Cold email subject lines that don't sound AI-written

Open the spam folder in your personal Gmail right now and read the subject lines. Notice the pattern? "Boost Your Revenue with AI-Powered Solutions." "Transform Your Marketing Strategy in 2026." "Discover the Future of Customer Engagement." Every one of them sounds like it was generated, because it was. After three years of GPT outbound, recipients have a finely-tuned filter for that voice. The opens go to people who write differently.

Here's what's working in 2026, based on the per-prospect subject lines we generate at Outreach and the ones I've watched friends hit 6%+ reply rates with.

Ten subject lines that work in 2026

1. quick question about [their specific product / new launch]
2. saw the [specific page on their site]
3. 1 thing on the [pricing / about / careers] page
4. [mutual connection] mentioned you
5. thought re: your [specific blog post / podcast appearance]
6. 30 seconds — [specific detail you noticed]
7. [their company] + [your category] = ?
8. noticed you're hiring for [specific role]
9. wrong person? [specific topic]
10. is this a good time to ask about [specific thing]?

Pattern: lowercase, short, references something specific that proves you actually looked at their company. No buzzwords. No "unlock," no "transform," no "revolutionize." Read those out loud — they sound like things a busy human would actually type.

Ten subject lines to never write again

1. Boost Your Revenue with Our Solution
2. Transform Your Business in 2026
3. Unlock Your Team's Full Potential
4. Discover the Future of [Their Industry]
5. Quick Question (no context)
6. Increase Your ROI by 300% Today
7. Hi [First Name] — A Quick Introduction
8. Re: Our Conversation (when you've never spoken)
9. Following Up — Did You See My Last Email?
10. Important: Action Required

Pattern: title-case, generic, no specificity, transparent fake-urgency. The "Re:" and "Important: Action Required" ones in particular have been getting people permanently blocked since 2023. Don't.

If your subject line could be sent to 10,000 different recipients without editing, it's already in spam. Make it impossible to send to anyone else.

The three formulas that hold up under scale

1. The question hook

"quick question about [specific]" — works because it's lowercase, short, sounds curious not pitchy, and proves you know what they do. The specificity is doing 90% of the work. "Quick question" with no specificity is a spam-trigger; "quick question about your new pricing page" reads like a customer.

2. The mutual-context hook

"saw your post on [topic]" or "loved the [podcast] appearance" — works because you're not selling, you're referencing. Even if the body of the email pitches, the subject earns the open. Only works if it's true. Pretending to have read something you didn't is how you become the villain in their next LinkedIn post.

3. The brutally specific ask

"30 seconds — does [specific thing] still go through [specific channel]?" — works because it gives the recipient a tiny clear action and proves you've done research. The 30-second framing also defuses the "this is going to be a long pitch" fear. Pair with a body that actually takes 30 seconds.

How Outreach handles this

When you launch a sequence in Outreach, Claude reads each prospect's homepage and pulls out one specific detail — a new product launch, a hiring page, a recent blog post — and writes the subject line around it. No two prospects in the same sequence get the same subject line. That's the difference between "AI-personalized" (a merge tag) and actually personalized (a paragraph of context the model wrote about this specific company).

If you're still typing subject lines by hand, the formulas above are the cheat sheet. If you want the model to handle it at scale, that's what the personalization layer is for. Either way: stop sending "Boost Your Revenue with Our Solution." It's 2026.

Related: the 1-line opener that gets a 40% reply rate for what to write in the body, and 2026 open-rate benchmarks for what numbers to actually aim for.