting every page a ghostwriter drafted. The manuscript is yours, and before a professional editor sees it you want a second reader.

So you paste chapter three into a chat tool and type “edit this”. What comes back is smoother, tidier and no longer quite you. Your blunt opening line has turned polite. A number has changed. The phrase your clients quote back to you has gone.

The AI did the job you gave it, which was rewriting. Used as a reviewer that comments instead of a writer that replaces, it is a useful second reader at three of the four editing levels.

Two different jobs: editing your book or editing an AI draft

This article covers a manuscript you wrote, alone or with a human collaborator. AI is the reviewer; the text is yours.

A draft the AI wrote, which you now have to make your own, has different problems: generic phrasing, invented examples, tidy structure that argues nothing. Our main-site guide to editing an AI-assisted book covers that six-pass process. If most of your sentences started life in a chat window, start there.

Amazon also treats the two cases differently, as you will see below.

What each editing level is, and how much of it AI can do

Editors’ associations do not quite agree on how many levels there are. The US-based Editorial Freelancers Association lists four: developmental editing, which deals “with content, organization, and genre considerations”; line editing, at “the sentence or paragraph level”; copyediting, covering spelling, grammar, usage, punctuation and cross-references; and proofreading, “among the last stages” before publication (EFA, Editorial Service Definitions). The UK’s Chartered Institute of Editing and Proofreading describes three main stages, with sentence-level work inside copyediting, and places proofreading “after editing and layout” (CIEP, About proofreading and editing).

Here is our view of where AI earns its place at each level for a business book.

Level	What a human editor does	What AI does well	Where AI misleads	Our verdict
Developmental / structural	Tests the argument, chapter order and gaps	Describes your argument back to you; lists what each chapter does; finds the story told twice	Praises the structure by default; reads a logic into the book that is not on the page	Good first reader, poor judge
Line	Improves clarity and rhythm sentence by sentence	Flags long, tangled or vague sentences and buried points	Rewrites your sentences into its own average style	Useful if it only comments
Copyediting	Corrects mechanics, applies a style sheet, queries errors	Checks against a style sheet you supply	Misses errors, “fixes” correct text, alters quotes and numbers	A pre-check, not the copyedit
Proofreading	Checks the laid-out pages before print	Very little, because it never sees your layout	Lets you believe a manuscript check was a proof	Leave it to a person

AI is better at describing your book than at deciding what it should be. Build your passes around that.

The rule for every pass: comments, not replacements

Ask a chat model to “improve” a passage and it will change more than you asked. A 2023 study comparing ChatGPT with Grammarly on a grammar-correction benchmark found that ChatGPT “goes beyond one-by-one corrections” and prefers to change “the surface expression of certain phrases or sentence structure”; human raters found it made more over-corrections than the other tools (Wu et al., ChatGPT or Grammarly?). Models have moved on since; the habit has not.

The second risk is slower. In a controlled study with 118 participants from India and the United States, AI writing suggestions pulled the Indian participants towards Western writing styles, “altering not just what is written but also how it is written” (Agarwal, Naaman and Vashistha, CHI 2025). That study was about culture, not business books, but the mechanism carries over: accept enough small suggestions and the book drifts towards the model’s default voice, the one described in why AI-written books sound generic.

So every pass starts with the same instruction block, pasted at the top of each session.

You are reviewing a business book manuscript that I wrote. Act as an editor
writing margin comments, not as a co-author.

Rules:
- Do not rewrite my text. Do not return a revised version.
- For each issue, quote the exact words (15 words at most) and say what the
  problem is.
- Suggest the fix in principle ("split this sentence", "cut the second
  example"), not replacement wording, unless I ask about a specific line.
- Never correct a fact, figure, name or quotation. If something looks wrong,
  raise it as a question for me.
- If you are not sure something is a problem, say so.
- My voice is deliberate: short sentences, plain opinions, British spelling.
  Do not flag these as errors.

Replace the last line with your own habits, and attach your style sheet if you have one.

Four passes, each with a prompt you can copy

Run them in order, in separate sessions, each starting with the instruction block.

Pass 1: a structural review of the whole book

Structural feedback is the most expensive help to buy. The EFA’s 2026 rate chart gives a median range of 4.0 to 5.0 cents per word for nonfiction developmental editing (EFA, Editorial Rates, survey run November 2025 to mid-January 2026), which works out at roughly $2,000 to $2,500 for a 50,000-word book. AI will not replace that judgement. It can make sure you are not paying someone to notice that chapters four and seven make the same point.

Run it in two steps: a fixed-format summary of each chapter, then all the summaries together. Some tools accept a whole manuscript, but long inputs are read less evenly (see can ChatGPT or Claude write a business book), and summaries show you when the model has misread a chapter.

Step 1, one chapter at a time:
Summarise this chapter in this format, using my wording where possible.
- Main claim (one sentence)
- What the reader can do or believe after reading it
- Stories and examples used (two to five words each)
- Statistics and named sources used
- Terms I define or rely on

Step 2, all summaries together:
Below are summaries of every chapter in my book. The book's promise to the
reader is: [one sentence].
1. Set out the argument as a numbered chain: what each chapter claims and
   how it depends on the one before. Mark any link that does not follow.
2. Which chapters do the same job? Which stories or statistics appear twice?
3. Which promise made in the introduction is never delivered?
4. Which chapter could be cut with the least loss? Explain; do not decide.
Do not praise the structure. List problems and questions only.

Keep that last line, or most models open with praise. For the idea of one job per chapter, see how to structure a business book.

Pass 2: a line review, one chapter at a time

Line work is where voice is kept or lost, so keep the AI on a short lead: one chapter per session and a narrow brief.

Line review of the chapter below. List up to 20 issues in order of
appearance. Look only for:
- sentences over 35 words that would be clearer split
- sentences where the main point is buried behind a long clause
- vague words where I probably meant something specific ("things",
  "various", "a number of")
- a word or phrase used three or more times on one page
- paragraphs that say the same thing twice
For each: quote, problem, principle of the fix. No rewritten sentences.

The cap of twenty is deliberate. Ask for everything and you get 140 comments, half of them taste, and by comment ninety you accept changes out of tiredness.

When a comment is right but the fix escapes you, ask about that line only (“three ways to split this sentence, using my words”) and then write your own version. We would never paste a model’s sentence into a client story; those are the passages readers remember.

Pass 3: a consistency check against your style sheet

The CIEP lists “applying consistency to the language and formatting” and creating a style sheet among a copyeditor’s routine tasks (CIEP, What is copyediting?). AI handles the dull half of this well if you hand it the sheet. Without one, it applies its own preferences.

Below are my style sheet and glossary, then one chapter. List every place
the chapter departs from the style sheet: spelling, capitalisation,
numbers, framework and step names, and how I refer to clients. Quote each
instance and name the rule it breaks. Do not flag anything the style sheet
does not cover.

Then search the manuscript yourself for your framework’s name and its stages. A model can miss an instance; your word processor’s search will not. If you do not have a style sheet yet, keeping your voice consistent across chapters shows how to write one in an afternoon.

Pass 4: a list of claims for you to verify

This pass asks the AI to find claims, not to check them.

List every factual claim in this chapter that a sceptical reader could ask
me to back up: statistics, dates, research findings, quotations, named
companies or people, and results I attribute to clients. For each, quote
the sentence and say what kind of evidence it needs. Do not tell me whether
the claim is true. Do not suggest sources.

Then work through the list with your sources open. Even a human copyeditor only checks that “basic facts and arguments are plausible, consistent and reasonable” (CIEP, same page); full fact-checking is yours. Asking a model whether a statistic is right invites a confident answer and sometimes a “corrected” figure with nothing behind it. Our guide to stopping AI inventing facts in your book covers how that happens.

A worked example: one paragraph, edited two ways

This is a hypothetical example. Imagine an operations consultant whose chapter on meetings contains this paragraph:

Most leadership teams don’t have a meeting problem. They have a decision problem that shows up in meetings. In 2022 I sat in on a quarterly review at a 300-person logistics firm where the same pricing decision had been “parked” four times. Nobody was lazy. Nobody owned it.

Asked simply to “edit this paragraph”, a chat model might return something like:

Many leadership teams believe they face challenges with meetings, but the underlying issue is often decision-making. For example, at one logistics company, a pricing decision was repeatedly postponed, highlighting the importance of clear ownership.

It is grammatical, and worse in every way that matters. The opinion has gone soft (“often”), the specifics (2022, 300 people, four times) have gone, and the author’s two short closing sentences have become a moral.

With the instruction block, the line-review brief and one extra line asking it to flag identifiable client details, a good run returns two comments instead. One: “‘parked’ four times: over what period? A reader may want the timescale.” Two: “‘a 300-person logistics firm’ in 2022: could the client be identified from this?” The author adds “over three quarters”, writes “a mid-sized logistics firm”, and moves on. Most good AI edits look like that.

The four risks, and how to contain each one

Voice flattening is the slow risk; the comment-only rule fixes it. Introduced errors are the quiet one: asked to tighten a paragraph, a model can turn “four” into “several” or drop the qualifier that made a claim true. Invented corrections are the dangerous one: the model “fixes” a statistic you had right, or a quotation it remembers differently. Pass 4 exists so it never gets the chance.

Confidentiality is the one people forget. A manuscript full of client stories is client material. As of November 2026, Anthropic’s consumer terms update (announced 28 August 2025) lets Claude Free, Pro and Max users choose whether their chats are used for training, with five-year retention when the setting is on and the existing 30-day retention when it is off; Team, Enterprise and API use run under commercial terms instead. OpenAI’s data-control pages returned an access error when we tried to open them, so check ChatGPT’s settings on OpenAI’s own help pages on the day you decide. Whatever the tool, anonymise client stories before they leave your machine. More in protecting your ideas and your clients’ data when writing with AI.

Does AI editing count as AI content on Amazon KDP?

Not if the words were yours to begin with. KDP’s content guidelines define AI-generated content as text, images or translations “created by an AI-based tool”, and it stays AI-generated “even if you applied substantial edits afterwards”. If you wrote the content yourself and “used AI-based tools to edit, refine, error-check, or otherwise improve that content”, it is AI-assisted, and “You are not required to disclose AI-assisted content” (Amazon KDP, Content Guidelines, as of November 2026).

Two cautions. Paragraphs the model wrote and you pasted in are moving towards the generated side of that line. And KDP makes you responsible for checking that both kinds of content follow its guidelines. For the rest of the process, see publishing a business book on Amazon KDP.

Why we would still pay a human copyeditor

The AI passes make a human edit cheaper and better; they do not replace it. A copyeditor queries what they cannot confirm and is accountable for the result in a way a chat window is not.

The EFA’s median ranges are 3.0 to 4.0 cents per word for nonfiction copyediting and 2.0 to 3.0 cents for nonfiction proofreading (EFA, Editorial Rates), so roughly $1,500 to $2,000 and $1,000 to $1,500 for a 50,000-word book. The EFA does not set rates. If your budget stretches to only one paid edit, we would buy the copyedit and skip a paid line edit: careful AI line passes catch much of what a business book needs at that level, while copyediting slips are what readers notice and reviewers quote. Proofreading the laid-out file is the step we would least want to hand to AI, because it never sees the pages.

What to do this week

Pick one chapter. Write the instruction block with three lines describing your voice, run Pass 2, and count how many of the twenty comments you would act on. If it is more than five, the tool is earning its place, so run Pass 1 across the book next. If it is fewer than three, your sentences are in better shape than you feared, and your money is better spent on a copyeditor.