AI for Copywriting Cannot Replace the One Stage Most Teams Wrongly Hand It

By Frederik Baun, conversion copywriter and founder of Words Sounds Visuals

I've run a five-stage AI copywriting process on every client project since 2018, the year I started Words Sounds Visuals.

What Is the AI Copywriting Workflow, Stage by Stage?

A copywriting project runs through six stages: research, brief, draft, variation, edit, and fact-check.

I use AI at four of those six on every client project. The other two, the brief and the final edit, still come from me, because both decide what the copy needs to say rather than how it says it.

Most guides to AI for copywriting skip straight to a tool list. That’s useful once you know which stage you’re solving for, but it answers the wrong question first.

The question that changes your results is which stage you’re handing over, since one of those six stages is the one most teams give away without noticing.

Which Stage of Copywriting Does AI Handle Best?

Drafting and variation writing, without much competition.

I ask Claude for a rough first pass on a client brief and I’m editing inside two minutes, instead of staring at a blank page for twenty.

Ten ad headlines that used to eat an hour now come out of one prompt, and my job shrinks to picking the three worth testing.

Research assembly sits close behind. Before I write for a new client, I pull competitor pages, reviews, and forum threads to find the exact words their buyers use, and a model summarizes that pile faster than I can read it start to finish.

Those three tasks used to eat roughly half a working day on a mid-sized project: the blank page, the variations, and the research pile.

I cover the full breakdown of what moved and what didn’t in my post on what AI copywriting replaces for a working copywriter.

None of those three tasks decide anything. They produce raw material for a decision someone still has to make.

Which Stage Should You Never Hand to AI?

The brief.

A brief decides what the copy needs to say before a single sentence gets written.

Which objection kills the sale if you skip it, which proof point the reader will believe, and which angle a tired buyer responds to today rather than the one that worked last campaign.

A model wasn’t on the sales call where a prospect said, almost in passing, the exact phrase that would close them. It wasn’t in the support inbox when three customers used the same three words to describe the same frustration in the same week.

Feed it a topic instead of a brief, and it fills the gap with the most statistically likely answer for that industry in general. That answer is competent and belongs to nobody, because it was built for anybody.

I’ve watched a version of this brief go missing on a launch sequence a founder ran without a copywriter last year. The emails went out on schedule and read fine at a glance.

They converted at roughly a third of what the previous launch pulled, because nothing in them said anything specific enough to move a buyer who’d already read four competitors that week.

Why Do Teams End Up Handing AI the Brief Without Meaning To?

Because the brief doesn’t look like a stage. It looks like a step you can skip.

Drafting has a visible product: a paragraph appears, so it feels like the work happened. The brief has no visible product beyond a page of notes that never ships, so it’s the first thing a rushed team cuts.

Open a chat window and type “write me a landing page for a project management tool,” and the model doesn’t refuse. It answers, confidently, using the average landing page for that category rather than the one true thing that makes your tool worth switching to.

That’s the trap. The prompt reads like a request for a draft, but it’s a request for a decision underneath, and the model makes that decision for you by default because you never told it what to decide.

How Do You Know If Your Team Has Already Skipped the Brief Stage?

Five signs I check for on any client account before I touch a single sentence.

1. The prompt starts with the topic, not the objection. “Write a landing page for our project management tool” has no objection in it anywhere.
2. Drafts read fine and convert to nobody in particular. Grammar is clean, and nothing on the page would make a specific reader stop scrolling.
3. Nobody named the “who” before the first prompt went out. Ask who the reader is mid-scroll, and the answer is a demographic, not a moment.
4. The same proof line shows up across three unrelated products. “Trusted by thousands” fits a project management tool, a dog leash, and a mattress equally badly.
5. Editing a draft means fixing sentences, never questioning the angle. The edit pass polishes phrasing and never once asks whether this was the right message to send.

Two or more of those on a page I’m reviewing, and I go back to the brief before I touch the copy itself.

What Happens When AI Writes the Brief Instead of a Person?

The copy gets smoother and less specific at the same time, which is the part most people don’t expect.

A model, left to fill in strategy on its own, reaches for the safest version of every claim: “trusted by thousands,” “built for growing teams,” “the platform that scales with you.” Every sentence is grammatically fine and true of roughly a thousand other products.

Readers pattern-match that smoothness to marketing voice within a sentence or two, then skim for the offer and ignore the rest.

I go through why polish costs more than it saves in my full definition of AI copywriting, where the rougher, brief-driven draft consistently outperforms the polished one with no brief behind it.

Writing better prompts doesn’t solve it. Writing the brief before you open the chat window does, so the model has a decision to execute instead of one to invent.

Can AI Help With the Research Stage of a Copywriting Brief?

Yes, and this is where it earns its keep before the brief gets written.

A model can pull the recurring phrases out of a pile of reviews, forum threads, and competitor pages faster than a person can read the pile once. That’s raw material, not a finished brief.

The difference matters. Research tells you what buyers say. The brief decides which of those things to lead with, and that call still needs a person who’s talked to the client and knows which objection is killing the sale right now, not last quarter.

I run this stage the same way on every new client: pull the raw phrases with AI, then sit with a highlighter and pick the three that map to a real objection I’ve heard on a call.

A SaaS client’s reviews once handed me forty near-identical complaints about a competitor’s onboarding taking “half a day.” A model surfaced that phrase in a summary within minutes.

Deciding to lead the whole homepage with “set up in twenty minutes, not half a day” instead of the client’s original headline about features was a call the summary couldn’t make for me.

Which AI Model Should Handle Which Stage?

I cover the Claude-versus-ChatGPT question in more depth in my full comparison of the two for copywriting work, since the table above is the short version.

How Do You Write a Brief AI Can’t Write for You?

Four questions, answered before you open a chat window:

1. Who is reading this, and what did they just do? Not a demographic. The exact moment: they just abandoned a cart, just read a competitor’s pricing page, just got off a sales call where they said “let me think about it.”
2. Which objection kills this if you skip it? Name one. Not three, one, and put it in the first paragraph the model sees.
3. What proof backs the claim you’re making? A number, a named result, a specific detail from a client conversation. Vague proof produces vague copy, because the model can’t source something you never gave it.
4. What should the reader do next, in one sentence? Not “learn more.” The actual next click, and what happens on the page they land on.

Write those four answers in plain sentences, hand that to the model instead of the topic, and the draft that comes back has somewhere to stand.

I keep this same four-question structure in the free marketing skills I built for Claude Code, so the offer architect skill runs this exact check before any drafting starts.

What Belongs in a Brief for AI That a Junior Copywriter Wouldn’t Need Spelled Out?

A junior copywriter carries context a brief never has to state. They sat in on the last client call, remember the campaign that flopped in March, and know the founder hates the word “solution.”

A model carries none of that unless the prompt spells it out, so a brief written for AI needs three things a brief for a person could skip.

The client’s own phrases, quoted directly from a call or review, not paraphrased into marketing language. A model paraphrases by default, and paraphrasing is where the specific detail dies first.

The competitor being compared against, named. “Better than the alternatives” gives a model nothing to differentiate from, while “faster setup than Monday.com” gives it an actual claim to prove.

The angle that already failed, stated plainly. Without it, a model has a real chance of proposing the exact angle a client rejected two campaigns ago, since it has no memory of that meeting.

How Long Should Writing a Brief Take?

Fifteen to twenty minutes, for most single-asset projects like an email or a landing page section.

That’s the time it takes to answer the four questions above in full sentences, not bullet points, and pull two or three real client phrases from a call transcript or a review.

Longer than that, and the brief is probably trying to do the drafting’s job too, with full paragraphs of suggested copy instead of decisions. Shorter than that, and at least one of the four questions got a guess instead of an answer.

I time this on client work because it’s the easiest stage to let slide when a deadline is close, and it’s the fifteen minutes that decides whether the next two hours of AI drafting produce something worth sending.

Where Does the Edit Stage Fit After AI Drafts?

Right after the draft, and it’s not optional.

Every AI draft goes through the same checklist I’d run on a junior writer’s work: does the opening line earn the second sentence, does every claim have a source, does the rhythm sound like a person instead of a smoothed-out average.

I keep that process in a 25-point editing checklist that catches sourcing gaps and paragraph density in one pass.

Before anything client-facing ships with a factual claim in it, it goes through the same fact-check process I’d apply to a stat a junior handed me.

A model will state a percentage with total confidence and no source behind it, and that’s the sentence that gets a client sued or embarrassed if it ships.

Both of those checks assume the brief was solid going in. Run them on a draft with no brief behind it, and you’ll catch the sourcing problems and still ship copy that says nothing specific, because the checklist edits sentences, not strategy.

Does Skipping the Brief Show Up in Conversion?

Every time I’ve seen it happen, yes, and it shows up as a flat page rather than a broken one.

A retail client of mine ran a batch of 40 product descriptions through a model with a one-line prompt per product and no brief. They read fine. Nothing was wrong at the sentence level.

They also converted at roughly the same rate as the generic manufacturer copy they replaced, because nothing in them said anything a shopper couldn’t already read on the box.

I rewrote the top-selling ten with an actual brief, three lines each on who buys this product and what they’re worried about before purchase, and those ten outsold the other thirty within the month.

Same model, same afternoon of writing. The only variable that changed was whether a person decided what to say before the model said it.

I’ve seen the reverse mistake too, where a good brief gets undercut by a rushed edit.

A B2B client’s brief named the exact objection, but the draft that shipped kept a generic opening line the model had defaulted to before the brief’s angle kicked in a paragraph later.

Cutting that first paragraph on the next send lifted the click rate on the same list, because the brief’s work had been sitting in paragraph two the whole time instead of paragraph one.

What a Full AI Copywriting Process Looks Like End to End

Six stages, and AI does real work on four of them. The two that stay with you, the brief and the final call on what ships, were always decisions rather than writing tasks.

I go through what that means for pricing your own work in what AI copywriting replaces for a working copywriter, because selling that decision pays differently than selling typing speed.

That’s the whole answer to what AI for copywriting can and can’t do. It drafts. You decide.