We Analyzed Every Cold Email Our AI Sent for 100 Days. The Conventional Wisdom Is Wrong.

Emails over 125 words got zero opens. Not 2%. Zero.

We tracked every cold email our AI agents sent over 100 days — 1,520 emails, 33 campaigns, 16 real businesses, 12 industries. We logged the send time, word count, subject length, whether the opener used a first name, whether it was a follow-up or initial outreach, whether it opened, whether it bounced.

The playbook said Tuesday morning, personalize the subject, keep it short. Some of that held up. A lot of it didn't.

Here's the full breakdown with raw numbers.

The quick version

• 1,520 clean sends across 33 campaigns (February to May 2026)
• 57.6% overall open rate
• Emails under 75 words: 60.4% opens. Emails over 125 words: 0% opens. Not a typo.
• "Hey [Name]," opener added 11 points. Nothing else moved the needle.
• Saturday beat Tuesday by 7 points. Tuesday came in fifth out of seven days.
• Fintech opened 94% of emails. Wedding venues: 0%.
• Follow-ups outperformed first sends by 9 points.

How we ran this

We pulled every outbound email from our production database between February 17 and May 25, 2026. We filtered out 26 bounced sends. What remained: 1,520 emails sent by AI agents on behalf of 16 businesses across construction, fintech, SaaS, bookkeeping, restaurants, real estate, and several niche verticals.

Open rates are pixel-tracked. Some prefetch from inbox protection services gets counted — we're not claiming these are pure human opens. The relative patterns between buckets are what matter here. A 60-point spread between industries and a 30-point swing between word count buckets doesn't disappear in the noise.

We didn't analyze reply rates. Our reply ingestion was incomplete during this window, so we're reporting what we measured cleanly: opens, bounces, timing, and structure. We'll rerun at 10,000 sends with reply data.

Finding 1: The 75-word cliff

This was the starkest finding and it wasn't close.

We split every email into word count buckets and computed open rate per bucket.

Word count	Sends	Open rate
Under 75 words	1,392	60.4%
75 to 124 words	113	31.0%
125+ words	15	0.0%

The 75-word line isn't a slope. It's a cliff. Cross it and open rate gets cut in half. Cross 125 and it goes to zero.

Cold inboxes get scanned, not read. The recipient sees the preview text and decides in under a second.

A four-sentence email reads like a message from a person. An eight-sentence email reads like a pitch deck. They don't open pitch decks.

Your cold email isn't competing on quality. It's competing on preview length.

Cut your email to under 75 words. Then read it again and cut more.

Finding 2: One personalization signal added 11 points. Nothing else moved.

Our agents read each prospect's website before drafting. Some emails open with "Hey [Name]," using a name pulled from the site. Some don't, either because the agent couldn't find a name or chose a different opener.

We split on that one signal.

Opener	Sends	Open rate
"Hey [Name],"	783	63.0%
No personalized opener	737	52.0%

11 points from one change. The agent doing the obvious thing outperformed the agent doing fancier work.

We also looked at subject line personalization. Subjects with a question mark opened at 58.0% versus 57.5% without. Flat.

So it's not personalization in the abstract that matters. It's the part the recipient sees first. Address them by name. That's it.

Finding 3: Saturday beat Tuesday. By 7 points.

Every tool, every guide, every "best practices" post says send on Tuesday or Wednesday morning. Our data says otherwise.

Day	Sends	Open rate
Sunday	151	53.6%
Monday	280	59.6%
Tuesday	234	55.1%
Wednesday	228	55.7%
Thursday	273	55.7%
Friday	212	61.8%
Saturday	142	62.7%

Tuesday came in fifth. Saturday came in first.

Our read: the conventional advice was written for a world where everyone ignored it. Now everyone follows it. If every cold email tool fires at 9am Tuesday, that's the highest-competition moment of the week. A short, well-written email on Saturday morning lands in a near-empty inbox.

Tuesday is crowded. Saturday is quiet. Send where there's less competition.

Try it for two weeks. If your audience looks like ours, the difference will show up fast.

Finding 4: Subject lines over 50 characters tank opens

Same bucketing approach, applied to subject line length.

Subject length	Sends	Open rate
Under 30 chars	466	61.6%
30 to 49 chars	985	57.7%
50 to 69 chars	69	30.4%

Mobile preview cuts off around 50 characters. Past that, the recipient sees a fragment. Fragments look like marketing. The opens drop by half.

Our agents learned this on their own. Only 4.5% of emails had subjects over 50 characters. The ones that did paid for it every time.

Finding 5: Fintech opened 94% of emails. Wedding venues opened 0%.

We computed open rate by industry for every vertical with at least 30 sends. The spread was not what we expected.

Industry	Sends	Open rate
Fintech	89	94.4%
Podcast production	46	89.1%
AI and accelerators	42	85.7%
B2B SaaS	55	78.2%
SaaS	103	66.0%
Fractional CFO	169	64.5%
Bookkeeping	138	59.4%
Real estate	41	56.1%
Construction	329	54.1%
Luxury residential	63	52.4%
Restaurants and bars	93	46.2%
Wedding venues	67	0%

A 94-point spread between the top and bottom industry. Same AI agent. Same email structure. Different people on the other end.

The pattern is obvious once you see it. Industries where the decision-maker lives in their inbox (founders, SaaS execs, podcast producers) open everything. Industries where the recipient is physically busy (restaurants, wedding venues, construction sites) don't engage with email at all.

If you're selling to wedding venues or restaurants, cold email isn't your channel. Call them. Show up. Email won't move them.

Finding 6: The follow-up outperformed the first email

Email type	Sends	Open rate
Initial outreach	845	54.7%
Follow-up	651	63.6%

Follow-ups won by 9 points. Even though every follow-up went to someone who didn't respond to the first email.

Here's why: a follow-up lands in an existing thread. The recipient pattern-matches "I know this sender" and opens. A first cold email has no thread, no context, no prior recognition to borrow.

Most teams send one email and stop. The data says the second email is the more valuable send. Don't stop at one.

Finding 7: Bounce rate held at 1.68%

26 bounces across 1,546 sends. That's 1.68%.

The safe threshold is under 2%. Above that, inbox placement starts to erode. We held well under it with no external verification service beyond what the agent does at lead intake.

Most cold email teams hitting scraped lists see 4 to 8% bounce and slowly damage their sending domain. The agents pre-verify before sending. The domain stays healthy.

What to do with this

If you're running cold outreach manually and want to apply this tomorrow:

1. Cut to under 75 words. Then read it again and cut more. The cliff is real and it's steep.
2. Open with the person's first name. "Hey Sarah," not "Hi there" or "Dear Decision Maker." Pull the name from their site or LinkedIn profile.
3. Try Friday and Saturday for two weeks. Quieter inboxes. Less competition from every other tool firing Tuesday morning.
4. Keep subjects under 50 characters. Your constraint is the mobile preview window, not the email client.
5. Send the follow-up. One email and done is leaving opens on the table.
6. Be honest about your industry. If you're selling to restaurants or wedding venues, find a different channel.

A note on AI agents

We didn't write these emails. Our agents did. And the patterns they converged on without being told: short, personalized opener, well-timed, follow-up heavy. That's exactly what a good salesperson doing this by hand would write.

The interesting thing isn't that AI can write cold email. It's that AI given a real objective (get a reply) will evolve toward the same playbook a human would use. Just at higher volume and without needing to be reminded to follow up.

If you want an agent running this on your list, LeadClaw offers a 7-day free trial. The agent reads your site, researches the prospect, drafts the email, sends it, follows up on schedule, and stops when it gets a reply. Same patterns. Your business.

---

Study covers 1,520 outbound emails sent by LeadClaw AI agents between February 17 and May 25, 2026. Real production data across 33 campaigns and 16 businesses. Reply rates weren't included because our ingestion was incomplete during the window. We'll rerun at 10,000 sends.