How to Avoid Generic, Obviously-AI Cold Emails

A cold email that reads as obviously AI-generated usually isn’t failing because it was written by a model, it’s failing because it’s generic: no specific fact, a handful of tell-tale phrases, and a structure that could apply to any recipient. Fixing those three things fixes the “this feels like AI” problem, regardless of what actually wrote the first draft.

The phrases that give it away

Certain words and phrases have become shorthand for generated, unedited copy. Cut these on sight:

• delve, leverage, seamless, robust, cutting-edge, game-changer, revolutionise, unlock, elevate
• navigate (used figuratively), tapestry, testament, landscape (“today’s landscape”), realm
• foster (as filler), moreover, furthermore
• “it’s worth noting”, “I hope this email finds you well”, “I wanted to reach out”
• “in today’s fast-paced world”, “at the end of the day”, “circle back”, “touch base”
• “synergy”, “holistic”, “empower”, “supercharge”, “take it to the next level”

Also worth avoiding: three-part lists everywhere (“faster, smarter, and more efficient”), and the “not just X, but Y” construction, both of which show up disproportionately in generated copy and read as filler once you notice them.

The structural tell: nothing specific enough to be wrong

The strongest signal of a generic email isn’t any single phrase, it’s the absence of anything that could be factually incorrect. A real, researched detail carries risk (it could be outdated or slightly off), and that risk is exactly what makes it convincing. A generic email plays it safe by saying nothing that specific, which is also what makes it forgettable.

How to fix it, in order

1. Add one true, dated fact about the recipient’s company. See how to research B2B prospects.
2. Cut every phrase from the list above. Read the draft once specifically hunting for them.
3. Shorten every sentence that’s doing two jobs. One idea per sentence reads as human far more reliably than any individual word choice does.
4. Read it out loud.If it doesn’t sound like something you’d actually say to a person, it’ll read that way too.

Why this matters more if you’re using AI tools to help draft

If a tool is doing some of the drafting, the fix isn’t avoiding AI assistance entirely, it’s making sure the research step happens first and the phrase-level edit happens last. A model given a real, specific fact to work from produces a much better first draft than one given only a name and a company. The editing pass is what catches anything generic that slipped through either way.

Phrases to cut on sight

None of these are wrong in isolation. All of them are load-bearing filler: sentences that occupy the position where a specific fact should be.

• “I hope this email finds you well.”
• “I came across your profile / your company and was impressed.”
• “In today’s fast-paced business landscape…”
• “I noticed you’re growing / scaling / expanding.”
• “revolutionise”, “seamless”, “cutting-edge”, “game-changing”, “leverage”.
• “I’d love to pick your brain.”
• “Let me know if this resonates.”
• Any sentence beginning “As a [job title], you probably…”

Why rewriting does not fix it

The common response to a generic draft is to ask for it again with more personality. This produces a generic email with personality, which is not an improvement. It is the same absence of information wearing a better outfit, and it often reads as try-hard on top of empty.

Genericness is an input problem. If the only thing the writer knew about the company was its name and its industry, no instruction about tone can conjure a fact that was never gathered. The fix is upstream: supply one verified, specific, current thing, and the output stops being generic without anyone mentioning tone at all.

Three constraints that do the work

1. Require a sourced fact. No draft begins without one verifiable observation and a link to where it came from. If there is no fact, there is no email. drop the account.
2. Cap the length. Ninety words. Padding is where generic language goes to live, and a hard cap starves it.
3. Ban the abstractions explicitly.An instruction to avoid the list above is far more effective than an instruction to “sound human”, because it is checkable.

Is it dishonest to use AI to write cold email?

No, and the question is slightly misplaced. What would be dishonest is asserting things nobody verified, and a human writing from a template does that just as readily. The obligations are the same whoever types it: the facts are true, a person takes responsibility for what goes out, and the recipient can tell who sent it and opt out.

Which is why a review step is not a nicety. It is where responsibility actually attaches to a named human, and it is the thing that makes the rest defensible. See the structure that survives this.