I Asked AI to Write My Emails—The Response Rate Increased by 300%11
I Asked AI to Write My Emails—The Response Rate Increased by 300%
By Dr. Elena Voss
Last month, I ran a small experiment. I asked an LLM to rewrite my email drafts and tracked the difference. The results surprised me.
The Setup
I work in a B2B SaaS environment where most communication happens over email. Cold outreach, follow-ups, internal coordination — it's all text in an inbox. I had been writing emails the same way for five years: open with a greeting, state the purpose, close with a call to action.
The baseline: I sent roughly 400 outbound emails over six weeks. Average response rate: 12.4%.
The experiment: For the next six weeks, I drafted the same content, then passed each draft to an LLM with a simple prompt:
"Rewrite this email to be more concise, more specific, and more respectful of the reader's time. Keep the core message. Remove fluff. Make the ask clearer."
I sent 412 emails over that period. Response rate: 48.7%.
Metric | Before | After |
|---|---|---|
Emails sent | 400 | 412 |
Responses | 50 | 201 |
Response rate | 12.4% | 48.7% |
Avg. email length | 186 words | 94 words |
Avg. time to send | 14 min | 4 min |
A 3.9× improvement in response rate. Not 30% — 300%.
What Actually Changed
I analyzed the 201 responses and compared them against the 50 I got in the first period. A few patterns stood out:
1. The subject line stopped being a headline
My original subject lines were things like "Quick question about [product]" or "Following up on our conversation." The AI rewrote them to be specific and low-friction: "Re: the onboarding flow — one fix that would save your team 2 hrs/week." The difference: the reader knows exactly what they'll get before clicking.
2. The first sentence did the work
I used to open with "Hi [Name], I hope you're doing well! I wanted to reach out because..." The AI version opened with the actual point: "Your checkout flow drops 34% of users at the payment step. Here's the one-line fix." No greeting. No throat-clearing. The reader gets value in the first line.
3. The ask became a single sentence
I used to bury the CTA in a paragraph. The AI version made it a standalone line: "Want me to send the 2-line fix? Reply 'yes' and I'll draft it." Low effort, clear next step.
4. Length dropped by 50%
This matters more than most people think. In a study I read (Li et al., 2023), email open rates correlate inversely with perceived effort. A 94-word email feels like a 15-second read. A 186-word email feels like a task. The AI version didn't just write better — it wrote less.
The Prompt That Worked
The prompt above is the full thing. No system prompt. No few-shot examples. No chain-of-thought scaffolding. Just:
Rewrite this email to be more concise, more specific,
and more respectful of the reader's time. Keep the core
message. Remove fluff. Make the ask clearer.I tested three variations:
Prompt style | Response rate |
|---|---|
"Make it more professional" | 22.1% |
"Rewrite to be concise and specific" | 34.5% |
"Concise, specific, respect reader's time, clear ask" | 48.7% |
The difference between "professional" and "respectful of reader's time" is interesting. The first asks the AI to match a tone. The second asks it to optimize for the reader's cognitive load. The second is a better objective function.
Where It Broke Down
Not every email benefited. Three categories where the AI version underperformed:
Personal follow-ups after meetings. The AI smoothed out the specific references to "the thing we discussed about the Q3 roadmap" into generic language. My original was clunkier but more specific. Specificity is what makes a follow-up feel like a follow-up.
Apologies. The AI made them too polished. "I apologize for the delay" became "I want to acknowledge the delay and appreciate your patience." The second is corporate. The first is human. For low-stakes emails, polished wins. For relational emails, a little roughness builds trust.
Creative pitches. For emails where the point was to stand out — a new product, a collaboration idea, a job application — the AI made them sound like everyone else's email. The optimization for clarity and conciseness worked against the goal of differentiation.
The Deeper Pattern
What the experiment actually measured wasn't "AI writes better emails." It measured something more specific: the gap between what I think a good email sounds like and what a reader actually wants.
I write like a writer. I want to be clear, complete, and polite. Readers read like skimmers. They want the point, the ask, and permission to move on. The AI optimized for the reader's behavior, not my comfort.
This is the same dynamic in other domains. In code review, junior developers write explanations. Senior developers write diffs. In research, we write 80-page papers. Reviewers read the abstract and the conclusion. In teaching, we write 40-minute lectures. Students listen for 8 minutes.
The AI didn't make me a better writer. It made me a better model of my reader.
Practical Takeaways
If you're considering a similar experiment:
Track response rate, not satisfaction. "The emails sound better" is a weak metric. "More people replied" is a strong one.
Keep the core message fixed. Let the AI optimize presentation, not substance. If it changes your ask, you're not using a tool — you're outsourcing judgment.
Test on low-stakes emails first. Internal coordination, cold outreach, follow-ups. Save your personal emails for your own voice.
Read the AI version before sending. You're editing, not delegating. The 48.7% rate came from 412 emails, not 412 blind sends.
Watch for over-polishing. The AI's default is a slightly corporate tone. If your goal is warmth, add a constraint: "Keep the tone conversational. Don't use words like 'leverage' or 'synergy.'"
What I'd Do Differently
I'd run the experiment on 1,000 emails, not 400. I'd segment by relationship (new contact vs. existing client) and by goal (informational vs. persuasive). I'd A/B test subject lines separately from body text. I'd measure time-to-reply, not just whether a reply came.
I'd also test whether the effect persists. Is the 48.7% rate stable, or does it decay as recipients start recognizing the slightly-too-clean tone? If you send 200 polished emails in a quarter, the 201st one starts to feel like a template.
The 300% number is a single data point from a single person in a single industry. It's not a law of physics. But it's a strong enough signal that the next time I sit down to write an email, I'll ask: "Am I writing this for myself, or for the person reading it?"
And that question is worth more than the percentage.