How to stop your cold emails sounding like AI

Cold emails read as AI because of sentence SHAPE, not word choice. Swapping "revolutionize" for "improve" changes nothing while the structure stays the same: a meta-observation about having looked at the account, a compliment shaped around the recipient's size, a sympathy line about how hard their position is, and a closing question that names the topic back. A reader clocks that pattern in about two seconds, and no amount of merge-field personalization survives it.

We know this because we built a machine that writes first messages and then spent months deleting what it produced. The file that came out of that, our humanization protocol, is a list of constructions that are banned outright, each with the date it was banned and the draft that earned it. What follows is the useful part of it.

The three tells, in the order they give you away

1. The meta-observation

The message opens by narrating the research instead of stating the finding. Every variant of this is dead on arrival, because no human writes this way about something they genuinely noticed.

I took a look at your account and one thing that stood out was the engagement.

your posts reach about 3% of your followers.

State the observation as a plain fact. The reader does not need to be told that looking happened, and saying so is how they learn a tool did the looking.

Banned in our protocol as a family, not as a phrase: "one thing I noticed", "stood out", "worth flagging", "caught my eye". Ban one and a generator reaches for the next. You have to ban the move.

2. The stature compliment

The second tell is flattery built out of the recipient's own bio. It reads as backhanded, because it is: the sentence works by implying they should be doing better than they are.

Pretty quiet for a project this size, especially with the backing you have.

you have got the audience already, the replies just are not coming with it, so there is real room to grow here.

Same observation, framed as upside instead of as a shortfall. One makes a stranger defensive, the other makes them curious.

Our banned list carries seven variants of this construction, added on one day in July after a review of a sent batch: "for a project this size", "for a project this big", "with this kind of backing", "for a team this", "given the backing". They all came from the same generator reaching for the same move through different words.

3. The sympathy line

A standalone sentence of diagnosis or empathy placed between the observation and the pitch. It is meant to soften. It reads as a label being applied to them.

That kind of gap usually means the social side is not getting attention while the team is heads-down on the product. Makes sense though.

(nothing. go straight to what you would do about it)

Observation, then pitch, then ask. A first message that stops to explain the recipient's own company to them has spent its one paragraph of attention on nothing.

Why personalization is not the fix

The standard advice is to personalize harder: mention their funding round, their recent post, their new hire. That advice is not wrong, it is just aimed at the wrong layer. A merge field inserted into a template built out of the three moves above produces a message that is specific AND obviously generated, which is worse than a short generic note, because it proves a machine did it at scale.

The thing that actually changes the read is whether the message contains something the sender could only know by looking, stated the way a person states a fact. "Your last six posts averaged four replies" is not personalization. It is evidence.

The shape that works

Three beats, and nothing between them:

One specific observation. Something visible from outside, stated flat. One number at most in the whole message, and only if it carries weight.
One pain, named plainly. Not their product, their distribution. Never recite their features back at them.
One light ask. Short and generic. The moment the ask names the topic back ("worth a quick chat about what is holding the reach down?") it sounds like a subject line.

A few smaller rules that came out of the same review and cost nothing to apply:

Do not quote follower counts. Describe the audience without sizing it.
Raw counts beat percentages. "two to five replies a post" lands; "0.8% engagement rate" reads as a dashboard export.
Never cite a likes number above ten as a problem. The founder knows twelve likes is not a disaster, and saying it is costs you the whole message.
No em dashes. Ours is rule one, and it is the single cheapest tell to remove.

How to check your own

Take a message you sent last week and mark every sentence that is not doing one of the three jobs above. Openers that describe your research, compliments that restate their bio, and lines that explain their situation back to them are all deletions. What is left is usually about forty percent of the original, and it is the part that would have been answered.

Then read the remainder out loud. Anything you would not say to a person standing in front of you goes too.

Questions people actually ask

Can AI write cold emails that do not sound like AI?

Yes, but not by being told to sound human. It works when the model is constrained to state verified observations in a fixed shape and is blocked from the constructions that give it away. The intelligence goes into the rules, not into the prompt.

What is the single biggest giveaway?

The meta-observation opener, some version of 'I looked at your account and noticed'. It is the most common and the easiest to remove, because the fix is to delete the clause and keep the fact.

Do spam filters detect AI-written email?

Not directly, and that is the wrong thing to optimise for. Deliverability is decided by sending infrastructure and recipient engagement. What AI-shaped writing costs you is the reply, from a human who already read it.

Is a longer message more personal?

No. Length reads as effort to the sender and as work to the reader. Every rule above removes sentences rather than adding them.

Try Coldpitch and skip the manual part

It finds the accounts worth writing to, spots the weak point, digs out who to talk to on Telegram or X, and writes the first message from what it actually found. Less work than doing this by hand, and better leads at the end of it.

Get my first messages