Estimate AI drafting from the complete task you expect to perform: input preparation, scene generation, review, revisions, and any discarded attempts. Test a small representative sample and record actual usage when the provider reports it. Keep uncertain assumptions visible and include human review effort separately. A first-pass word count is not a reliable estimate of the whole drafting process.
Define the output you are estimating
Specify whether the tool will brainstorm, draft complete scenes, revise human-written passages, or perform another bounded task. Those workflows send different amounts of context and produce different kinds of output. Record the target deliverable and the review standard it must meet. A scene that merely reaches the requested length is not necessarily a scene the author can use in the manuscript.
Choose a representative sample from the novel's actual demands. Dialogue, action, exposition, and a scene with several continuity constraints may behave differently. Avoid estimating a complex book from one easy descriptive passage. Keep the sample small enough to review carefully, but varied enough to expose the work you expect to repeat. The estimate should describe your intended process rather than a provider's most favorable demonstration.
Record inputs as well as generated text
Note the context supplied for each task, including the scene brief, relevant continuity facts, style guidance, and any text being revised. Repeatedly sending a large manuscript can change usage compared with a focused context packet. Use the provider's actual reported usage and current pricing terms where available, rather than assuming that visible output words represent the entire billed work.
Record unsuccessful or incomplete attempts as well. Depending on the service and circumstances, an unusable response may still involve usage or charges, so do not automatically assign it zero cost. Preserve the reported information and mark unavailable amounts as unknown. Keep your own estimate distinct from the provider's billing record. That distinction matters when a task fails or a tool presents only partial usage information to the author.
Measure the revision path through acceptance
For each sample scene, record whether it was accepted, revised, regenerated, or discarded. Note why another pass was needed: continuity error, weak motivation, unsuitable voice, missing action, or a changed brief. Different causes suggest different improvements. Better context may reduce factual contradictions, while a weak scene goal may require an editorial decision rather than another attempt with the same instructions.
Include the author's review time in a separate column. A low direct generation cost can still produce a large editing burden. Conversely, a more carefully prepared brief may take longer initially and reduce later repair. Compare the total path to an acceptable scene, not just the speed of the first response. Do not count substantial human rewriting as free simply because it does not appear on the provider's invoice.
Build scenarios instead of a single promise
Use the sample to create a baseline and a more demanding revision scenario. Vary the number of attempts and the amount of context where those inputs are uncertain. State which scenes are unlike the sample and may need a separate estimate. Avoid multiplying one unusually successful result across the entire novel as though every scene will have the same complexity and acceptance rate.
Check the current provider terms before relying on the figures. Pricing, model behavior, available limits, and the way usage is reported can change. Keep the verification date and the assumptions with the estimate. If exact usage is unavailable, label the calculation as approximate rather than presenting an artificially precise total. The purpose is to make the likely work visible, not to promise a fixed production cost before the process is understood.
Set a review checkpoint and a stopping rule
Decide when to reassess the estimate, such as after a representative set of scenes has passed human review. Also define what happens when repeated attempts do not solve the same problem. Pause that scene, revise the brief, or write it directly rather than running an indefinite loop. A stopping rule protects both the budget and the manuscript from accumulating many unusable alternatives without an editorial decision.
Preserve accepted text, task records, and origin notes as the project proceeds. Actual AI-created prose may have disclosure implications under current KDP guidance even after later editing. Keep that production history separate from the cost worksheet, while linking both to the same scene identifiers. Update the estimate with observed results without erasing the original assumptions, so the author can understand what changed and improve the remaining workflow.
Frequently asked questions
Clear answers for this publishing decision.
Can I estimate the whole novel from output word count alone?
No. Context, revisions, discarded attempts, provider billing terms, and human review all affect the process. Use representative tasks and preserve the assumptions behind the estimate.
Should an unusable response always be counted as free?
No. Check the provider's actual usage and billing information. When the amount is unavailable, mark it unknown instead of assuming that a failed or discarded attempt incurred no charge.
What should happen after repeated weak drafts of the same scene?
Reassess the scene goal, brief, and context before requesting more attempts. A defined stopping rule helps avoid spending on an unchanged problem that requires an editorial decision.
What does this book cost to print and produce, and what should the list price be?
Printing cost follows from the trim, ink, and page count, the list price sets the royalty after that cost, and the production budget should be estimated before generation starts, so compare these numbers before you commit to a format or a plan.
See what a book project costs in credits