Why Your AI Emails Sound Like AI, and How to Fix It

People try to fix this by banning words. No more delve, no more leverage, no more I hope this email finds you well. The email still reads as generated, because the vocabulary was never what gave it away. What gives it away is the shape: the way it opens, how evenly it is balanced, how carefully it avoids committing to anything, and the fact that it contains nothing only you could have known.

Word-level tells exist and they move. As soon as enough people complain about a word, models are tuned away from it and the tell disappears, which is why the list of banned words keeps needing updating.

The structural tells do not move, because they are not stylistic choices. They are consequences of how the text is produced. You can fix the words in a minute and the email will still be recognisably generated.

Any one of these is forgivable. Humans write mirror openings on a bad day. All five in one message is unmistakable, and readers do not need to consciously identify it. They just file it as bulk.

Understanding the cause makes the fix obvious, and it is not a defect being patched.

A model produces text near the centre of everything similar it has seen. Asked for a professional outreach email, it produces the average professional outreach email, and the average is balanced, hedged, polite and non-specific, because averaging over thousands of emails smooths away everything idiosyncratic.

Which means every fix is the same fix: give it something specific enough that averaging cannot produce it.

People spend enormous effort on prompt wording and almost none on examples, and the ratio should be reversed.

Telling a model to write conversationally moves it toward the average conversational email, which is a different uniform register rather than yours. Giving it twenty emails you actually sent gives it a distribution it has never seen, which is the only thing that can pull it off centre.

Pick emails you were pleased with, across different situations, and include the short ones. The short ones carry the most information about how you actually write, because they show what you leave out.

One rule underneath all of this: every message needs at least one thing in it that could only be in that message.

Not personalisation tokens. A first name and a company name are available to everybody and read as a mail merge, because they are one. A specific means something that required somebody or something to actually look: a line from their careers page, the thing they posted about last week, the fact that they are hiring for a role that implies the problem you solve.

This is where an agent genuinely helps rather than hurts. Finding one real specific per prospect across two hundred prospects is exactly the work a person cannot sustain and a system can. The failure is not using AI for outreach; it is using it to generate the wrapper and skipping the specific, which produces two hundred well-written emails that are obviously about nobody.

The other half of the rule: cut the first sentence. Almost always the mirror opening, almost always removable, and the email is better without it.

One, feed it your writing, not adjectives about your writing. Twenty real emails beats any description of your tone.

Two, require a specific and let it fail. If nothing real can be found about a prospect, that is information. Either they are not worth the email or the list is wrong. Filling the gap with a generic compliment is how the whole batch becomes obvious.

Three, ask for shorter than feels right. Models pad toward the average length, which is longer than a good email. Asking for three sentences produces something closer to what a busy person would actually write.

Four, read every one before it sends. Not for quality control alone, but because your edits are the training signal. This is the part the approval gate exists for, and the reason it makes the output better rather than just safer.

Five, cut the first sentence. Mechanically, every time. It is right often enough to be a rule.

The wider question of what AI should and should not be doing in your marketing is in what AI still cannot do, and the deeper version of feeding it your own material is in training an AI on your brand voice.