How to write a cold email that doesn't sound like AI wrote it

Open your inbox. Filter to the cold pitches. Read three of them.

You can spot the AI in under five seconds. The tells are everywhere. The em dash where a comma would do. The "I hope this email finds you well" opener nobody has used sincerely since 2008. The triple parallel pitch ("we help you streamline outbound, accelerate pipeline, and unlock revenue at scale"). The compliment that reads like it was scraped from LinkedIn ("impressive growth at Acme"). The request for "15 minutes for a quick chat."

Every single one of those patterns is a fingerprint. Not because GPT chose them, but because the people prompting GPT think this is what professional sounds like. It is not. It is what professional sounded like in a 2014 email training course that got fed into a million blog posts that got fed into a model.

Buyers learned to filter for these patterns before the AI wave even started. Now that AI has cranked the volume, the filter is faster and harsher. If your email opens with one of the tells, it gets archived in the time it takes to scroll past it.

Here is what to do instead.

The tells

The fastest fix is recognizing what to delete. These are the patterns that mark an email as AI-generated, even when a human wrote it.

The opener apology. "I hope this finds you well." "Sorry to interrupt." "Hope you're having a great week." These signal that the sender has nothing specific to say and is buying time. Cut them entirely.

The triple parallel pitch. "We help you do X, Y, and Z." Real value props are not three things. They are usually one specific thing. If you are listing three benefits, you do not know which one matters to this prospect.

The em dash. GPT pours these into every sentence because the training data did. A comma, a period, or a colon almost always works better. If you find yourself reaching for an em dash in a cold email, the sentence is probably too long anyway.

The vague compliment. "Loved your recent post." "Impressive growth." "Big fan of what you're building." If the prospect cannot tell whether you actually read the post or just saw it existed, the compliment costs more credibility than it earns.

The 15-minute ask. Nobody believes the meeting will be 15 minutes. The phrase has been so overused it now reads as "I am about to take 30 minutes of your time."

The closing flourish. "Looking forward to your thoughts." "Excited to connect." "Best regards" with a 12-line signature block. None of this adds anything. A name on its own is fine.

Real before and after

Here is a cold email written the AI way. This is roughly what you get if you prompt GPT with "write a cold email to a VP of Sales at a Series B SaaS company."

Subject: Quick question about Acme's outbound Hi Sarah, I hope this email finds you well! I've been following Acme's impressive growth and noticed your recent Series B. Congratulations on the milestone! I wanted to reach out because I think there might be some real synergies between what we're doing at Orcha and what you're building at Acme. We help fast-growing teams like yours streamline outbound, accelerate pipeline creation, and unlock revenue at scale through our AI-powered platform. Would you have 15 minutes for a quick chat next week to explore how we might support Acme's growth journey? Best regards, Danny

Now the same email written by someone who has read three of Sarah's actual posts and understands what is happening at Acme.

Subject: hiring 3 SDRs after the Series B? Sarah, Saw the Series B and the three SDR job posts on LinkedIn. Curious how you're planning to ramp them, since reply rates across most outbound right now are sitting near 5% and a lot of teams in your spot are leaning harder on signals than on headcount. If it's useful I can share what we're seeing from teams who just hit the same point. No pitch, happy to send notes. Danny

Same length, roughly. Completely different signal. The second one demonstrates research in the subject line, names a specific tradeoff the prospect is actively weighing, and removes the meeting ask entirely. It reads like one operator messaging another, because that is what it is.

The first one will get archived. The second one gets a reply, even if the reply is "not now."

The principles underneath

Specificity beats personalization. Personalization is the surface ("Hi Sarah, I saw you went to Northwestern"). Specificity is the substance ("you posted last week that your team's connect rate dropped 30% after the Apple privacy update"). Personalization can be faked by any AI with a LinkedIn scraper. Specificity cannot, because it requires you to actually understand what is happening to this person right now.

Brevity is a credibility signal. Long emails read as uncertain. They imply the sender is throwing everything in because they do not know which thing will land. A short email implies the sender knows the one thing that matters to this prospect and is willing to bet the email on it.

The unspoken value of "no pitch." Removing the meeting ask in the first email does more than save space. It changes the frame. The prospect is no longer being sold to. They are being given the option to engage or not. The reply rate goes up because the cost of replying just dropped to zero.

Tone matches the channel. Cold emails should read like a Slack message from a peer, not a press release from a vendor. If you would not type the sentence into a Slack DM, do not type it into a cold email.

Subject lines work like text messages. Lowercase. Specific. Curious. "hiring 3 SDRs after the Series B?" works. "Quick question about Acme's outbound" does not, because it is the same subject line every other cold email used today.

Why this matters more in 2026

Cold reply rates have collapsed to around 5% from roughly 7% the year before. The reason is not that buyers got meaner. The reason is that AI made it cheap to send a hundred mediocre emails for every one good one, and buyers calibrated their threshold accordingly.

The math has flipped. Volume used to be the lever. Now relevance is. A focused SDR sending 25 well-researched emails outperforms a high-volume SDR sending 80 templated ones, by a wide margin. The teams hitting quota in 2026 are running fewer touches with sharper context.

Which means the AI is not the problem. The AI is fine. The problem is that most teams are using AI to generate more of the same patterns buyers already learned to filter.

The fix is not to stop using AI. It is to give the AI better inputs. The signal that triggered the outreach. The prospect's recent posts. The specific tradeoff they are weighing. The same context a senior SDR would gather in 20 minutes of LinkedIn scrolling, fed in at the start.

That is the bet behind everything we are building at Orcha. Not AI that writes emails from firmographic fields, which is what most tools do. AI that writes emails the way a thoughtful human would, because the same context that triggered the outreach is the context the email is built from. (See our piece on bracket personalization for the longer version.)

The shortcut for now: read the prospect's three most recent posts before you write a single word. If you cannot reference one of them in the email, you are not ready to send it.

That single habit will outperform every prompt-engineering trick you can find online.