← Back to Blog
The 20-Minute De-AI Pass: What to Do When a Client Says Your Deliverable Sounds Like a Robot (Without Rewriting It From Scratch)

The 20-Minute De-AI Pass: What to Do When a Client Says Your Deliverable Sounds Like a Robot (Without Rewriting It From Scratch)

November 14, 2026·6 min read

A designer I know sent a proposal last month. The client replied with one line: "This reads like ChatGPT wrote it." Nothing specific. No notes on the numbers or the scope. Just that sentence, and a week of trust gone.

What's becoming clearer to me is that this is almost never a writing problem. It's a specificity problem wearing a writing costume.

The tell is not vocabulary, it's symmetry

Everybody assumes AI-sounding writing means big words. It doesn't. Humans reach for fancy words all the time, and it's annoying, but it still sounds like a person. What machines do that people don't is pull every sentence with the same weight.

Look at a paragraph and count words per sentence. Human writing swings. A 4-word sentence sits next to a 38-word one because the thought actually swung that way. Machine drafts average out to 15 or 20 and stay there.

Then there's the balance problem. Machine drafts give both sides. They hedge. They list three items where the third is filler. They write conclusions that restate the intro because that's what the pattern calls for. Nothing is allowed to matter more than anything else.

It's not that the writing is bad. It's that it's evenly good. Evenness is the fingerprint, and you can usually fix it in 20 minutes without touching the structure you already like.

The 20-minute pass, minute by minute

Do this on the document, not in your head.

Minutes 0 to 4: read it out loud. Use Word's Read Aloud, the Google Docs Read Aloud extension, or just your voice in an empty room. Mark every spot where you stumble. A stumble is almost always a seam where two writers, you and the model, got stitched together. Don't fix anything yet. Just mark.

Minutes 4 to 9: break the symmetry. Pick your three flattest paragraphs and cut one of them down to two sentences. Then go find every list of three and check item three. If item three is "and it saves time," delete it. A three-item list where the third is filler is the loudest AI tell on the internet right now, and it takes eight seconds to kill.

Minutes 9 to 14: add one fact only you could know. This is the highest-yield block in the whole pass. A date, a name, a number, something they said on the call. "Your ops lead is out until March, so we're staging this in two phases" does more work than any amount of style editing. Clients don't hear AI in specifics. They hear AI in the absence of them.

Minutes 14 to 18: run the voice pass. Paste your draft into Claude or ChatGPT with two or three real samples of your own writing and one instruction: flag sentences that don't match this voice, do not rewrite them. You want the flags, not the rewrite. Let a model rewrite and you get a different machine voice, which is worse because now it's confident. Hemingway App helps too, but ignore the grade level score and look at sentence length variance instead.

Minutes 18 to 20: hand it to someone who isn't you. A teammate, an assistant, a friend who reads your emails. One question: "Where did you get bored?" The bored spot is where the machine voice lives.

That's the pass. It costs one coffee and no rewrite.

Build the voice file once so you stop doing this every time

The 20-minute pass is a last line of defense. It shouldn't be your strategy.

The agencies I've watched get good at this have one shared document doing most of the work. Mine has four parts. Ten samples of writing I actually like, pulled from real emails and Slack messages, not from my website, because website copy is already a costume. A banned-words list I keep adding to. Three paragraphs on how I talk to clients when something goes wrong. And a short set of hard rules, like "never open with 'In today's fast-paced world'" and "no lists of three where the third item is filler."

Put that document where the model can see it. Claude Projects and ChatGPT's custom instructions both accept it and hold onto it across a whole project. From there your AI brand voice editing workflow is simple: every prompt for client work starts with that file attached. The first draft arrives most of the way there and your 20-minute pass turns into five.

The agencies that struggle have 12 people prompting 12 different ways with no shared standard for what "sounds like us" means. If you're weighing whether to build that in-house or have it built for you, the Mastermind vs page lays out the honest tradeoffs.

The 7-item checklist that runs before anything reaches a client

Every deliverable gets this pass before it leaves. It takes 90 seconds, and it is the whole of your AI output review process for consultants who don't want a process.

  1. Is there a specific number, name, or date on page one?
  2. Could a competitor send this same document to a different client without changing more than a logo? If yes, it's generic.
  3. Read the first sentence of every section in a row. Do they sound like one person?
  4. Does any sentence start with "In today's" or contain "it's important to note"?
  5. Is there a list of three with a filler third?
  6. Do the shortest and longest sentences in the document differ by at least 20 words?
  7. Does the ask at the end sound like the person who will actually do the work?

That last one matters more than people expect. AI writing gets vague at the moment of commitment, because commitment carries risk. Real people write more precisely when they're asking for something. We build versions of this checklist live in the workshop, using your own documents instead of sample ones.

When the client says it, don't defend

The worst reply is a paragraph explaining your process. They don't want a defense. They want to trust the next thing you send.

Say this: "Fair. Which parts sounded off to you?" Then listen. Half the time they point at something you wrote without a model anywhere near it, and now you know how their ear is calibrated. Ask what the last piece of work from a peer felt like, and write down their words, not yours.

Then fix it and send it back inside a day with three lines on what changed. Fast and specific beats perfect and defended.

I've trained more than 90,000 people since 1998, and the pattern never moves. The founders who keep clients are not the ones with the best AI tools. They're the ones who treat the last 20 minutes before sending as part of the job.

If this resonates and you want to build this kind of infrastructure in your own business, the Mastermind is where we do the work live. You can learn more at mastermindshq.business.

Ready to put this into practice?

Join Joe Che's AI Business Mastermind, a small cohort for founder-led service businesses that want to systematize their operations with AI.