Self-Publishing Workflow With AI in 2026: Where It Helps, Where It Must Not, and What to Disclose

This guide lays out the pipeline in seven stages and, for each, what AI does well, what it must not do, and what you have to disclose. It assumes you have a complete draft that has been through at least one structural revision; if not, start with the revision checklist. Prices and policies are as of September 2026, and policies change; verify each before you rely on it.

The pipeline

Stage 1: Final revision

The last full pass before outside eyes. AI's contribution is the mechanical part of the checklist:

• Continuity. Chapter by chapter, against the story bible: physical facts, timeline, who knows what, setting rules. The prompt is in our consistency guide.
• Tension map. A score per scene from the outline, plotted, to find flat stretches.
• Tell-finding. If any of the draft was AI-scaffolded, the search pass for stock phrases, named emotions and uniform paragraphs. Do this before an editor sees it; editors charge by the hour and should not be paid to remove "a testament to."
• Critique. The beta-reader prompt: where would a reader of this genre skim, stop, or be confused.

What AI must not do here: decide whether the ending works, whether the protagonist is likeable, whether the book is finished. Those are the author's calls and the model will say yes to all of them.

Stage 2: Editing

Three levels, in order: developmental (structure), line (sentences), copy (errors). AI is a reasonable first pass at copyediting: spelling, punctuation, consistency of names and terms, repeated words. Ask for a list of issues with line references rather than a rewritten text, so that you approve each change.

For a book you are charging money for, a human proofread after the AI pass is still the standard. Models miss things (homophones in context, a dropped word that produces a grammatical sentence with the wrong meaning) and introduce things (changes to voice presented as corrections). Budget for a proofreader; use the AI pass to make their job faster and cheaper.

Stage 3: Formatting

The manuscript needs to become at least two files: a DOCX or PDF for print and an EPUB for ebook stores. Our export guide covers formats in detail. The formatting requirements that trip up first-time publishers:

• Chapter headings as real heading styles, not bold text, so the EPUB's table of contents is generated correctly.
• Scene breaks as a consistent symbol or blank line, never as multiple blank lines.
• No tab indents; use paragraph first-line indent styles.
• Front matter (title page, copyright page) and back matter (about the author, also by, newsletter link) as separate sections.
• Straight quotes converted to curly, consistently.

AI helps by drafting the front and back matter text (the copyright page boilerplate, the "also by" list) and by checking an exported file's structure: paste the table of contents and ask whether chapter numbering is consistent. Tools with a built-in EPUB export do most of the mechanical work; if yours exports only DOCX, a formatting tool (Vellum, Atticus, Reedsy's editor, Kindle Create) is the next step.

Stage 4: Cover

Covers sell books, and they are the stage where AI use is most visible and most contentious. Two honest observations. First, a professionally designed cover for your genre is the single highest-return expense in self-publishing, and the market for designers is efficient; a premade or commissioned cover is not expensive relative to the alternative. Second, if you generate a cover with AI, it is AI-generated content under KDP's guidelines and must be disclosed, and some readers and communities react strongly to generated art on covers.

Where AI helps without either issue: writing the design brief. A model can turn your premise, genre and comparable titles into a brief a designer can work from, and can produce rough concept thumbnails to discuss composition, which the designer then replaces with their own work.

Stage 5: Blurb and metadata

The blurb (book description) is the most-read piece of writing you will publish, and it is the one where the AI default voice is most damaging, because browsing readers see hundreds of blurbs and have learned the pattern. Use AI for range, not for the final:

Here is my premise, genre and the first chapter: [paste]. Write ten book
descriptions of 120–180 words, each with a different first line strategy:
a question, the protagonist's situation, the antagonist's threat, a line of
dialogue, the setting, the stakes, a comparison to the genre's expectations,
a single striking image, the ticking clock, the emotional promise. Do not
use: "in a world where," "will she," "everything changes," "a journey."

Read the ten. Take the first line that works, the stakes sentence that works, and write the blurb yourself in the book's voice. Then run it through the tell finder.

Keywords and categories are a research task where AI is useful and cheap: ask for thirty keyword phrases a reader of your subgenre would search, then check each against the store's own search suggestions, which are the actual data. Categories: ask for the five most specific categories your book fits, then verify they exist in the store's current list.

Stage 6: Listing and disclosure

The listing itself is a form. The part that needs care is the AI disclosure.

Amazon KDP's content guidelines (as of September 2026) require you to state whether your book contains AI-generated content, and they define the terms:

• AI-generated: text, images or translations created by an AI tool, even if you subsequently edited them. Must be disclosed.
• AI-assisted: content you created yourself and then used AI to edit, refine, error-check or improve. Need not be disclosed.

The test is who created the content. Classify your own book honestly, chapter by chapter if necessary. A workflow where the model scaffolded scenes and you rewrote every sentence is a judgement call; a workflow where the model drafted and you lightly edited is AI-generated. Our copyright guide covers the ownership side of the same distinction, which points the same way: the more you wrote, the clearer both your disclosure and your claim.

Other retailers and aggregators have their own questions or none; check each. Korean platforms have stricter rules at the contest and community level (Munpia's 2026 contest bans generative AI outright; reader enforcement on Kakao Page and Naver Series is documented); see the Korean platforms guide.

A pre-listing checklist AI can help you run:

Check this listing before I publish: title and subtitle match the cover;
series name and number are consistent with the previous book; the blurb has
no unresolved placeholder text; categories exist in the store list;
keywords do not repeat words already in the title; the "also by" list is
current; the AI disclosure matches the classification I recorded for each
chapter.

Stage 7: Launch and after

Two legitimate uses and one prohibition.

Ad copy. Short ad text for the platforms you use, generated in quantity and edited by you, in the same way as the blurb. Never the AI's default voice in an ad; readers see too many.

Reader questions. A model can draft answers to common reader emails (series order, content warnings, where to buy in a particular country) that you review and send.

Never reviews. Do not generate reviews, reader comments, star ratings, social posts pretending to be readers, or "reader questions" you then answer. It violates every retailer's terms, is illegal in many jurisdictions as a deceptive practice, is detectable, and is the fastest way to lose an account. Nothing in a legitimate workflow involves generating reader-facing testimony.

The disclosure record

One practical habit that makes stage 6 painless: keep a per-chapter log as you write.

CHAPTER | DRAFTED BY | AI USED FOR                    | CLASSIFICATION
1       | me         | continuity check, tell finder  | assisted
2       | me         | expand 3 paragraphs, rewritten | assisted
3       | scaffold   | scene draft, every line rewritten | judgement: assisted
4       | scaffold   | scene draft, light edit        | generated

At publishing time, the classification is a lookup rather than a memory exercise, and if a retailer or a reader ever asks, you have the record. This is the same process record the copyright guide recommends for a different reason.

What the whole pipeline costs, roughly

Not in money (that depends on your cover and editor choices) but in where the time goes. With AI in the mechanical slots, stages 1 and 2 shrink from weeks to days, stages 3 and 5 from days to hours, and stages 4, 6 and 7 stay about the same because they were never the bottleneck. The stage that does not shrink is the one before the pipeline: writing a book worth publishing. AI moves the ratio of time spent from logistics to craft, which is the right direction.

Try it in SparkLore

SparkLore's project view covers stages 1 and 3 directly: continuity and consistency checks against the bible, a tension score per scene, a tell finder for scaffolded passages, and exports to DOCX and EPUB 3 (with heading-styled chapters and a generated navigation file) alongside TXT, Markdown and Fountain, at every tier. The per-chapter disclosure record above fits in the project notes. The export guide explains what each format is for; the free AI synopsis generator will draft blurb options from a premise without an account. Sign up to run the checks on a finished manuscript.