Give the AI a client pack (brief, style guide, facts sheet, voice samples), let it draft and list every claim it made, then do the real work yourself: edit for voice, check each claim against a source, run a plagiarism scan, and deliver with a claims log and an honest AI-use line.
Take one invented job: a 900-word landing page for a 12-person payroll-software company, billed at a flat $850. The shape of the work changes. Drafting shrinks, while editing and checking grow, and that shift is what protects both quality and voice.
The job, the contract line and the client pack
The client wants a landing page for a new holiday-pay calculator, three headline options and a meta description, in three days. Before touching a tool, do two things. First, put a sentence about AI in your proposal: you use AI tools for first drafts and research, you edit and fact-check everything, and you will follow the client's own AI policy if they have one. Second, ask what the client is happy for you to paste into a tool, because their product notes may be confidential. Consumer plans have a model-training switch in their privacy settings, while business plans such as ChatGPT Business and Claude Team don't train on your content by default, though both need two seats at minimum.
Then build the client pack, a single folder you attach to a project or chat:
- the brief, with the page sections, the audience and the call to action;
- a product facts sheet of up to two pages, which is the only place the model may take facts from;
- a dozen customer comments, quoted exactly, with permission noted;
- three paragraphs of the client's best existing copy, as voice anchors;
- the style guide below.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowA one-page style guide the model can follow
A style guide the model can follow has to be specific enough to fail. "Friendly but professional" cannot be checked; "average 12 to 15 words a sentence" can. This is the filled-in version for the job, drawn from the client's anchors and brief:
| Item | Rule for this client |
|---|---|
| Voice in three words | Plain, practical, a little dry |
| Reader | Payroll managers at firms of 20 to 200 staff, busy and sceptical of software promises |
| Sentences | Average 12 to 15 words, none over 25, one idea each |
| Person and tense | "You" for the reader, "we" for the company, present tense |
| Words we use | Pay run, timesheet, holiday pay, "working shown" |
| Words we never use | Streamline, solution, hassle-free, best-in-class, next level, exclamation marks, plus the client's pinned list of 15 banned words |
| Numbers | Numerals with units; every figure must trace to the facts sheet or a customer comment |
| Structure | Headline of nine words or fewer, a subhead naming the outcome, three proof points, one call to action using the verb "Try" |
| Legal lines | Never "compliant" or "guaranteed"; never name a law |
The anchors do more than the rules. Paste the three paragraphs the client likes most and tell the model to match their rhythm, because examples tend to steer a model more reliably than adjectives do. For a deeper look at building this layer for each client, see protecting each client's voice step by step.
The draft pass: prompt, first output and the claims list
The invented facts sheet for this job says the calculator handles fixed-hours, part-time and irregular-hours staff, shows its working for each employee, imports timesheets in two formats, and that a six-customer pilot cut holiday-pay checks from about 40 minutes to about 10 a pay run. The client supplied that last figure. Everything the page may say comes from that sheet or the customer comments. The prompt enforces it:
You are drafting a landing page for the client described in the attached pack:
brief, style guide, product facts sheet, customer comments and three voice anchors.
Write: three headline options (nine words or fewer), the page (about 900 words)
in the sections the brief lists, and a meta description of 150 characters or fewer.
Rules:
- Follow the style guide exactly, including the words we never use.
- Match the sentence rhythm of the voice anchors.
- State only facts that appear in the product facts sheet or the customer
comments. If a sentence needs a fact you do not have, write [FACT NEEDED].
- No statistics, studies or customer names unless they are in the pack.
After the page, give a table of every factual claim in it, with the pack
document and section it comes from. If a claim has no source in the pack,
write "none".This is the opening of a first pass that breaks several of those rules, which is typical, and exactly why the later stages exist:
Holiday pay, sorted in seconds
Streamline your payroll with our powerful, hassle-free holiday-pay solution, built to take your business to the next level. Whether your people work fixed hours or irregular shifts, you can calculate holiday pay accurately and keep your records in one place. Join over 500 businesses that trust us to save hours every pay run.
And these are the first rows of the claims table the same run returned, the ones covering that opening:
| Claim in the draft | Source the model gave |
|---|---|
| Handles fixed-hours and irregular-shift staff | Facts sheet, section 2 |
| Keeps records in one place | Facts sheet, section 3 |
| Used by over 500 businesses | None |
| Saves hours every pay run | None |
| Sorted in seconds | None |
The table earns its place by exposing three problems. The page uses four terms from the client's banned list. It carries a customer count that appears nowhere in the pack. And two of its benefits, "hours" and "seconds", have no source, because the facts sheet says 40 minutes down to 10. The model told you which claims were unsourced, which is the useful behaviour. It did not stop itself from writing them, so the edit stage has to.
Run a rhythm check as well, because voice drift shows up in sentence length before anyone can name it. Take ten sentences from the voice anchors and ten from the draft, and read each set's average words per sentence from your editor's readability statistics, or count them in a spreadsheet. Suppose the anchors average 13 and the draft averages 21, with the longest at 38. Nothing in the draft is wrong as such, but it will read to the client like someone else wrote it, which is the complaint you most want to avoid.
The human edit: six changes with before and after
Edit in a fixed order so effort goes where it counts: claims first, so you don't polish a sentence that will be cut, then the lead, then rhythm, then specifics, and a read-aloud last. Six of the changes made to the first pass are worth showing, because each is a pattern you will meet on every job. For a longer run of rewrites, a freelance copywriter's edit workflow goes through more.
| Change | AI first pass | After the edit |
|---|---|---|
| A plain lead | "Streamline your payroll with our powerful, hassle-free holiday-pay solution, built to take your business to the next level." (18 words, four banned terms) | "Holiday pay, worked out for you. Enter the hours, and the calculator shows its working." (two sentences, 15 words, using the client's own phrase) |
| A real benefit | "...to save hours every pay run." | "In a pilot with six customers, holiday-pay checks fell from about 40 minutes to about 10 per pay run." The limits of the evidence stay in the sentence |
| Proof from a customer | "Join over 500 businesses that trust us." | Cut. Replaced by customer comment 4, quoted exactly, with the permission note from the pack |
| Rhythm | "Whether your people work fixed hours, part-time patterns or irregular shifts, the calculator works out what each person has earned in holiday pay and shows you the full calculation so that you can check every figure before the pay run is approved." (42 words) | "Fixed hours, part-time or irregular shifts: the calculator works out holiday pay for each person. It shows the sum, so you can check every figure before you approve the run." (two sentences, 15 words each) |
| The call to action | "Get started today and transform the way you handle payroll!" | "Try the calculator on your next pay run." The guide's verb, no exclamation mark |
| A legal line | "Stay fully compliant with holiday-pay rules." | "The calculator shows how each figure was reached." It says what the product demonstrably does |
Notice what the edit is not. It is not cosmetic polishing of a draft that was basically fine. Four of the six changes remove a claim the page could not support or a phrase the client has banned, and the other two bring the rhythm and the call to action back in line with the style guide. After the full edit, which covers more than these six changes, the page's average sentence length is down from 21 to 14 words, close enough to the anchors' 13 that the client's own team would not notice the seam.
The fact-check log, claim by claim
Start from the model's claims table, then read the page for claims the table missed, because the model lists only what it knows it asserted. On this job the model's table for the whole page had eleven rows. Reading the page found a twelfth, "the most accurate holiday-pay calculator available", a superlative the model had not listed. The log below shows the seven claims that needed a decision; the other five were product facts copied straight from the sheet and verified in a minute. Give every claim one of four statuses: verified with a named source, corrected, cut, or queried with the client. "Probably fine" is not a status.
| Claim | Source checked | Status |
|---|---|---|
| Handles fixed-hours, part-time and irregular-hours staff | Facts sheet, section 2 | Verified |
| Shows its working for each employee | Facts sheet, section 2, and the demo screenshot in the brief | Verified |
| Imports timesheets in three formats | Facts sheet says two formats | Corrected to two |
| Checks fell from about 40 to about 10 minutes | Client's pilot summary, six customers | Verified by the client's email; wording keeps "pilot with six customers" |
| Used by over 500 businesses | Nothing in the pack | Cut; client asked whether a real figure exists |
| The most accurate calculator available | Nothing, and unprovable | Cut |
| Stay fully compliant | A legal claim with no source | Cut |
The client query is worth sending the same day, because it is the one that can hold up delivery. Phrase it so it can be answered in one line: "The draft would like to say how many businesses use the product. Is there a figure you are happy to publish, and where does it come from?" Where the answer is no, the page is better without it, and nobody has been misled.
The plagiarism scan: what it costs and what a hit looks like
Run the edited copy, not the raw draft, because the edit is what you are delivering. Copyscape's published notice sets its pay-per-search price at 3 cents for the first 200 words plus 1 cent for each additional 100 words or part of 100, so the 900-word page costs 3 + 7 = 10 cents. Grammarly Pro also includes a plagiarism checker, at $12 per member a month billed annually or $30 monthly, and it is the cheaper habit if the subscription is already in your budget. Use whichever you will actually run on every job, and save the report to the client folder.
A scan like this returns hits, and most are harmless. Here is an illustrative result for the page, with the triage for each:
- The client's own product page. Expected, because the facts came from it. No action; note it in the log so nobody wonders later.
- A nine-word phrase on 14 pages. "Automatically calculates holiday pay for part-time and irregular-hours staff" is stock category language. Reword it, because a distinctive page is better than a generic one.
- A 14-word run matching a competitor's feature list. Reword and re-run, then note the re-run in the log. A client should never find their page echoing a rival's.
A workable triage rule: the client's own text is fine, a few common words are ignored, a run of eight or more words matching a third party gets reworded, and anything matching a competitor is reworded and checked again. The wider routine is in checking AI content for plagiarism before you publish. Models also tend to reuse favourite constructions, so every so often run your last two or three deliveries through the same check. A headline pattern that repeats across clients is a quality problem even when it is not plagiarism.
If the client asks for an AI-detector score
It happens, and the useful response is a calm one that doesn't hinge on a number. Detectors are probabilistic, and their own vendors say so: Originality.ai's guidance states that "no AI detector is perfect", that it can produce false positives, and that scores "should be treated as a signal, not as proof of misconduct". Its advice to writers who are questioned is to gather evidence of their writing process and to review their writing agreement and the project's AI policy.
That advice translates into a delivery habit. Offer a process log with every job: the brief, the prompts you used, the first draft, your edit history, the claims log and the plagiarism report. Agree in the contract what evidence you will provide if a score is ever disputed. A score can be wrong in either direction, while a dated trail of how the copy was made answers the question the client is actually asking, which is whether a person stood behind the words.
The delivery pack, filled in
Deliver one folder, not an attachment and an apology. For this job it holds the clean copy and a version with the changes visible, the claims and sources log, the plagiarism report, the three headlines with a recommendation, the open queries, and the process folder. The covering note carries the AI-use line in plain words:
Hi [first name], the landing page is attached with three headline options. I'd choose option 2 because it names the outcome. Every factual claim is in the attached log with its source, and two items need your confirmation: the customer count, which I've left out, and permission to show the working screenshot. The copy was drafted with AI assistance from your brief and facts sheet, then written, edited and fact-checked by me, and the final text was plagiarism-checked on [date]; the report is attached. As agreed, you have two revision rounds within seven days. [your name]
Nothing in that note apologises for using AI or hides it. It says what the tool did, what you did and what the client still needs to decide, in that order, and the third part is the one the client has to act on.
Where the 335 minutes went, before and after
The point of the workflow is not speed alone, so look at where the time moved. These are the invented minutes for this job, against how the same page might have gone without AI:
| Stage | Without AI | With AI |
|---|---|---|
| Reading the brief and research | 60 min | 40 min |
| Outline and angle | 20 min | 10 min |
| First draft | 150 min | 35 min (10 setting up, 25 running and reading) |
| Edit for voice and rhythm | 60 min | 75 min |
| Fact-check and client queries | 20 min | 35 min |
| Plagiarism scan and triage | 5 min | 10 min |
| Delivery pack | 20 min | 15 min |
| Total | 335 min | 220 min |
On the flat $850 fee that is about $152 an hour without AI (335 minutes is 5.6 hours) and about $232 an hour with it (220 minutes is 3.7 hours). Drafting fell by 115 minutes, while editing, fact-checking and scanning together grew by 35. That ratio is the healthy one. If your editing and checking time falls as drafting time falls, the quality is going somewhere you will eventually hear about.
The tools for a month like this cost little. One assistant is $20 (ChatGPT Plus, or Claude Pro at $17 billed annually or $20 monthly), Grammarly Pro is $12 a month billed annually, and Copyscape runs about 10 cents a page. Spread across six jobs of this size that is roughly $5 a job. If you need a client's confidential material kept out of model training and want a business plan, note that ChatGPT Business and Claude Team both need two seats. The breakdown of a freelancer's monthly AI tool budget compares the options in more detail.
What the job taught
- Measurable rules beat adjectives. Sentence length, a banned-word list and "Try" as the call to action caught more drift than "plain and practical" ever would.
- Budget for the fact-check. The model produces the claims that need checking, so the checking grows with the draft. Thirty-five minutes was about right for a 900-word page.
- Send client queries early and short. One unanswered question can hold the whole delivery, and one-line questions get one-line answers.
- Keep the trail. Prompts, first draft, edit history and claims log cost nothing to save and answer most doubts about authorship before they become disputes.
- Don't sell hours you didn't work. A flat fee rests on the page's value and your checks. If a client is billed by the hour, bill what you did.
None of this needs a special tool. A client pack, a style guide with rules you can test, a claims table, a plagiarism scan and a delivery note work with any of the major assistants, and the habits transfer when one product changes or gets replaced. To see how different assistants handle the same brief, one brief run through three assistants compares their drafts side by side.
Questions freelance copywriters ask about AI and clients
Should I tell clients I use AI?
Yes, in your terms and in the delivery note, and check the client's own AI policy before you start. Some clients restrict or forbid it, and some contracts say what you must disclose. A short line in the proposal and the delivery pack costs nothing and removes any surprise later, whatever the client decides.
Can an AI-detector score prove my copy is human?
No. Originality.ai's own guidance says no detector is perfect and that scores are a signal, not proof. If a client asks for one, explain that limit and offer a process log instead: the brief, your prompts, the first draft and your edit history show how the copy was made, which a single score never can.
How should I price AI-assisted copy?
Price the outcome and the scope, not the hours the draft happened to take. A flat fee for a landing page rests on the value of the page, your edit and your checks. If you bill by the hour, bill the hours you actually worked, and say in your terms which tools you use.
Do I need a plagiarism check if the draft is AI-written?
Yes. A model can echo phrases that exist elsewhere, and a client will assume you checked. Run the final edited copy, not the raw draft, and keep the report. At Copyscape's published rate, a 900-word page costs about 10 cents, so the check is cheap against the risk.
Further reads
- AI Copywriting Tool Examples: Before-and-After Rewrites With Prompts — More before-and-after rewrites with the prompts used
- How to Write a House Style Guide Your Team and AI Both Follow — Write a style guide that people and models both follow
- Best AI for Copywriting Ads: Variants, Testing and Client Sign-Off — Ad variants, testing and client sign-off
- ChatGPT Prompts for Copywriters: Research, Angles and Edits — Prompts for research, angles and edits
- Voice of Customer Research With AI: Mine Reviews and Calls for Copy — Mine reviews and calls for the client's own words
- How Freelance Writers Use AI Without Losing Their Voice — Protecting your own voice as a writer
- Best AI Tools for Freelance Writers: Research, Editing, Client Admin — Tools for freelance writers sorted by client-trust risk, with prices, sample prompts, a source ledger and wording for telling clients.
- Can AI Help Copywriters? Research, Drafts and Edits for Client Work — A three-stage workflow for freelance copywriters: AI for research, drafts and edits, with a voice guide per client and a fact-check before delivery.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Copyscape Premium page, Originality.ai guide to AI content detector false positives, OpenAI ChatGPT and ChatGPT Business pricing pages, OpenAI help article on data controls, Claude pricing page (all read on 4 November 2026), and Grammarly's plans page in dollars (archived copy of 3 November 2026).