Writing Cold Outreach Emails With ChatGPT That Don't Sound Like a Bot

You can usually spot an AI-written cold email in the first five words. "I hope this email finds you well." "I wanted to reach out because." A subject line that's a complete, grammatically tidy sentence. Three tidy paragraphs, each one a little too balanced, ending with "Would love to connect!" and an exclamation point that nobody actually feels. Prospects have seen thousands of these, and the pattern-match happens before they've even read what the email is offering.

The fix isn't avoiding ChatGPT for cold outreach. It's giving it the inputs that make the difference between a template and a specific message, then editing out the handful of phrases that give it away. If you haven't used ChatGPT for sales writing before, the Complete Beginner's Guide to ChatGPT covers the basics of getting a useful first draft out of it.

What actually gives a cold email away

It's rarely one obvious tell. It's an accumulation of small habits that, individually, look like decent writing, and together, read as generic.

Overly balanced structure: an opener, exactly three benefits in parallel phrasing, a closer. Real cold emails from a real person are lumpier than that. Vague personalization that could apply to five hundred companies: "I saw your company is doing great things in the [industry] space" is not personalization, it's a fill-in-the-blank that happens to have the blank filled in. Enthusiasm with nothing behind it: "excited," "thrilled," "love to," used at a volume no one uses in a first message to a stranger. And the closer that asks for too much too fast: "Let's schedule a 30-minute call this week to discuss synergies" from someone who has given the recipient zero reason yet to spend 30 minutes on anything.

Mechanically, this happens because those patterns are the statistical average of how cold emails get written. A model asked for a "cold email to a VP of Sales" with no other input has nothing to draw on except the most common shape that phrase takes across every cold email template it's ever seen, and the most common shape is exactly the generic one: balanced paragraphs, safe enthusiasm, a big ask at the end. That's not a flaw specific to ChatGPT, it's what happens whenever a prompt gives a model a category (cold email, VP of Sales, CRM tool) instead of a specific situation. The fix isn't a trick phrase, it's replacing the category with a real, unrepeatable detail the average template has no way of guessing.

Watching the same email get less generic, stage by stage

Take one scenario through three attempts, so the difference isn't described, it's visible.

Bad prompt: "Write a cold email to a VP of Sales pitching our CRM tool."

That produces something close to: "Subject: Boost Your Sales Team's Performance. Hi there, I hope this email finds you well! I wanted to reach out because I noticed your company is doing great things in the sales space. Our CRM platform helps teams close more deals with less busywork, thanks to robust automation, real-time reporting, and easy integrations with the tools you already use. Would love to schedule a 30-minute call this week to discuss how we could help. Looking forward to connecting!" Nothing in that email would need to change if you swapped the recipient's name and industry. That's the tell.

Better prompt: "Write a cold email to Marcus, VP of Sales at a 40-person SaaS company that just closed a $12M Series A. Pitch our CRM tool, which helps sales teams track deals without manual data entry."

That's closer: "Subject: Congrats on the Series A! Hi Marcus, congratulations on closing your $12M Series A, that's a huge milestone! As you scale your sales team, keeping deal data accurate can get messy fast. Our CRM automatically logs calls and emails so your reps spend less time on data entry and more time selling. Would you be open to a quick call this week to see if it's a fit? Excited to connect!" There's a real, specific hook now (the funding round), which is genuine progress. But it still asks for a call in the first message, still ends on forced enthusiasm, and the exclamation points are doing the same performed-excitement work as the bad version.

Excellent prompt: everything from the better version, plus explicit constraints on tone and the ask.

The only new information between the better and excellent prompts is a set of constraints (length, no exclamation points, a smaller ask), and the output changes from "acceptable but pushy" to something that reads like a specific person wrote it to Marcus and no one else.

The same two levers, a real hook and explicit constraints, are what separate the two worked examples below from a generic template, in two different situations.

Worked example one: a founder pitching a niche tool to an agency owner

Say you run a small QA testing tool for e-commerce sites, and you want to reach the owner of a Shopify design agency, whose team you noticed just shipped a redesign for a mid-sized skincare brand.

A reasonable output from a prompt like that reads close to: "Subject: quick QA question for Glowstate launch. Hi Priya, the Glowstate redesign your team shipped looks great, especially the mobile checkout flow. Curious how you handle QA on launches like that, since it's usually the first thing that gets rushed when a deadline is tight, and the first thing a client notices if something breaks post-launch. We built a tool that catches exactly that (broken checkout paths, mobile bugs) before a site goes live. Not trying to book a call, just curious if pre-launch QA is a real headache for your team or something you've already got handled. Two lines either way would be genuinely useful to hear."

Notice what makes this different from the generic version: the specific client name, a specific observation about the redesign, and a low-effort ask instead of a calendar invite.

Worked example two: a salesperson following up after a conference

Different scenario: a sales rep met someone briefly at a conference booth, got a business card, and wants to follow up without pretending the conversation was longer or deeper than it was.

A good draft here should openly reference the brevity of the real interaction rather than inflating it, something like: "Subject: the 3-location syncing thing. Hi Daniel, we only got a couple minutes at the booth last week, but the inventory syncing issue across your three warehouses stuck with me, that's almost exactly the problem we built our tool around. Happy to send over a short one-pager on how we handle it if that'd be useful, no pressure either way. If it's not a priority right now, no worries at all."

Editing the draft before you send it

Even a well-prompted draft usually needs one more pass. Read it out loud once. If a sentence doesn't sound like something you'd actually say to a person, cut it or rewrite it by hand, even just changing a couple of words. The goal isn't a perfect AI output, it's a draft that gets you most of the way there fast, followed by five minutes of making it sound like you specifically wrote it.