How to Personalize Cold Emails With AI

Three months ago I was prepping a 60-prospect outbound push and did something I now consider embarrassing. I opened ChatGPT, pasted in a prospect's LinkedIn URL, and typed "write me a personalized cold email to this person." What I got back was a paragraph that mentioned their job title, their company name, and the phrase "your impressive track record." I sent it anyway because I was tired. No reply. Obviously.

Here's the thing though — the AI wasn't the problem. I was. I used it like a lazy intern: no real data, no constraints, no direction. Just vibes and a LinkedIn URL. AI email personalization actually works. I've watched it pull 8-12% reply rates on lists that would've flatlined at 2% with templates. But only — and I cannot stress this enough — when you feed it actual inputs and tell it precisely what to do with them. What follows is the step-by-step version of everything I figured out after that embarrassing first attempt.

Step 1: Gather the Right Data (LinkedIn Is the Gold Mine)

Your CRM knows someone's title and company. Congrats, so does everyone else's CRM. To write an email that doesn't get instantly trashed, you need to know what this person cares about right now — this week — not whenever they last bothered updating their LinkedIn bio.

Here's my checklist before any AI touches the copy:

• Last 3-5 LinkedIn posts. What topics are they publicly invested in? If a VP of Sales posted twice about pipeline coverage ratios, that's your opening — not their job title.
• Recent job change. Someone 45 days into a new VP gig is ripping apart their predecessor's tool stack. That window lasts about 90 days. After that, they've already committed.
• Company news. Funding round, acquisition, product launch, splashy customer win. Money just moved. Priorities just shifted. That's your in.
• Mutual connections or shared groups. Not for name-dropping (that's weird). For calibrating tone and context. If you share a Pavilion membership, you're writing a different email than if they're a total stranger.
• Their company's G2 reviews or customer complaints. If their customers are publicly saying the onboarding is slow, and you sell onboarding software, that's a gift.

Most people skip all of this and ask AI to personalize from thin air. That doesn't work. AI can't invent context any more than you can. Feed it nothing, get "I noticed your impressive growth" back. You deserved that.

Step 2: Structure Your Prompt Like a Brief

This is where it all goes sideways for most people. They paste raw data into a prompt, type "write an email," and wonder why the result reads like it was written by a LinkedIn influencer having a stroke. You wouldn't hand a copywriter a random pile of papers and expect a finished ad. Same principle.

Here's the prompt structure I use now:

Context block: One paragraph about who this person is, what their company does, and what's happening in their world right now. I write this from the data I gathered. It takes about 30 seconds per prospect.

Constraint block: Tell the AI what NOT to do. This matters more than telling it what to do. My constraints:

• No compliments about their "impressive" anything
• No lines that could apply to anyone ("I see you're passionate about growth")
• Under 65 words total
• One specific observation, one sentence connecting it to the problem I solve, one question
• No exclamation points

Example block: Give it two examples of emails you've sent that actually got replies. Real ones. The AI will pattern-match on tone and structure without copying the content.

Night and day. An unprompted AI writes emails that all sound like the same LinkedIn comment bot — polished, vague, weirdly complimentary. A properly constrained one writes something that reads like you personally sat down, spent ten minutes on research, and banged out a thoughtful note. Took eight seconds though.

Step 3: Build the Assembly Line

Cool, your prompt works. Now what? Doing this one-by-one for 60 prospects is still painfully slow if you're running campaigns every week. You need a repeatable pipeline.

Here's mine:

1. Pull your prospect list from Apollo or your CRM — 40-60 people per batch
2. Run an enrichment pass that grabs LinkedIn activity, company news, and recent triggers for each prospect. This is the step that used to take me two hours and now takes about four minutes with an agent.
3. Feed enriched data into your prompt template — one per prospect, with the context block auto-filled from the enrichment data
4. Review the output. This is not optional. I reject about 20% of AI-drafted emails because the observation is weak or the connection to my product feels forced. That's fine. A human veto gate is what keeps this from turning into fancy spam.
5. Load approved emails into your sequencer — Instantly, Lemlist, Smartlead, whatever you use

Start to finish, 50 prospects, about 30 minutes. That's it. Most teams I talk to are stuck choosing between two bad options: burn three hours hand-writing everything (nobody keeps that up past week two) or skip personalization and blast templates (1.5% reply rate, congrats). This is the third option nobody told them about.

When AI Personalization Backfires

Nobody writing about this stuff wants to admit the failure modes. I will, because I've hit all of them.

The uncanny valley email. AI sometimes nails the structure but whiffs the meaning. I once sent an email referencing a prospect's LinkedIn post about "cutting headcount to stay lean." The AI interpreted it as a growth signal. The prospect had just laid off 40 people. She did not respond warmly. Read the AI's output. Actually read it. If the observation is even slightly off, trash it and move on.

Over-personalization. There's a line between "I noticed your post about pipeline metrics" and "I read your last fourteen posts, looked at your company's Glassdoor reviews, and found your podcast appearance from 2023." The second one makes you sound like a stalker, not a salesperson. One observation per email. One. That's the rule.

Scaling past your review capacity. I got cocky once and ran 200 prospects through the pipeline on a Friday afternoon. Didn't review half of them. Reply rate cratered to 3%. Went back to 50 with proper review the following week — bounced right back to 9%. Your throughput ceiling lives wherever your ability to spot bad emails breaks down. Respect it.

Why Use an Agent

Every step I just walked through lives in a different tool. LinkedIn in one tab. Apollo in another. ChatGPT in a third. Your email sequencer in a fourth. A spreadsheet trying to hold it all together. It's a mess, and honestly it's the part I hated most before I started using agents.

The Instantly + LinkedIn Personalized Sequences agent pulls LinkedIn data for each prospect and feeds it directly into Instantly with personalized context already embedded. No copy-pasting between tools. The Lemlist + LinkedIn Lead Enrichment agent does the same thing for Lemlist users — it grabs LinkedIn activity and recent signals so your campaigns reference real things the prospect did, not database fields. And if you want the enrichment layer without being locked to a specific email tool, the Lead Enrichment agent assembles company data, social profiles, and buying signals into a research brief you can use anywhere.

The agents don't replace your judgment. They replace the data-gathering gruntwork that sits between "I should email this person" and "here's what I should say."

The Short Version

Treat AI like a junior copywriter who needs a brutally specific brief — not a magic box you throw names into. Pull real data first: LinkedIn posts, company news, job changes. Write your prompt with constraints that prevent the sycophantic slop. Build a pipeline: enrich, draft, review, send, 40-60 at a time. Throw away the bad ones without guilt. And if the phrase "impressive track record" ever appears in your outbox, you've lost the plot entirely.

Try These Agents

• Instantly + LinkedIn Personalized Sequences — Build personalized cold email sequences using LinkedIn data piped into Instantly
• Lemlist + LinkedIn Lead Enrichment — Enrich leads with LinkedIn data for hyper-personalized Lemlist campaigns
• Lead Enrichment — Enrich any lead with company data, social profiles, and buying signals