How to Use AI on Client Work Without Shipping Unedited Drafts
A practical QA checklist for freelancers who use AI on client writing: what you may generate, what you must check, and what belongs in the agreement—before the draft goes out.
Using AI for client writing is no longer unusual. Shipping the first fluent draft is still a mistake. The sentences can be smooth, and the problems that get you into trouble—an invented statistic, a misspelled product name, a claim the client cannot defend—sit in the same paragraph.
This is a quality and ethics guide for freelance writers and other solo operators who already use, or are about to use, AI on paid work. It is not a buying guide, and it does not argue that you should or should not use AI. It argues that you remain the author of record.
For choosing a writing tool, see Best AI writing tools for freelance writers. For where writing sits next to research, email, and the rest of the business, see The solo operator AI and ops stack.
The real risk is not “using AI.” It is shipping unreviewed work
Clients do not usually object to a tool that helps you outline or tighten a paragraph. They object to work that is wrong, generic, or written in a voice they did not hire. They also object when a policy they thought was obvious—no AI, or no pasting confidential material into a third-party system—was never discussed.
The failure mode is predictable. You paste a thin brief into a tool, accept a complete-looking article, and send it because the deadline is close. The tool has filled every gap. Some of those gaps were facts. Some were tone. Some were legal or reputational claims you would never have typed yourself if you had been starting from a blank page.
A process beats a vow to “be careful.” The rest of this article is that process: what the tool may do, what you must still do, and what you should write down with the client before the first assignment.
What you may generate, and what you still have to write or check
Use this as a default split. Tighten it for regulated industries, named testimonials, or clients who forbid AI entirely.
| Job | Reasonable to generate, then edit | Do not ship without your own work | Do not generate at all |
|---|---|---|---|
| Outline | A first structure from a brief you already wrote | The final outline, including what you will not cover | An outline that assumes facts you have not sourced |
| Draft | A first pass from an approved outline, with the client’s terms and constraints in the prompt | The voice, examples, and transitions you would stand behind in a meeting | A draft you have not read against the brief |
| Evidence | A list of questions to research, or a summary of sources you provided | Every number, date, price, and citation, checked at the source | Statistics, studies, quotes, or testimonials the tool invented |
| Voice and claims | Alternate headlines or a tighter version of a paragraph you wrote | Product names, legal claims, pricing, and competitor statements | Implied endorsements, medical or legal advice, or a person’s words they did not say |
| Confidential material | Nothing, unless the client has approved the tool and you have read its retention settings | Your judgment about what may leave the client’s environment | Unpublished strategy, personal data, credentials, or anything the contract marks as confidential |
The middle column is the work you are paid for. If a tool makes the left column faster and you skip the middle, you have not saved time. You have moved the risk to the client, and then back to you when they notice.
Writesonic can sit in the left column as a drafting or SEO-workspace example, in the same way ChatGPT, Claude, or Gemini can. It does not move a fact into the “checked” column. On Writesonic’s official pricing page on 14 November 2026, even the Starter plan is sold as an SEO and AI-visibility platform with an article writer included, at $79 a month billed annually or $99 billed monthly. A more expensive generator is not a more reliable source.
A QA checklist you can run on every AI-assisted draft
You can run this on a 1,000-word article in a single sitting once it is a habit. Use a printed or digital copy of the draft, and mark the document, not your memory.
Before you open the tool
- Confirm, in writing, whether the client allows AI, and which tools or data are off limits.
- Write a brief you would give a junior writer: audience, offer, proof, tone, words to avoid, and sources you already trust.
- Collect the source material yourself: the client’s site, the brief, interviews, and any URLs you will actually open.
- Decide what must not go into the tool: personal data, unpublished strategy, credentials, legal memos, or anything you would not put in an unsecured email.
- If you use a saved prompt or a project, put the constraints there, not only in a chat you will lose.
If you cannot write the brief, you are not ready to generate. The tool cannot respect limits you never stated.
After the first draft
- Read the draft once without making small edits, only to mark anything that surprised you.
- Highlight every specific claim: numbers, dates, prices, “studies show,” testimonials, competitor statements, and product capabilities.
- Check each highlighted item against a source you opened. If there is no source, cut it or research it yourself.
- Search for leftover prompt language, hedging that hides a missing fact, and generic examples that could belong to any client.
- Confirm names, including the client’s spelling, product capitalization, and the names of people you quote.
- If the tool added a section that was not in the approved outline, either justify it or delete it.
This is the step most people skip because the draft already looks like an article. It is also the step that prevents the expensive revision.
Before the client sees it
- Read the piece aloud, or at least slowly, as the deliverable, not as a generated file.
- Check links, headings, and any internal promises (“as shown in the table above”) that may no longer be true after your edits.
- Make sure the voice matches what the client hired, including reading level and claims they are willing to make.
- Remove anything you would not be comfortable defending on a call.
- Save the brief, the sources, and a note about which tool you used, in case the client asks later.
If you also deliver audio, the same standard applies to a script before it is voiced. A confident mispronunciation of a product name is still your error. See Best AI voice generators for solo creators for the license and consent side of that job.
What belongs in the client agreement
You do not need a long AI addendum. You need three decisions in writing, ideally in the kickoff form or the statement of work, before the first draft.
Disclosure. Some clients forbid AI. Some do not care. Some want to know which parts were assisted, or which tools you use. Ask. Do not wait for a dispute to reveal a policy. If the client forbids AI, do not use it on that account, including “just for an outline,” unless they revise the rule in writing.
Confidentiality and tools. Name the kind of material that will not be pasted into a third-party system, and say that you will use only tools you have reviewed for retention and training settings. If the client’s industry is regulated, get the constraint in the brief. If they require a specific workspace with no training on their data, use that workspace or decline the work.
Accountability. You remain responsible for the final copy. If a draft includes a fake quote, a wrong price, or a defamatory claim, that is your error. State that you will fact-check and edit before delivery, and that the client still reviews for factual and brand accuracy on their side.
A short clause is enough. For example: you may use AI for outlining and first drafts; you will not use it to invent sources, testimonials, or legal advice; confidential material stays out of tools the client has not approved; and you remain responsible for the deliverable. If you later include affiliate links in content you publish on your own site, disclose those separately.
A one-assignment habit, not a one-time audit
Checklists fail when they are a special event. Attach this one to the kind of work you already do.
Pick one assignment this week that used, or could have used, AI. Run only the “after the first draft” highlight pass, even if you already delivered the piece. Count the claims you would not want to defend. That number is your baseline.
On the next assignment, add the pre-brief and the confidentiality check before you open the tool. On the assignment after that, add the final read-aloud pass. You do not need a new system. You need the same three gates on every piece until they are boring.
If a tool consistently produces drafts that fail the highlight pass, the tool is not saving time. Change the prompt, change the workflow, or stop using it for that job. The comparison in Writesonic vs a ChatGPT workflow vs Semrush is about buying the right layer. This checklist is about not embarrassing yourself after the purchase.
Common ways a “helpful” draft goes wrong
- “Studies show” with no study. Cut it, or name the study you opened.
- A testimonial or quote the client never collected. Invented social proof is a trust problem, not a style problem.
- Yesterday’s price, or a price the tool guessed. Check the client’s current page, or ask.
- A competitor claim you cannot source. If you cannot open a citation, you cannot publish the sentence.
- The client’s name, product, or legal entity spelled the familiar way, not the correct way. Search and replace is not optional.
- A confident paragraph on a topic that was not in the brief. It may be padded. It may also be wrong.
- Confidential strategy, pasted into a consumer chatbot to “help it sound like us.” The draft is not worth that leak.
None of these require a 30-product tool review to prevent. They require a slow read and a source pass.
Want the prompts and a printable version of this checklist? The Solo Operator AI Toolkit has 40 prompts, a client-work QA checklist, and a solo stack cost sheet. Get the toolkit →
FAQ
Can I use AI for client writing? Often yes, if the client allows it and you still brief, fact-check, edit, and take responsibility for the final copy. Some clients forbid AI or restrict tools and data. Put that in the agreement before you start.
Do I have to disclose that I used AI? You have to follow the rule you agreed with the client. If there is no rule yet, ask. Disclosure is not a substitute for quality, and hiding a forbidden process is worse than asking early.
Is a specialist writing tool safer than a general assistant? Safer is the wrong word. A specialist workspace may hold a brief and an outline more cleanly, but it can still invent a fact. Official prices also differ: ChatGPT Plus is $20 a month, while Writesonic’s official Starter plan was $79 a month billed annually on 14 November 2026. Price does not replace a QA pass.
What should I never paste into an AI tool? Anything the client has not approved for that tool: unpublished strategy, personal data, credentials, legal or medical files, and anything marked confidential. If you have not read the tool’s training and retention settings, treat client material as off limits.
What if the client wants more volume and less review time? Then the offer has changed. Unreviewed volume is easy to replace and easy to get wrong. You can use AI to go faster through outlining and first drafts; you should not sell speed by dropping the highlight pass. If the fee cannot cover that pass, change the scope or the fee.
How does this apply if I also produce audio or video? Check the script with the same list before you generate a voice, and do not clone a voice without documented consent. See Best AI voice generators for solo creators.
