How to Write a Cold Email That Doesn't Sound Like AI

Recipients have seen enough AI-drafted outreach by now to recognize the pattern on sight: the em dash, the forced compliment, the three-item list that wraps up too neatly. When a cold email trips those signals, it gets skimmed and archived before the offer even registers. Fixing this isn't about banning AI tools, it's about editing what they produce until it reads like one person wrote it to one other person.

What makes a cold email sound like it was written by AI?

A handful of patterns give it away almost instantly, and most of them come from the same root cause: language models default to symmetry and safety. They hedge, they balance, they wrap every paragraph in a bow. Real people don't write that way under time pressure, which is exactly the context a cold email is supposed to simulate.

• Em dash overuse. One or two per email reads fine. Five or six, especially stacked in consecutive sentences, is a stylistic fingerprint most large language models default to when asked to sound "polished."
• Stock openers. "I hope this email finds you well" and its cousins ("I hope you're having a great week") say nothing and cost the reader three seconds they didn't agree to spend.
• Generic flattery. "I noticed your impressive work at [Company]" is a template with a mail-merge field. It could describe any company in the sender's list, which is precisely the tell.
• Triplet lists. Three benefits, three pain points, three reasons to reply. Symmetrical structure signals a model optimizing for "complete-feeling" output rather than a person who actually has one specific reason to write.
• Overly polished, symmetric sentences. Every sentence is roughly the same length, every clause balances the one before it. Humans write unevenly, especially in a first draft they intend to send fast.
• Vague personalization. A line that references the industry, the funding round, or the "growth trajectory" without a fact specific enough to prove the sender actually looked.

None of these are disqualifying on their own. Stack three or four in one email and the reader's pattern-matching kicks in before they've read the ask.

Here's what the difference looks like on the same account. Template version: "I hope this email finds you well. I noticed your impressive work at Acme and wanted to reach out. We help companies like yours streamline operations, boost efficiency, and drive growth." Rewritten version: "Saw Acme dropped the annual plan option in June, only monthly left on the pricing page now. That usually means someone's rethinking churn math upstream. Worth a 15-minute call?" Same recipient, same ask, one of them gets deleted and one gets a reply.

Why do AI-sounding cold emails get lower reply rates?

Recipients increasingly report noticing when outreach was clearly templated by a model, and once they notice, they stop reading for content and start reading for pattern. That shift changes the entire transaction. An email that reads as mass-produced signals the sender didn't spend real attention on this specific account, so why would the recipient spend real attention responding?

This compounds with inbox fatigue that predates AI drafting tools. Reply rate benchmarks have been under pressure for two years running, and volume is part of that story, but so is quality erosion: more senders are shipping first-draft AI copy with zero edit pass. See our cold email reply rate benchmarks for 2026 for where the numbers actually sit by industry.

There isn't a clean, sourced statistic for "X% of recipients can detect AI email" worth repeating here, and inventing one would be its own kind of dishonesty. The qualitative signal is strong enough on its own: spend five minutes in any B2B founder's LinkedIn comments and you'll find people openly mocking the em-dash-and-triplet template. That's a reader base actively training itself to filter this exact pattern.

How do you personalize a cold email so it doesn't feel generic?

Swap the compliment for a fact. "I noticed your impressive work at Acme" personalizes nothing, because it's true of every company on the list. "You shipped a self-serve pricing page in June and pulled your enterprise tier off the homepage" personalizes everything, because it's true of exactly one company, and it proves someone actually looked before hitting send.

The test is simple: could this line describe your top three competitors as easily as it describes the recipient? If yes, it's not personalization, it's a variable. Pull from a specific trigger instead: a recent hire in a relevant role, a product change visible on the site, a review left on G2, a talk given at a named conference. One true, specific detail outperforms three vague ones, every time.

This is a deeper rabbit hole than one section can cover well. Our guide to personalizing cold email at scale walks through how to keep this specific without hand-writing every line, which is the actual constraint teams run into once volume passes a few hundred sends a week.

What sentence patterns make emails sound robotic or overly polished?

Uniform rhythm is the biggest structural tell, bigger than any individual banned phrase. When every sentence runs twelve to eighteen words, subject-verb-object, no fragments, no interruptions, it reads like it was assembled rather than typed by someone thinking in real time.

Fix it by breaking the symmetry on purpose. Follow a longer sentence that sets up context with something short. Just three words, sometimes. Let a sentence start with "And" or "But" occasionally, the way people actually talk. Cut connective tissue like "furthermore" and "it's worth noting" entirely, they exist to smooth a machine's transitions, not a human's.

Punctuation matters too. A model asked to sound professional will reach for the semicolon and the em dash by default. A person dashing off twelve emails before lunch reaches for a period, or just starts a new sentence without one. Read your draft back and count how many sentences are within two words of the same length. If it's more than half, rewrite until it isn't.

Should you avoid using AI to write cold email entirely?

No, and anyone telling you to hand-write every send at real volume is giving advice that doesn't survive contact with a 2,000-send-a-month motion. The mistake isn't using AI to draft. It's sending what it drafted.

Treat a model output the way you'd treat a first draft from a junior copywriter: useful for structure, unusable as final copy. Read it out loud. Cut anything that sounds like it's performing professionalism instead of communicating. Replace every generic line with a specific one, even if that means the email takes five extra minutes per segment instead of zero. The senders getting burned right now aren't the ones using AI, they're the ones shipping its first pass unedited at scale.

Where teams actually get in trouble is volume math. It's fast to generate five hundred drafts and slow to rewrite five hundred personalization lines by hand, so the rewrite step is the first thing that gets skipped under deadline pressure. Build the edit pass into the workflow before the campaign launches, not as a cleanup step after reply rates come in soft. A shorter list with a real edit pass beats a longer list with none.

How do you test whether a cold email sounds human before sending it?

Read it out loud before it goes anywhere. If a line makes you wince, or if you'd never actually say it to the person's face on a call, cut it. This single habit catches more AI-sounding copy than any checklist, because the ear notices rhythm problems the eye skims past.

Run a second pass against a short list: does any sentence pair run the same length back to back? Is there a triplet list doing the heavy lifting instead of one real point? Does the personalization line survive a swap test against a competitor's name? Is there an em dash in every paragraph? Any single "yes" is fixable in under a minute. Two or three "yes" answers in the same email means it needs a real rewrite, not a polish.

Modern Inbound has delivered 6,000+ warm leads.

The habit scales better than it sounds. Once a rewrite pass becomes routine, it adds minutes per email, not hours, and it's the difference between copy that reads as effort and copy that reads as output.

What do people ask about writing cold email that doesn't sound like AI?

Is it okay to use ChatGPT or another AI tool to draft a cold email?

Yes, drafting with AI is fine and most experienced senders do it. The problem isn't the tool, it's sending the first output unedited. Treat the draft as a starting point: rewrite the personalization line with a real fact, break up symmetric sentences, and cut stock phrases before it goes out.

What's the single biggest AI tell in cold email copy?

Generic flattery paired with a triplet list. A line like "I noticed your impressive work at [Company]" followed by three neatly balanced benefits reads as templated because it could apply to almost any recipient on the list, not just this one.

Does removing em dashes fix an AI-sounding email?

It helps but it's not sufficient on its own. Em dash overuse is one signal among several, alongside symmetric sentence length, stock openers, and vague personalization. Fixing punctuation without fixing sentence rhythm and specificity leaves the underlying pattern intact.

How short should a cold email be to avoid sounding AI-generated?

There's no fixed word count that fixes this on its own, but shorter drafts tend to force more specific language because there's no room for filler. Aim for a length where every sentence is doing a job: one hook, one proof point, one ask. If you can cut a sentence and lose nothing, cut it.

If you'd rather have someone else run this discipline on your outbound instead of building the review process yourself, that's the execution layer Modern Inbound handles. See how the setup works, or get in touch to talk through your current sequences.

By Rishabh Ambasta, Founder, Modern Inbound.