Can AI Help Copywriters? Research, Drafts and Edits for Client Work

Yes: use AI for three jobs, research that condenses the client's own material, first drafts built from their real words, and narrow edit passes against a written voice guide. You keep the angle, the fact-check and the final read. Work on a plan that does not train on client material, and tell clients how you use it.

Most copywriters who try AI hit the same wall: the draft is fluent and wrong for this client. It sounds like an average of every brand, because that is what a blank prompt produces. A cleverer prompt rarely fixes that. A short voice guide per client, built once from real material, does, together with a fixed order of work so the machine never decides the message.

Follow me on Instagram@sagnikteaches

Where AI fits in a paid copy project

A client project has four stages: understanding the business, deciding what to say, writing it, and tightening it. AI is useful at the first, weak at the second, decent at the third if it has good raw material, and very good at certain parts of the fourth. The second stage, the decision about what the page should claim and to whom, is the one clients actually pay for, so it stays yours.

Connect on LinkedInSagnik Bhattacharya

The times are illustrative. The point is where the saving sits: in reading and reshaping material, not in deciding what to say.

Subscribe on YouTube@sagnikteaches

The client file that makes everything else work

Before any drafting, build one document per client and keep it in a Claude Project or ChatGPT Project (or a plain document you paste in). Both let you keep project instructions and uploaded files together, so the voice guide travels with every request in that project. On training, OpenAI and Anthropic both state that content on their business tiers (ChatGPT Business and Enterprise, Claude Team and Enterprise) is not used to train models by default. Consumer plans differ: OpenAI's help centre says the training setting is on by default for ChatGPT Free, Plus and Pro, and Claude asks each consumer user to choose. Check the model-training setting in your privacy controls before the first client brief goes in.

A usable client file has five parts, none longer than a page:

1. The promise. One sentence on what the client does for whom, in their words.
2. Voice samples. Three to five real pieces: a newsletter they wrote themselves, an email reply to a customer, a voice-note transcript. Longer is not better; variety is.
3. Do and don't list. Words they use, words they avoid, how formal, whether they use contractions, how they greet readers.
4. Proof points. Facts you have verified: years trading, certifications, numbers they have confirmed in writing.
5. Banned claims. Anything legal or compliance has ruled out, plus superlatives they cannot back up.

Here is a filled-in extract for an illustrative client, a furniture maker with a six-person workshop:

CLIENT: Workshop that builds solid-wood dining tables to order
PROMISE: "A table that outlasts the people who buy it, made by the people you can phone."
VOICE: Plain, warm, a bit dry. Contractions always. Says "we" not "the company".
Addresses reader as "you". Short sentences, then one long one for rhythm.
USES: "made to order", "oak", "walnut", "proper joints", "come and see it"
NEVER: "bespoke", "artisanal", "curated", "luxury", exclamation marks
PROOF (verified by owner 12 Oct): 22 years in the workshop; 5 to 7 weeks lead time;
tables finished with hard-wax oil; delivery is by the makers, not a courier.
BANNED CLAIMS: "lifetime guarantee" (it is a 10-year guarantee), "sustainably sourced"
(no certificate yet).

Notice the banned claims. A model will happily write "lifetime guarantee" because it sounds right. A line in the file stops it. This is also the cheapest place to prevent the mistakes that cost you a client.

Stage one: research that points at the customer

AI is at its best here because the task is reading a lot and condensing it. Give it material you have gathered, not an open request to "research the market". Material you have: the client's reviews, support emails, competitor landing pages you saved, a recorded kickoff call.

A prompt that works for review mining:

Below are 40 customer reviews of an online clothing store. Do not invent anything.
1. List the five reasons customers say they bought, ranked by how many reviews mention them,
   with a count.
2. List three recurring complaints, with a count.
3. Pull out 10 short phrases that customers used themselves (exact wording, in quotes)
   that could work in headlines.
4. Say which reviews you were unsure about and why.
[reviews pasted here]

A realistic output (illustrative) might say that 17 of 40 reviews mention fit being true to size, 11 mention fast returns, and 9 mention fabric feeling heavier than expected. It would quote phrases such as "fits like it was cut for me". What you fix: the counts. Models miscount, so open the reviews, search for "fit" and "size", and confirm the 17. If it is 14, the headline "most customers" becomes "many customers". Counts are exactly the sort of detail a client's sceptical finance director will check.

Competitor scan without copying

Paste three competitor pages and ask for the claims each makes, grouped by theme, and the claims nobody makes. The second list is the useful one: a gap such as "nobody mentions repair" is a positioning idea. Do not ask for a rewrite of a competitor's page. That produces copy that is too close to someone else's and puts both you and the client at risk.

Interview transcripts

If you record kickoff calls (with permission), have the assistant pull out stories, numbers, and phrases the founder repeats. A pest-control firm owner may say "we come back for free if they come back" three times in an hour. That is the guarantee line, in the owner's voice. The transcript is also where the best proof points live, so ask the model to mark any figure with the timestamp so you can verify it with the client.

Stage two: choose the angle yourself, then let AI argue with it

Write the angle in one sentence before you open a chatbot. Then use the assistant as a sparring partner: ask for the three strongest objections a sceptical reader would raise, and the one thing in the brief that does not fit your angle. The decision stays yours; the assistant's job is to look for holes.

My angle for a car repair garage's homepage: "You'll know the price before we touch the car."
Give me:
- three objections a nervous first-time customer would have to that promise
- two things in the garage's own materials (pasted below) that contradict it
- one claim I should not make because I cannot prove it from this material
[client materials pasted here]

A sample answer might point out that the garage's own FAQ says prices can change after diagnosis. That is a real contradiction, and it is better to find it in your office than on the client's live site. The fix is a rewritten promise: "A written estimate before work starts, and a call before anything extra."

Stage three: drafting with the client's own words

The draft prompt has a fixed shape: role, client file, brief, constraints, and a request to leave visible gaps rather than invent. The last part matters. Ask the assistant to put square brackets where it needs a fact you have not supplied.

You are drafting copy for the client described in the file below. Use the voice samples and
the do/don't list. Use only facts from the PROOF section; where you need a fact that is
not there, write [NEED: what is missing] instead of guessing.
Task: homepage hero (max 14 words), three benefit blocks (40 words each), one closing line.
Angle: [my one-sentence angle]
[client file pasted here]

What comes back (illustrative, for the furniture maker): a hero line, three blocks, a closing line and probably two or three [NEED: ...] flags, such as "[NEED: lead time for walnut]". Those flags are a gift. They are your list of questions for the client, produced in 30 seconds.

What you fix in a draft, in order:

1. The first sentence. Models open with scene-setting. Cut to the sentence that carries the promise.
2. Any adjective doing the work of a fact. "Exceptional craftsmanship" becomes "dovetail joints, cut by hand".
3. The ending. Drafts tend to close with a summary line that restates the page. Delete it or replace it with the single action you want.
4. Rhythm. Read it aloud. If three sentences in a row have the same length, break one.

If you write on behalf of a business owner who also writes themselves, put a paragraph they wrote in the prompt and tell the model to imitate its sentence length and punctuation, not its topic. Imitating style from a real sample works far better than describing tone with words such as "friendly" or "professional".

Worked project: a product page and emails for a clothing store

Take an illustrative engagement: a small online clothing store wants a product page for a heavy cotton overshirt and a three-email launch sequence. The fee is a flat price, and you have a 4-day window.

Day one, research (about 90 minutes). You paste 40 reviews of similar items from the store's own catalogue and last year's customer service emails. The assistant returns the five buying reasons and three complaints. You verify counts, find that "feels heavier than expected" appears in 9 reviews, and decide the page should lead with weight and warmth, framing "heavy" as a feature. You send two questions to the client: the fabric weight in grams, and the return window.

Day two, angle and draft (about 2 hours). You write the angle: "The overshirt that replaces your light jacket from October to March." You run the objection prompt and learn that the sizing chart contradicts the "fits true" claim in the reviews, because the chart runs a half size small. You raise it with the client before it goes live. Then you draft: product page, three emails. The assistant leaves [NEED: fabric weight] and [NEED: return days] flags, filled once the client replies.

Day three, edit (about 90 minutes). You run the edit pass described below, accept roughly half the suggestions, and read the lot aloud.

Day four, delivery. You send the copy with a short note listing every factual claim and where it came from (review count, client's email, spec sheet). That claims table is the thing the client's team will actually use to sign off, and it covers you if a number is later challenged.

A rough sum, illustrative: without AI you might budget 14 hours for this project; with the workflow it comes to around 9 to 10 hours. The saving is in reading, first drafts and the mechanical edit. Do not count on halving the time, and do not count on the first draft being close.

Stage four: edit passes that protect the voice

Run separate passes rather than one "improve this" prompt. A single instruction to improve invites the model to rewrite everything towards its own average. Narrow passes keep your hands on the wheel.

Pass one: voice drift

Compare the draft to the voice samples and the do/don't list in the client file.
List every sentence that sounds unlike the client, with a one-line reason.
Do not rewrite anything. Do not comment on content.

Reading the list is quicker than hunting through the draft yourself, and you decide what changes. A typical result for the furniture maker: "Sentence 4 uses 'crafted with passion', which does not appear in any sample." You replace it with a concrete detail.

Pass two: claims and proof

List every factual claim in the draft (numbers, guarantees, comparisons, "best", "only").
For each, say whether it appears in the PROOF section. Mark unsupported ones UNSUPPORTED.

This pass catches the "lifetime guarantee" problem. Also check any "first", "only" or "most" claim by hand. Advertising rules in many places require substantiation, and a comparative claim without evidence is your client's problem and then yours.

Pass three: clarity and length

Ask for sentences over 25 words, repeated words within a paragraph, and passive constructions. Then fix them yourself. Tools such as a plain readability checker do this too; the benefit of asking the assistant is that it can group the issues by paragraph.

Pass four: your own read

Read the final copy aloud, out of the chat window, in the document the client will receive. Fix what your ear catches. No model replaces this pass, and it is where the copy starts to sound like a person.

Facts, checks and what to never paste

Assistants make confident errors, and in copy the errors are usually specific ones: a dated statistic, a misquoted feature, an invented testimonial. Adopt three habits.

• No claim without a source in your file. If it is not in the PROOF section, it is either verified now or deleted.
• Never accept a quote or a statistic the model produced. Find it at the original source or leave it out. "Studies show" is a red flag in a draft, and you should not see it in a delivered page.
• Treat product facts as perishable. If copy mentions a software feature or price, open the vendor's page the day you deliver.

On privacy: check what the client's contract says about third-party tools before you paste anything. If you use a chat assistant at all, use a plan that excludes your content from training (business plans do by default; on a consumer plan, check the training setting and switch it off) and paste only what the task needs. Strip customer names, order numbers and anything under NDA. If a client says no AI tools, that is the end of the discussion, and you can still use your own checklist-based editing.

Telling clients, and putting it in writing

Disclosure is moving from courtesy to contract term. Some clients ask for a clause, some ban AI-generated text, and some publishing platforms set their own rules. Amazon's Kindle Direct Publishing is a useful example for anyone who ghostwrites books: its content guidelines require publishers to declare AI-generated text, images or translations, count text as AI-generated even after substantial edits, and do not ask for a declaration when the work is AI-assisted, meaning you wrote it and a tool helped edit, refine or check it. If a client will publish there, note in your project file which of the two your process produced, because the client makes the declaration and needs to know.

Decide your own position before clients ask. A workable approach is a two-line paragraph in your proposal or terms:

I use AI tools for research, outlining and editing passes. Every word delivered is written or
rewritten by me, checked against sources you approve, and I remain responsible for it. I do
not paste your confidential material into tools that train on it, and I'll tell you which tools
I use if you ask.

That wording is a starting point, and your contract may need something firmer. See the tutorial on whether freelancers should tell clients they use AI and the one on writing an AI disclosure statement for fuller options. Some clients also want assurance on ownership of the work; if a contract clause worries you, ask a lawyer rather than relying on a chatbot's reading of it.

Five quick cases from other kinds of client

A pest-control firm's quote follow-up email. Paste the firm's three best-performing past emails and ask the assistant for three subject lines in that style. You pick one, check it contains no guarantee the firm does not offer, and test it against their last open rate.

A car repair garage's service pages. You use the assistant to turn a mechanic's rambling voice note into a list of what the service includes. Gaps become [NEED: ...] questions, such as whether the oil change includes a filter. The mechanic answers in two minutes.

A before and after. A draft line reads: "Our passionate team delivers exceptional results." The edit pass flags it as unlike the client. The final reads: "Two of us answer the phone, and we both know your car." The change came from a voice sample, not from the model.

A realistic mistake. A writer lets the assistant draft a testimonial block "based on reviews". The block combines two customers' comments into one invented quote with a first name. The client's customer sees it and asks who she is. Rule: only real, attributed quotes, with permission, copied exactly.

A subject-line test. You ask for 20 subject lines for the overshirt launch and a score for each against the client's do/don't list. Three survive. You send two to the client to pick, and the list becomes the starting point for an A/B test the client can run in their email tool.

How to tell the workflow is working

Track three numbers over your next five projects: edit time per finished page, number of client revision rounds, and number of factual corrections caught after delivery. If edit time drops and revisions stay flat or fall, the workflow is helping. If factual corrections rise, tighten the claims pass or slow the draft stage. If a client says the copy "doesn't sound like us", your voice samples are too thin. Add more real writing from the client and rerun.

A once-a-quarter check is enough: take one delivered project and compare the final copy with the first AI draft. If most of the first draft survives, either the voice guide is excellent or you are editing too lightly. If almost none survives, the draft stage is adding little, and the research and edit passes are where the value is. Both outcomes are fine information. If you want to go further on edit passes alone, read a freelance copywriter's edit workflow, and if client voice is your main worry, protecting each client's voice step by step goes into the voice file in more depth.

For wider context on how freelancers handle the admin around the writing, invoices, proposals and chasers covers the parts that also tend to eat unbilled time. The AI hub collects the rest of the series.

Questions freelance copywriters ask before using AI on client work

Do I have to tell clients I use AI?

Check the contract first, because some clients now ban or require it. Even where nothing is written down, a short line in your proposal on what AI does in your process (research, outlines, edit passes) avoids awkward conversations later. Say what you do, and that you are accountable for every word delivered.

Should client briefs and customer data go into a chatbot?

Only on a plan whose terms say business content is not used for training, and only what the task needs. Strip customer names, order numbers and anything covered by an NDA before pasting. If the client's contract forbids third-party tools, believe it and work without.

Which AI model is best for copy?

None wins every time, and versions change monthly. Test two assistants on the same voice guide and the same brief, then keep the one whose draft needs the fewest edits. Judge by edit time per page, not by how polished the first paragraph looks.

Can I charge less because AI speeds things up?

Most copywriters charge for the outcome and the judgement, not the hours typed. If a project takes less time, the saving can go into extra rounds of testing or a second angle. Tell clients you charge per deliverable so speed does not become a discount by default.

How do I stop AI copy sounding like everyone else's?

Feed it the client's own words, not adjectives about their tone. Real customer emails, past newsletters and the founder's voice notes carry vocabulary and rhythm that a description cannot. Then edit against a list of phrases the client would never say.

Further reads

• How Freelance Writers Use AI Without Losing Their Voice — Keeping your own voice while using AI.
• AI Clauses for Agency Contracts: Disclosure, Ownership, Approvals — Contract clauses for disclosure, ownership and approvals.
• Best AI for Copywriting Ads: Variants, Testing and Client Sign-Off — Ad variants, testing and client sign-off.
• AI Tools for Copywriting: A Freelancer's Draft-to-Delivery Workflow — Tool choices from first draft to delivery.
• How to Mention AI Use in Client Contracts and Proposals — Mentioning AI use in contracts and proposals.
• Should You Tell Customers When You Use AI in Your Marketing? — Whether to tell audiences about AI in marketing.
• Best AI Model for Copywriting: How to Test ChatGPT, Claude and Gemini on One Brief — An owner's brief-test scorecard: one brief, one voice sample and one edit round through ChatGPT, Claude and Gemini, and where each needs heavy editing.
• Voice of Customer Research With AI: Mine Reviews and Calls for Copy — From a folder of reviews and call transcripts to ranked themes, verbatim phrases and a swipe file, with checks that stop the AI inventing quotes.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI business-data page and ChatGPT data-controls and Projects help articles; Anthropic privacy centre articles on model training (commercial and consumer) and Claude's Projects help article; Amazon KDP content guidelines on AI-generated and AI-assisted content. All read 4 November 2026.