How to write cold emails with AI that still get replies

Open your inbox and count how many cold emails sound exactly the same. “I noticed your impressive growth.” “I wanted to reach out because…” AI wrote most of them, and you can tell in the first five words.

Here’s the part that surprised me when I pulled the numbers: AI isn’t the problem. How you use it is. In our data, some AI-assisted campaigns hit a 7.6% reply rate. Others, using the same model, barely cracked 1.5%. Same tool, five times the gap.

In this article, I’ll show you how to write cold emails with AI in a way that actually gets replies, not silence. I’ll walk through the two ways teams use AI right now, what the reply data says about each, and the exact workflow (plus prompts) that separates the 7.6% campaigns from the rest.

If you’ve ever wondered whether AI is quietly tanking your reply rate, the answer is in the numbers below.

Two ways teams use AI in cold email

The best way to write cold emails with AI is to use it for research and one personalized line per lead, inside a human-written template. The worst way is to let AI write the entire email from a name and a job title.

That split matters more than which model you pick or how clever your prompt is. We see it in the data every week.

AI for context means you give the model an account, a signal (a funding round, a job post, a product launch), and a reason to reach out. The AI researches the account and writes one specific sentence you drop into your email as a variable. The list is small or signal-based, usually under a few hundred leads.

AI for volume is the opposite. The model writes the full email, subject line to sign-off, and runs on a list of 1,000+ leads with little more than a name and a title as input. Both approaches use the same underlying technology. The difference is what you ask it to do. For a deeper look at the full spectrum, see this guide to AI cold outreach.

What the data says about AI and replies

We analyzed campaigns across 412.6M messages sent between October 2025 and November 2026. We grouped every campaign by how it used AI and measured the reply rate of each group.

The gap between the top and bottom row is five to one. Same tools, same platform, same inbox infrastructure. The only variable is how teams used AI.

Campaigns that used AI for context outperformed even campaigns with no AI at all. And campaigns that handed the entire email to AI performed worse than writing by hand. The tool is neutral. The workflow decides the outcome.

Why fully AI-written emails underperform

We looked at 5,645 campaigns where the entire email was generated by AI from a lead’s name and title. Here is what we found in lemlist internal data:

• •

  Email reply rate: 1.4% (against 2.7% for handwritten emails)

• •

  Positive reply rate: 0.4% (against 1%)

• •

  Bounce rate: 9.6% (against 7.8%)

Run those numbers across 1,000 emails. A fully AI-written campaign gets about 8.6 replies. A well-structured email with one AI-personalized line gets up to 30.

The reason is straightforward. When you give AI a name and a title, the only thing it can do is guess. It guesses the pain point, guesses the industry challenge, guesses the opener. Every guess sounds the same because the input is the same. Generic input produces generic output, and inboxes are already full of it.

The higher bounce rate tells a second story. Teams that fully automate the writing also tend to fully automate the list. Bigger, colder lists mean more invalid emails and more spam complaints.

1. Use AI to research the account

Before AI writes a single word of copy, it should read. Give it the account, a signal, and a reason you are reaching out.

The signal is the event that makes now the right time: a new VP of Sales hired, a Series B closed, a competitor mentioned in a review, a product launch on the changelog. The reason is why that signal matters to your offer.

Then point the AI at sources a rep would never have time to read manually:

• •

  Earnings call transcripts

• •

  Open job postings and their descriptions

• •

  G2 or Capterra reviews mentioning specific pain points

• •

  Product changelogs and release notes

• •

  LinkedIn posts from the buyer

The output is a short research brief, not an email. That brief becomes the raw material for the one line you actually insert. For a step-by-step workflow on setting this up, see this guide to AI prospect research.

2. Let AI write one line, not the whole email

The highest-performing pattern in our data is a human-written email with one AI-generated variable: a single researched sentence that changes per lead.

That sentence is what AI variables do inside lemlist. You write the email once. The AI fills in one dynamic line per lead based on the research context you gave it.

A good AI line meets three criteria:

• •

  Specific. It references something only true for that lead or account.

• •

  Verifiable. The prospect can confirm it with a quick glance (their own job post, their own LinkedIn post, their own product page).

• •

  About the prospect. It talks about their world, not yours.

A bad AI line is a compliment about “impressive growth” or “innovative approach.” Those phrases could apply to any company on earth. If the line works for 500 leads, it works for none of them.

3. Keep the list small and the reason clear

Context only works when there is context to use. A list of 5,000 leads scraped from a job title filter gives AI nothing specific to say. That is why AI for context campaigns run on small, signal-based lists.

The timing matters as much as the list size. In lemlist internal data, leads contacted the same day a signal fires reply at 21.3%. Wait a week and the signal is stale. Wait a month and someone else already had the conversation.

Build your lists around buying signals: a funding round, a key hire, a tech stack change, a website visit. Each signal gives AI a concrete fact to reference and gives the prospect a reason to care right now. The smaller the list and the fresher the signal, the better the AI line and the higher the reply rate.

4. Review before you send

AI hallucinates. It invents product features, misreads job titles, and confuses companies with similar names. A single wrong fact in your opener kills trust faster than a generic one.

Before launching any campaign with AI-generated lines, run through this checklist:

• •

  Verify the fact in the AI line. Open the source (the job post, the LinkedIn profile, the review) and confirm it is real.

• •

  Remove generic phrases. If the line says “I noticed your company is growing fast,” delete it and leave the slot empty. An empty variable is better than a fake one.

• •

  Read the email on a phone. Most prospects will. If it scrolls past one screen, cut it.

• •

  Keep one ask. One question, one next step. Two asks split attention and drop reply rates.

You do not need to check every lead. Sample 10 to 15 leads before launch. If more than two have a wrong or generic AI line, fix the prompt before sending to the full list.

Prompts that give AI enough context

Most AI cold email prompts fail because they are too open. “Write a cold email to this person” gives the model no constraints, so it fills the gap with filler.

A good prompt for a personalized line has five parts:

1. 1.

   Who the lead is. Their role, their company, and what the company does.

2. 2.

   The signal. The specific event or data point that triggered outreach.

3. 3.

   The campaign angle. What your product does for someone in that situation.

4. 4.

   A format limit. One sentence, under 25 words, no exclamation marks.

5. 5.

   A ban on invented facts. If the information is not in the provided context, leave the field blank.

Here are three prompt specs you can adapt:

The format limit is the most important part. Without it, AI defaults to a full paragraph. The ban on invented facts is the second most important. Without it, AI will fabricate a detail to fill the space.

How to write cold emails with AI in lemlist

lemlist gives you five features that map directly to the workflow above. Together they cover the full cold emailing process from research to send.

AI variables generate one researched line per lead inside your email. You write the prompt once, and the variable populates a unique sentence for every lead in the campaign based on their data. This is the core of the AI email personalization workflow.

AI agentic enrichment researches accounts before the email is written. AI agents pull structured context from websites, LinkedIn profiles, CRM records, and Claap sales call recordings, then return it as fields you can use in prompts, filters, and sequences.

lemAgent drafts a full multichannel sequence from a chat-style prompt. It builds the steps, writes the messages, and assembles the campaign. You review and edit everything before anything sends.

Intent signals track events like funding rounds, hiring changes, tech stack swaps, and website visits. When a signal fires, leads are automatically added to campaigns with messaging tied to that exact moment.

A/B testing lets you run two versions of a sequence and pick the winner by reply rate. You test the AI line, the subject line, or the full structure, and the data tells you which version to keep.

Final takeaway

AI makes cold email better when it researches and worse when it writes the whole thing. Use it to generate one specific, verifiable line per lead inside a human-written email, and run it on small lists built around fresh signals.

The data is clear: same tools, five times the replies.

Build your first signal-based sequence in lemlist.