The advice you normally get here is a list of banned words. Stop saying "streamline". Stop saying "I hope this email finds you well". That list is real, and following it produces a machine-written email with a slightly better vocabulary.
The tells are structural. A generated draft is symmetrical, risk-free, flattering, and written from the sender's side of the table. Real messages from real people are lopsided, specific, slightly blunt, and occasionally end in a way that is grammatically ugly.
The machine-written drafts that land in my inbox give themselves away in the first line, before anyone gets anywhere near the pitch.
So the fix is not a thesaurus. Below is the shape these drafts arrive in, assembled from the lines that turn up in nearly every one of them, then taken apart, then rewritten. What earns a reply once the message finally reads like a person wrote it is the larger question, and that one is answered in what actually gets a reply.
The draft, as it always arrives
Subject: Quick question about scaling your operations
Hi Martin,
I hope this email finds you well. I came across Nordwerk Systems and was really impressed by your work in the industrial automation space.
At Acme, we help companies like yours streamline their operations, reduce costs, and drive sustainable growth through our innovative end-to-end solutions. Our clients typically see significant improvements in efficiency within the first few months.
I thought it might potentially be a good fit and would love to hop on a quick 15-minute call to explore how we could support your goals.
Would you be open to a brief chat sometime next week?
Best regards, Tom
Nothing in that email is wrong. That is the problem with it. It is correct, complete, polite and completely synthetic, and a director who reads this shape every morning can feel it before the second line.
Line by line
"Quick question about scaling your operations." There is no question in the email. The subject promises one thing and the body does another, which is the first small withdrawal from an account you have not opened yet.
"I hope this email finds you well." Six words that carry no information and confirm you did not write this for Martin.
"I came across Nordwerk Systems and was really impressed by your work in the industrial automation space." This is the sentence generated drafts cannot resist. It is flattery with a variable in it. Swap the company name and it works for any company on the list, which is exactly what the reader suspects. The word "space" is doing nothing. "Really impressed" is doing less.
"we help companies like yours streamline their operations, reduce costs, and drive sustainable growth" Three benefits in parallel clauses, ordered from shortest to longest. That rhythm is the single clearest fingerprint of a generated sentence. No person writing at speed produces a balanced triple.
"through our innovative end-to-end solutions" Nobody has ever bought a solution. Also, this is a sentence about the sender, sitting in the position where the reader's problem should be.
"Our clients typically see significant improvements in efficiency" A claim with every checkable element removed. How much, over what period, for whom, compared to what. If the reader cannot verify it or argue with it, it is decoration.
"I thought it might potentially be a good fit" Two hedges stacked in one clause. The writer has pre-apologised for the email, and the reader now knows nobody is confident about anything.
"would love to hop on a quick 15-minute call to explore how we could support your goals" "Quick" and "15-minute" say the same thing twice. "Explore" and "support your goals" mean nothing. The ask is large and vague, which is the worst combination available.
The rewrite
Subject: your second shift
Hi Martin,
You run three production lines in Brno and, from your job ads, you are hiring a second shift supervisor for the third time this year.
We work with mid-size manufacturers in Czechia and Slovakia on exactly that handover problem. [One real result from your own work goes here: what changed, by how much, over how long.]
Worth fifteen minutes on Thursday to see whether the same approach fits your lines? If it is not relevant, tell me and I will leave you alone.
Tom
Shorter, ruder, and written for one person. It contains something the reader can disagree with, which is the whole point. A generated draft cannot be argued with, because it never says anything.
The bracket in the middle is deliberate. I am not going to invent a manufacturing result to make an example look good, and neither should your email. That slot takes one real outcome from your own work, with a number, a scope and a timeframe. If you do not have one yet, the sentence comes out and the email gets shorter.
What actually changed
The first line is about them and checkable. It comes from their own job ads, it took ninety seconds to find, and it could be wrong. Being wrong in public is a risk no template takes.
The claim has edges, or it is not there. A number, a comparison, a scope, all of it yours. Borrowing somebody else's result is the one shortcut that ends the conversation permanently, and a bracket in the draft is better than a sentence you cannot defend on the call.
The rhythm is uneven. One long sentence, one short one, one fragment that a grammar checker would flag. Read your draft aloud. Where it sounds like a school report, break it.
The ask is small, dated and reversible. Fifteen minutes on Thursday, with an explicit way to say no. The exit line is doing work rather than being polite. You have made saying no cost nothing, and a fast no is worth more to you than a slow silence.
Nothing describes the sender's company. "At Acme, we help companies like yours" has been deleted rather than improved.
The numbers that circulate about personalization almost all come from companies selling personalization software, so there is none of that here. You do not need a statistic for this one anyway. Nobody has ever answered an email because it contained their first name and their company name. Believing otherwise is what produced the draft at the top of this post.
Where the machine actually belongs
None of this is an argument against using AI in outbound. It is an argument about which half. Automate what the prospect never sees, from sourcing and list checks to testing one angle against another, and keep a person on the words. The test I use is simple: would the prospect ever see it.
The one use of a model I actually recommend for copy is subtraction. Paste your draft in and ask which sentences the reader could not check. Then delete those yourself. The method for that pass, and what usually survives it, is in what I cut from cold emails.
Run one test before any batch goes out. Take your draft and imagine handing it to the recipient at a conference next month, with your name on it. If you would want to explain the first line before they read it, rewrite the first line.
