Loading tutorials…
Loading tutorials…
AI Personalization writes a custom opener per lead based on their LinkedIn + website. Done right, reply rate jumps 2-3x. Done wrong, AI generates generic 'I noticed you do great work' fluff that hurts reply rate. This walks the right setup.
Who this is forCold email operators on Hypergrowth+ who want to scale personalization beyond first-name. If you have 500+ leads per campaign and manual research is not feasible, AI is the leverage move.
What you'll need
Step 1
AI Personalization needs LinkedIn URL OR company website per lead. Without these, output is generic.
AI Personalization scrapes LinkedIn + company website to generate per-lead context.
Required: LinkedIn URL OR company website URL per lead. Both is better.
Pre-clean: verify URLs are valid (no typos, no broken links). Run a quick spot check on 20 leads.
If 30%+ of leads lack LinkedIn URLs, source from Apollo/Clay/Lead Finder to fill the gap before running AI.
Garbage in = garbage out. AI cannot generate good personalization from missing data.
Step 2
The prompt tells AI WHAT to generate. Be specific. "Write a 1-sentence opener referencing [their role + a recent achievement]" beats "Write a personalized intro."
Instantly's AI Personalization uses a prompt + LinkedIn/website data to generate output.
Bad prompt: "Write a personalized intro for this person."
Good prompt: "Write ONE sentence opening for a cold email to [lead]. Reference something specific from their LinkedIn (a role transition, a post they made, or company growth). Keep it under 20 words. Make it feel like you read their profile, not like AI."
Better prompt: same as above + give 2-3 examples of the format and tone you want.
The prompt also defines tone (conversational vs formal), length, and what to AVOID (generic phrases like 'I noticed your impressive work').
Step 3
Generate openers for 20-50 leads. Manually review every output. Iterate prompt based on what AI gets wrong.
Run AI on 20-50 leads as a test batch.
For each output, score: usable as-is / needs minor edit / unusable.
Target: 70-80% usable as-is. If lower, iterate the prompt.
Common AI failures: invents facts (says "you started X" when they didn't), reuses the same phrase across leads, gets the role/company wrong, generic fluff.
For each failure pattern, update the prompt: "Do not invent facts. Use only what is in the LinkedIn data provided." or "Do not use the phrase X."
After 2-3 prompt iterations, output quality stabilizes.
Step 4
After prompt is tuned, run AI on the full list. Each lead gets a custom opener stored in a custom field.
In Instantly: Lead list → AI Personalization → run on the full list.
Each lead gets a custom opener stored in a field like `ai_opener`.
For 1,000 leads, this takes 10-30 minutes (depending on credit pool and rate limits).
Cost: ~$0.10-0.30 per lead in AI credits. 1,000 leads = $100-300.
After generation, review a sample of 20-30 outputs to confirm quality stayed consistent.
Step 5
In your campaign template, use the `{{ ai_opener }}` variable as the first line. The rest of the email is standard template.
In your campaign's step 1 template, structure the email body as: opening line uses `{{ ai_opener }}` variable, followed by a blank line, then your standard value statement, another blank line, then your standard ask. The AI opener replaces the generic personalization line.
Rest of the email stays consistent across all leads — only the opener is dynamic per recipient.
For multi-step sequences: AI opener appears on step 1 only. Steps 2-6 use standard templates with first_name + company_name personalization.
Step 6
Send 5-10 test emails to internal addresses. Read each one as a recipient would. Look for awkward phrasings, hallucinations, or generic outputs.
Before launching: send test emails to 5-10 internal addresses (yourself, teammates).
Read each as a recipient would. Does the opener feel natural? Is it specific? Or does it read as obvious AI?
Common giveaways of AI personalization done wrong:
- Same sentence structure across emails ("Saw your work at... and was impressed by...").
- Vague specifics ("your recent growth," "your innovative approach").
- Generic praise without substance.
If outputs read as AI to internal readers, they read as AI to recipients. Iterate prompt or supplement with manual review.
Step 7
Compare reply rate of AI-personalized campaigns vs non-personalized. Target 1.5-3x uplift. If no uplift, AI quality is insufficient.
A/B test: same campaign, same audience, half with AI personalization, half with basic first-name personalization.
After 500+ sends per arm, compare reply rates.
Healthy uplift: 1.5-3x reply rate on AI arm vs non-AI.
If uplift is small (under 1.3x), AI quality is not earning its cost. Iterate prompt or supplement with manual research for highest-value leads.
Document the prompt + uplift data per campaign. Build a library of prompts that work for specific ICPs.
Common mistakes
Lead data missing LinkedIn URLs
What goes wrong: Half of leads have no LinkedIn URL. AI Personalization runs but generates generic output for those leads (no data to personalize from). Reply rate uplift evaporates.
How to avoid: Pre-clean: ensure 90%+ of leads have LinkedIn URLs. Source missing URLs via Apollo or manual research before running AI.
Vague prompt
What goes wrong: Prompt says 'write a personalized intro.' AI generates generic 'I noticed you work in tech' for every lead. Output is unusable.
How to avoid: Specific prompt with examples. Tell AI: tone, length, what to reference, what to avoid. Iterate based on actual output.
Not testing on a small batch first
What goes wrong: Run AI on full 5,000-lead list. Discover at send time that 60% of outputs are unusable (hallucinations, generic, wrong company). $1,000 of AI credits wasted.
How to avoid: Always test on 20-50 leads first. Iterate prompt. Only run full batch after quality is consistent.
Trusting AI output without review
What goes wrong: Skip the review step. AI hallucinates 'congrats on raising your Series A' for a bootstrapped company. Recipient calls out the error. Embarrassing + signals you sent without reading.
How to avoid: Review at least 10-20 outputs per campaign before launching. For high-stakes campaigns (enterprise outreach), manual review of every opener.
No A/B comparison vs non-AI
What goes wrong: Run AI on every campaign. Assume it helps. Don't measure. Pay $100-300/campaign in AI credits for marginal lift you do not actually have.
How to avoid: A/B test AI vs non-AI on similar campaigns periodically. Confirm uplift is real before scaling spend.
Recap
Done — what's next
How to build cold email templates in Instantly
Read the next tutorial
Hand it off
AI Personalization is a leverage move when prompts are tuned and quality is monitored. Most operators run it badly and abandon. EverestX specialists with AI experience tune prompts to 80%+ usable output and maintain quality across campaigns. Typically $400-1,000/mo at $14-16/hr.
See specialist rates
Roughly $0.10-0.30 per lead in Instantly AI credits. For a 1,000-lead campaign, $100-300 in AI costs. Hypergrowth+ plans include 5K-10K AI credits/month; beyond that, top-up credits are available.
When done right: 1.5-3x uplift over generic personalization. Done wrong: no uplift or slight decrease. Quality depends on prompt + data quality + review discipline.
Sophisticated recipients increasingly can. Telltale signs: generic phrasing, identical sentence structure, vague references. Good AI personalization is indistinguishable from manual research; bad AI is obvious.
Not necessarily. For high-touch enterprise outreach (< 500 leads), manual research often produces better output. For SMB outreach (> 1,000 leads), AI is the only scalable option. Match approach to campaign size + stakes.
Yes, via integration. Export leads → run through your own LLM via API → import enriched leads back with `ai_opener` field populated. More work, more control. Some teams prefer this for prompt flexibility and cost control.
Instantly
Most cold email templates are bad. Generic openers, weak value statements, vague asks. This walks the template structure specialists use to hit 3-8% reply rates instead of 0.5-1%.
Instantly
Instantly's Lead Finder is one of its biggest underrated features — built-in B2B lead sourcing without paying for a separate tool like Apollo or ZoomInfo. This walks the right setup.
Instantly
Most cold email "campaigns" are one cold email + one follow-up that lands in spam. Real campaigns have 4-6 step sequences, deliberate cadence, throttled sending, and A/B testing. This walks the right setup.
Instantly
DIY cold email is great until reply rates plateau under 1% and you cannot figure out why. This is the honest framework: when the cost of self-managing exceeds the cost of a specialist, and how to tell which side you are on.
Lemlist
Lemlist AI can write a full sequence in 60 seconds. The output is also instantly recognizable as AI to anyone who reads cold email all day. Here's how to actually use AI effectively — for ideation and personalization, not for finished copy.