AI cold email: the tells that get it deleted

I asked a model to write a cold email for me and it came back in nine seconds. I read it twice and could not find anything wrong with it. Correct grammar, sensible structure, polite close, no typos anywhere.

It was also the most ignorable thing I have ever put my name on, and it took me a while to work out why, because "wrong" is not the problem with it.

Same stand-in as the other posts here, because the message is the point and she is not: call her Dana, Head of Content at a 40-person SaaS, publishes twice a week, none of it ranking. I am selling her an AI tool that writes thirty SEO articles a month.

Hi Dana,

I hope this email finds you well. I came across your work at [Company] and was really impressed by the content you're putting out.

I noticed that many Heads of Content are struggling not just with volume, but with consistency. That's where we come in. Our AI-powered platform helps content teams streamline their workflow, boost organic traffic, and free up time for the strategic work that matters most.

Would you be open to a quick 15-minute chat next week to explore how we might be able to help? Happy to work around your schedule.

Best, Alex

Nothing in there is a mistake. That is the whole problem.

Why AI cold email stopped working

Every benchmark report published this year says the same direction: reply rates are down, and they have been falling for years. They disagree about the number by a factor of three or more depending on who ran the study and what they were selling at the bottom of the page, so I am not going to hang an argument on one of them. Pick the number you like and the shape is the same.

The cause is not mysterious. Writing a cold email used to cost something. Not much, but enough that a person could only send so many before they got bored, and boredom was doing a quiet job of rationing the inbox. That cost is now approximately zero. One person with a list and an API key can produce more outreach in an afternoon than an entire SDR team could in a quarter.

What happens next is arithmetic, not psychology. A senior buyer at a company worth pitching now gets dozens of these a week, all grammatical, all polite, all structurally identical, because they were all produced by the same handful of models answering the same handful of prompts. Dana does not need to think about any of this. She has just been trained, by volume, to recognise the shape of a message that was cheap to send.

And here is the part most posts on this get wrong: she is not detecting AI. She cannot run a detector on her phone at 8:40am and would not bother if she could. She is detecting that nobody looked at her before they sent it. Those two things correlate right now, which is why "sounds like AI" has become shorthand for it, but the thing being punished is genericness. A human writing from a template gets deleted for exactly the same reason and always did.

That distinction matters because it tells you what to fix. Rewriting a generic message in a chattier voice does not fix a generic message. It just makes it a generic message with contractions in it.

The six tells a buyer clocks before your first line ends

None of these are read consciously. Dana is not annotating your email. She is doing what all of us do with a full inbox, which is deciding in about two seconds whether this is a message or a broadcast, and these are the things that decide it for her.

1. It opens with a compliment that cost nothing

"I came across your work and was really impressed."

Impressed by what? A compliment that does not name the thing it is complimenting is not a compliment, it is a greeting wearing one. Dana has received this exact sentence enough times to know it survives find-and-replace, which means it was not written about her.

The same test from the cold DM post applies here: if you can put the next name on the list into the sentence and send it unchanged, it says nothing about the current one.

2. The personalization is a variable, not a fact

"[Company]" is not personalization. Neither is her job title, her city, or the fact that she works in content.

Those are merge fields. They prove you had a spreadsheet, which every sender has. Real personalization is a fact you had to go and find: a specific post she published, the keyword she is losing, the competitor sitting above her on it. Something that would be wrong if you sent it to anyone else on the list.

The gap between "personalized" and "personalized in a way that costs the sender something" is the entire game now, because the first kind became free at the same moment it stopped working.

3. The rhythm gives it away before the words do

"Struggling not just with volume, but with consistency."

"Streamline their workflow, boost organic traffic, and free up time."

Not just X, but Y. Three items in a list where two would do, each one roughly the same length. Every sentence landing on a similar beat. Models write in this rhythm because it is the average of everything ever written, and averages are smooth in a way people are not.

Read your own messages to colleagues sometime. They start with "so" and "actually," they run short then long, they have a fragment in the middle. Nobody talks in balanced triads. When a stranger's email does, it registers as a document rather than a message, and documents get filed, not answered.

4. Words nobody says out loud

Streamline. Leverage. Robust. Seamless. Unlock. Empower. That's where we come in. The strategic work that matters most.

Say any of those to a friend and watch their face. This vocabulary exists almost exclusively in marketing copy, which means its presence tells Dana she is reading marketing copy, which means she already knows how the rest of it goes.

There is a faster version of this check. Read the draft out loud. Every phrase you would be embarrassed to say to an actual human at an actual table is a phrase that came from the average, not from you.

5. The message is longer than the ask deserves

The draft above is roughly ninety words to ask for fifteen minutes.

Length is a promise about value. A long message says "there is something in here worth your time," and when the payload turns out to be a calendar link, the promise was false and Dana knows it by the second paragraph. She does not finish it. Nobody finishes a cold email that has already told them what it wants.

Models default to long because prompts default to "write a cold email" and length reads as effort to a model in a way it does not to a person. It is one of the few tells you can fix with the prompt alone, which is exactly why it is the least important one on this list.

6. Nothing in it is wrong

This is the strange one, and I want to be careful with it, because I have written the opposite before.

In the cold DM post a missing letter in "it save you time" cost me, because when you sell writing, the message is a work sample. That is still true. This is not an argument for typos, and if you take it as one you have made your email worse for no reason.

The tell is not correctness. It is that correctness is the only quality the message has. Nothing in it is risky, nothing is specific enough to be wrong, no claim is made that could be checked and fail. A message with no surface for disagreement is a message with no content, and flawlessness in the absence of content is itself information about where it came from.

A real message from a real person contains at least one thing they could be wrong about. That is what makes it worth reading.

How to make an AI cold email sound human

You do not do it by asking the model to sound human. I have tried that prompt, and every variation of it, and what comes back is the same message with "Hey" at the front and a contraction where the comma was. The rhythm survives. The genericness survives, because genericness was never a style problem.

You fix it by putting something in the message that could only have come from looking. Here is the same pitch after I did that:

Dana, your last four posts are all how-to pieces and all four sit on page two. The one competitor above you on every single one of them publishes half as often as you do. I write thirty SEO articles a month and I would rather be judged on one than described: pick the keyword you keep losing and I will send you that article, no call, so you can put it next to your own. Yes, and it is in your inbox Friday.

Four sentences, about eighty words, so barely shorter than the draft. Length was never the issue.

Line one is a fact about her that took me four minutes and a search to find. She can check it faster than she can delete it, and checking is a form of engagement. It also carries an implicit diagnosis without stating one, which is more respectful than "many Heads of Content are struggling."

Line two is the observation underneath the fact. Anyone can report a ranking. Noticing that the person beating her does it on half the output is the part that suggests I understand the problem rather than the metric. That is the sentence a model will not write for you, because it requires having looked at two things and connected them.

Line three replaces the pitch with an offer. "Our platform helps teams streamline their workflow" asks Dana to imagine the product. "Pick the keyword and I will send the article" hands her the product. I have no logos and no case studies, so the work itself is the only proof I actually have, and it is the honest version of that sentence.

Line four names the cost of saying yes and the date she gets it. Not "would you be open to," which is a question about her openness, a thing nobody has an answer to. One word, one deadline, no calendar link.

Where AI actually belongs in cold outreach

I am not telling you to stop using models. I use one every day and I built a product on top of one. The distinction is which part of the job you hand over.

Research is the obvious one and the most underused. Finding what Dana published, what she ranks for, who sits above her, what her competitor is doing differently, that is exactly the sort of tedious lookup work worth automating, and it is the input that makes line one of the rewrite possible. Most people use AI for the last mile, the sentence, and do the research themselves or not at all. That is backwards. The research is the expensive part and the part that shows.

List qualification is the other one. Deciding whether a company on your list is actually a fit, before you write anything, saves more replies than any subject line. Most bad outreach is not badly written, it is correctly written to the wrong person.

Pressure-testing your draft is the third. Handing a model your finished message and asking it what a busy buyer would object to is a genuinely different task from asking it to write the message. One returns criticism you have to act on, the other returns something you can paste.

What I would keep for yourself is the sentence. Not for craft reasons, for evidence reasons: the sentence is the only place the reader can see that you did any of the work above. If you outsource that, none of the research reaches her.

Should you use AI to write cold emails at all?

The honest answer is that it depends on which failure you are more afraid of.

If the alternative to an AI-written email is no email, send the AI-written email. A generic message that gets ignored is a worse outcome than a specific one and a better outcome than silence, and plenty of founders are stuck at zero because they are waiting to feel ready. Something is better than nothing, and you learn more from a message that got ignored than from one you never sent.

But do not confuse it with a strategy, and do not scale it. The volume approach is available to everyone at the same near-zero price, which means it has already been arbitraged away. You are not early to it. You are the four hundredth person this month to send Dana a well-structured email about unlocking content efficiencies, and the four hundredth copy of a thing is worth less than the first, no matter how well written it is.

The version I would actually defend: use the model for everything except the words the buyer reads, then write forty of those a week instead of four hundred. Forty is a number one person can genuinely research. Four hundred is a number that requires you to stop looking, and the moment you stop looking, all six tells come back regardless of who typed the sentence.

The inbox you are actually landing in

One thing worth knowing before you blame your copy. The flood is not evenly distributed. Senior titles absorb most of it, because every tool with a filter points at the same job titles, so the VP and the C-level inbox is where outreach goes to die. One or two levels down, the same message arrives in a much quieter room.

That is a targeting fix, not a writing fix, and no rewrite will save you from getting it wrong. If your reply rate is zero and your message is genuinely specific, the problem may not be the message at all. It may be that you are competing with sixty other senders for the same person while the person who actually owns the problem gets almost nothing.

Worth checking before you rewrite the email for the fifth time.

Five checks before you send an AI-assisted cold email

Run these on the draft, in this order:

1. Delete every sentence that would still be true if you sent it to the next name on your list. What is left is the message.
2. Is there one fact in it you had to go and find? Not a merge field, a finding.
3. Read it out loud. Does any phrase in it sound like something you would be embarrassed to say to a person at a table?
4. Is the message longer than the ask justifies? Ninety words to request a call is a broken trade.
5. Is there one thing in it you could be wrong about? If nothing in the message is checkable, nothing in it is worth checking.

Check one does most of the work. On the draft at the top of this post it removes everything except my product's name, which is a fair description of what Dana received.

And if none of it gets a reply, the follow-up is its own trap with its own wrong answer: what to say instead of "just checking in". If it does get a reply and the reply is three words long, that has a wrong answer too, and it is the one a model will hand you fastest.

The habit underneath

The reason a model cannot write this message for you is not that it writes badly. It writes better sentences than I do, faster, and it never leaves a letter out of "it saves you time."

It is that the message is not really made of sentences. It is made of the twenty minutes you spent looking at Dana before you wrote anything, and the sentence is just where that becomes visible. Ask a model to produce the visible part without the twenty minutes and you get a shape with nothing inside it, which is precisely what the six tells are detecting.

So the skill worth building is not prompting. It is noticing, in your own draft, the exact line where you stopped writing about her and started reciting your pitch. I have been doing this for a while and I still miss it, because by the time I read the draft back I already know what I meant, and my brain fills in the specificity that never made it onto the page.

That gap closes with reps, not with a better prompt. Write the draft yourself, then get it read back to you before Dana reads it.

That is what I built Sell a Pen to do: it marks the lines that are carrying nothing, names the habit behind them, and then you write the line again yourself, because the rewrite is the part that sticks. It will not hand you a better sentence, which is the whole point. Start free on a product I made up.