How to Use AI to Write Cold Emails That Don't Sound Like AI Wrote Them

Last month I sent 47 cold emails. The first 23 were written with ChatGPT using the obvious approach — paste in context, ask for an email, copy the output. I got two responses, both from people politely asking me to stop emailing them. The next 24 used a completely different method I stumbled into by accident. Nine responses. Three actual calls booked.

Something changed between batch one and batch two, and it wasn't the recipients or the timing. It was how I was using the AI. I'd been doing it backwards the whole time.

The Default Approach Is a Trap

Here's what I did wrong for those first 23 emails. I gave ChatGPT exactly what every tutorial tells you to give it: the recipient's company, their role, what I was selling, what I wanted them to do. Then I asked for a cold email.

The outputs were technically good. Professional tone. Clear value proposition. Proper structure with a hook, body, and call to action. They read like cold emails written by someone who had studied cold emails. Which is exactly the problem.

Real humans don't write emails that sound like they were optimized for conversion. They write emails that sound like a person had a thought and typed it out. The AI was giving me the platonic ideal of a cold email, and recipients could smell it instantly.

I kept tweaking the prompts. "Make it more casual." "Sound more human." "Don't use marketing language." Each revision made it slightly better and still unmistakably synthetic. The AI was trying to imitate humanity rather than actually reasoning like a human would.

What Actually Worked: Backwards Prompting

The breakthrough came when I stopped asking the AI to write the email and started asking it to think about the recipient first.

Instead of "Write a cold email to the VP of Marketing at [company]," I tried: "You're a VP of Marketing at a 200-person B2B SaaS company. It's Tuesday morning, you have 47 unread emails. What makes you immediately delete a cold email? What would make you actually read one?"

ChatGPT gave me a list of instant-delete triggers. Fake personalization. "I noticed your company is growing." Subject lines with [First Name]. Anything that opens with "I hope this email finds you well."

Then I asked what would make it actually stop and read. The answers were specific: something that shows the sender actually knows a real problem I have, not a generic industry problem. Something that doesn't ask for a meeting in the first email. Something short enough that I can process it in the preview pane.

Only after that conversation did I ask for help with the actual email. And I didn't ask for a complete draft. I asked for just the opening line — something that would pass the delete-or-read test we'd just established.

The Kick: Layer Your Drafts Through Different Personas

Here's the thing nobody tells you about AI-written cold emails, and I only figured this out after comparing my successful batch to the failures line by line.

The emails that worked weren't written by one AI conversation. They were written by three.

First conversation: I have ChatGPT roleplay as the recipient and tell me what would make them respond. Second conversation: I have it roleplay as the most cynical spam filter — a person who has seen every cold email trick and is looking for any excuse to delete. I run my draft through that filter and ask what needs to change.

Third conversation: I paste in my near-final draft and ask ChatGPT to identify any phrase that sounds like it came from a template, a marketing playbook, or an AI. Not to rewrite them — just to flag them. Then I rewrite those parts myself, by hand, in my own voice.

That third step is the one that makes the difference. The AI is surprisingly good at identifying its own tells. Phrases like "I wanted to reach out" or "I'd love to connect" or "simplify your process" — it flags all of them when you ask it to look for synthetic language. But it can't replace them with something genuinely human. That part is still on you.

The final emails ended up being maybe 40% AI-assisted and 60% my own words. But the AI did the heavy lifting of understanding the recipient and catching my blind spots.

The Specific Failures Worth Mentioning

Not everything worked even with this method. A few things I tried that made the emails worse:

Asking the AI to add humor. Every attempt at a joke landed awkwardly. Cold emails aren't the place for it anyway, but AI humor in professional contexts is especially painful.

Asking for "creative" subject lines. The AI generated things like "Quick question about [pain point]" and "Not your typical sales email" — subject lines so obviously templated that they might as well say "COLD EMAIL: DO NOT OPEN."

Using the "pretend you're [famous copywriter]" trick. I tried having ChatGPT write in the style of various direct response legends. The outputs were parodies, not imitations. Worse, they were parodies of what the AI thought those writers sounded like based on their most famous lines.

The subject lines that actually got opens were the boring ones I wrote myself. "[Specific thing] at [their company]" or just "quick thought." Nothing clever. Clever is a red flag.

What I'm Still Testing

I've been running a new experiment for the past two weeks where I have Claude write the first draft and then have ChatGPT critique it, and vice versa. The theory is that each AI might catch synthetic patterns the other can't see in its own output. Early results are interesting — the cross-critiqued emails feel slightly more natural — but my sample size is too small to know if it's real or if I'm just noticing what I want to notice.

The uncomfortable truth is that AI can help you write cold emails that don't sound like AI, but only if you're willing to do more work, not less. If you're using it to save time, you're going to get caught. If you're using it to think more carefully about the recipient before you write, that's when it actually adds something.

I still wonder whether the nine responses I got were because of my method, or because those particular people were just more likely to respond regardless. I'll never know for sure. That's the annoying thing about testing anything in sales — you're always working with incomplete data and trying to find patterns that might just be

Heads up: Some links in this post may be affiliate links. I only recommend tools I've personally tested. Opinions are entirely my own.