Rémi KokabiNovember 4, 20268 min read
AI email personalization that doesn't sound like a bot (prompts and examples)
“I loved your recent post” used to work. Now prospects spot it in half a second and delete.
AI made personalization cheap. Every rep can generate a custom opener in seconds, so every inbox fills up with lines that sound custom but mean nothing. The problem was never the AI. It’s the data going in and the rules guiding what comes out.
In this article, I’ll show you how to do ai email personalization in a way that still sounds like a human wrote it. You’ll see what to personalize and what to leave alone, what data actually moves reply rates, and the prompts that stop AI from inventing facts or sliding into flattery.
Get this right and your emails read like you did the research yourself. Get it wrong and you’re just automating the thing prospects already ignore.
What AI email personalization is (and isn’t)
AI email personalization means using AI to write specific parts of a cold email, like the opening line or the reason for reaching out, based on real data about each prospect. It does not mean letting AI rewrite your entire email from scratch for every lead.
The distinction matters. A fully AI-generated email often reads like a template wearing a costume. The structure, tone, and phrasing repeat across hundreds of sends, and prospects notice. What works is keeping your core message fixed and letting AI handle one or two lines that connect your offer to something true about the recipient.
That also rules out flattering openers. “I was blown away by your company’s innovative approach to X” is not personalization. It is filler that any AI can produce without knowing anything about the person.
Why most AI personalization sounds like a bot
Most AI personalized emails fail for the same handful of reasons:
The input data is too generic (job title and company name alone)
The prompt asks for a compliment instead of a relevant observation
The output is too long, burying the point in three sentences when one would land
Every email follows the same structure, so 50 prospects in the same industry get lines that feel identical
The AI invents a fact that does not exist, like referencing a blog post the prospect never wrote
Each of these problems traces back to the prompt or the data, not the AI itself. Fix the inputs and you fix the output.
What to personalize (and what to keep fixed)
Not every part of your email should change per lead. Personalizing the wrong section adds complexity without improving replies.
Part of the email | Personalize with AI? | Why |
|---|---|---|
First line | Yes | Shows you know something relevant about the prospect |
Reason for reaching out | Yes, tied to a signal | Explains your timing and makes the email feel less random |
Value proposition | Lightly, by persona | Keeps your message consistent while adjusting for role |
CTA | No | A simple question works best and does not need to change |
The first line and the reason for reaching out carry almost all the weight. If those two feel relevant, the rest of the email can stay the same across your campaign. If you try to personalize everything, you lose control over your message and make QA nearly impossible.
The data to feed the AI
AI can only personalize from what you give it. Vague inputs produce vague lines. Here are the data types that consistently lead to useful output:
Company news (product launches, expansions, partnerships)
Hiring activity (open roles that signal a new initiative or gap)
Tech stack (tools the prospect already uses or recently adopted)
LinkedIn activity (posts, comments, job changes)
Website content (messaging, case studies, recent updates)
The challenge is gathering this data for hundreds of leads without spending hours on manual research. lemlist’s AI agentic enrichment handles this by sending AI agents to research each lead across websites, LinkedIn, CRM data, and other sources, then structuring what they find into fields you can use directly in your campaigns.
Once you have that context, the next step is identifying which signals actually indicate a prospect might care about your offer right now. Not every data point is a buying signal. A company hiring three SDRs is a signal if you sell sales tools. A company redesigning their homepage is not.
How to set up AI personalization step by step
1. Pick one signal per campaign
Do not try to cover every signal in a single sequence. Pick one, like “recently raised funding” or “hiring for a new role,” and build the campaign around it. One signal keeps your prompt focused and your output consistent. You can always create a second campaign for a different signal.
2. Write the prompt
Your prompt is where most AI email personalization succeeds or fails. A good prompt includes the data the AI should reference, a strict length limit, and rules about what to avoid.
Here are prompts you can adapt:
Signal: Hiring
Write one sentence (max 20 words) noting that {{companyName}} is hiring
a {{jobTitle}}. Mention what that hire suggests about their priorities.
No compliments. No invented facts. Use only the data provided.
Signal: Funding
Write one sentence (max 20 words) referencing {{companyName}}'s recent
{{fundingRound}}. Connect it to a likely next step for the team.
No compliments. No exclamation marks. Do not invent details beyond
what is provided.
Signal: New role
Write one sentence (max 20 words) acknowledging that {{firstName}}
recently started as {{jobTitle}} at {{companyName}}. Reference one
challenge common in the first 90 days of that role. No flattery.
No invented facts.
Notice the pattern: every prompt caps the length, bans compliments, and tells the AI to stick to the data it has. These constraints are what separate a useful line from a generic one.
3. Test on 20 leads before scaling
Run your prompt on 20 real leads and read every output. Look for hallucinated facts, repeated phrasing, and lines that could apply to anyone. If more than two or three outputs miss the mark, rewrite the prompt before scaling.
This step takes ten minutes and saves you from sending hundreds of emails with lines that damage your credibility.
4. Review and keep a fallback line
Not every lead will have the data your prompt needs. When a field is empty, the AI either skips the line or makes something up. Neither is acceptable.
Write a fallback line that works without any signal data. Something like: “Saw {{companyName}} is growing the team, curious how you’re thinking about outbound right now.” It is less specific, but it is honest.
In lemlist, you can use AI variables to generate personalized lines for each lead directly inside your campaign. You create an AI variable in your lead list, set your prompt, and the output populates as a variable you insert into your email steps. When the data is missing, your fallback line takes over automatically. For a deeper look at how to personalize at scale across your full pipeline, that workflow is the foundation.
Examples of good and bad AI personalized lines
The difference between a line that earns a reply and one that gets deleted usually comes down to specificity. Here are six pairs, organized by signal.
Hiring signal
Bad: “I noticed your company is growing fast. Congrats on the momentum!”
Good: “You’re hiring two data engineers, which usually means the pipeline is outgrowing the current stack.”
Funding signal
Bad: “Congratulations on your recent funding round! Exciting times ahead.”
Good: “Post-Series B, most teams start scaling outbound before the new hires ramp. That’s where we come in.”
New role signal
Bad: “Congrats on the new role! I’m sure you’ll do amazing things at the company.”
Good: “Three weeks into a new VP Sales seat, you’re probably auditing the existing tech stack. Worth a quick look at how we handle that.”
Tech stack signal
Bad: “I see you’re using some great tools. We’d love to help you optimize your workflow.”
Good: “You’re running HubSpot and Outreach but no enrichment layer, which usually means reps are doing manual research.”
LinkedIn post signal
Bad: “I loved your recent post about leadership. Really insightful!”
Good: “Your post on SDR ramp time mentioned 90 days as the benchmark. We’ve helped teams cut that by running signal-based sequences from day one.”
Website visit signal
Bad: “I noticed some activity from your team and wanted to reach out.”
Good: “Someone from {{companyName}} looked at our deliverability page twice this week. If inbox placement is on your radar, here’s what we’d check first.”
In every good example, the line references a specific fact and connects it to a problem the product solves. In every bad example, the line could have been written without knowing anything about the prospect.
Go beyond text
AI personalization does not stop at the written word. Personalized images, like a screenshot with the prospect’s logo or name embedded, can increase engagement because they break the pattern of a plain-text inbox. You can learn how personalized images in cold emails work without hurting deliverability.
Video is another layer. A short, personalized video intro where you say the prospect’s name and reference their company adds a human element that text alone cannot match. If you want to test this channel, here is a guide on video email for sales with practical workflows.
Both formats work best when the personalization is visual and quick. A two-second name mention in a video or a dynamic logo on an image is enough. You do not need to record a unique video for every lead.
Final takeaway
AI email personalization works when you treat AI as a research assistant, not a ghostwriter. Give it one real signal, set strict rules in your prompt, and keep the rest of the email in your own voice.
The data backs this up. Sequences run end to end by AI with no human review reply at 1.9%, while the same sequences where a rep reviews the research and the first line reply at 4.4% (lemlist internal data). The gap is not about the AI. It is about the inputs and the review.
And volume alone will not close that gap. Teams that doubled their leads per month saw reply rates drop from 4.3% to 2.9%. Teams that kept volume flat and invested the time saved into better research went from 4.0% to 4.2% (lemlist internal data).
Better inputs, tighter prompts, human review. That is the formula. Build your first signal-based sequence in lemlist.
Build cold email campaigns that work
Find verified leads and launch personalized cold email campaigns.