AI cold email works when you use AI for the two things it’s genuinely great at: researching each prospect to personalize at scale, and handling replies instantly. It stops working the moment you use it for the thing it makes dangerously easy: sending more mediocre email to more people. Deliverability physics didn’t change because the copy is machine-written. Get the infrastructure right first, then let an AI SDR run the sequence. Here’s the whole system, including the 4-step sequence we’d start with.
Why does deliverability come before everything else?
Because since February 2024, Gmail and Yahoo enforce hard rules on senders: authenticated email (SPF, DKIM, DMARC), one-click unsubscribe for bulk senders, and a spam-complaint rate kept under 0.3% (Google Postmaster guidelines). Cross the line and your mail goes to spam for everyone, not just the people who complained. AI that triples your volume without fixing the foundation just gets you blacklisted three times faster.
The non-negotiables:
- Never send cold email from your main domain. Buy 2–3 lookalike domains (yourbrand-hq.com, tryyourbrand.com). If one burns, your real domain (the one your clients and invoices depend on) survives.
- Authenticate everything. SPF, DKIM, and DMARC on every sending domain. This is 30 minutes of DNS work and it’s mandatory, not optional.
- Warm up for 2–4 weeks. New inboxes start at 10–20 emails/day to engaged addresses, ramping gradually. Sending 500 cold emails from a week-old domain is how domains die.
- Cap volume per inbox. Hold each inbox to roughly 30–50 cold sends per day even after warm-up. Need more? Add inboxes and domains, don’t push the throttle.
- Verify every list. Run addresses through a verification tool before sending. A bounce rate above ~2–3% is a red flag to mailbox providers.
- Watch replies, not opens. Open rates are unreliable after Apple Mail Privacy Protection. Your health metrics are reply rate, bounce rate, and spam complaints.
Shared-pool horror stories are common in the CRM world. GHL users have documented open rates collapsing on shared sending infrastructure. Whatever platform you use, you want your own sending reputation, visible metrics, and throttled sending.
What does AI actually do well in cold email?
Three things, and they’re significant:
- Personalization at scale. The old trade-off was 10 researched emails a day or 500 templated ones. AI reads the prospect’s website, reviews, job posts, and news, and writes a specific first line and a relevant angle for each, hundreds of times a day. Specific first lines are what separate “deleted on sight” from “huh, they actually looked at us.”
- Reply handling and speed-to-lead. Half the meetings you’ll book die in the gap between a prospect replying and you noticing. An AI SDR answers questions, handles “what does it cost?”, and proposes meeting times within a minute, at 11pm, every time. (Stack Space’s Outreach SDR drafts or sends these replies and books straight into your calendar.)
- List intelligence. AI is good at filtering raw lists into “actually fits our niche,” like finding local businesses that are hiring or have no social presence, both buying signals worth more than any subject-line trick.
Hear the receptionist take a call — live demo on the homepage.
What does AI do badly (and how do campaigns die)?
- Spray-and-pray at machine speed. AI makes it free to email 10,000 people, which is exactly why you shouldn’t. Mediocre relevance × huge volume = complaint rates over 0.3% = every future email in spam. Volume amplifies whatever quality you have, including bad.
- Fake-personal ick. “I loved your recent LinkedIn post about growth!” reads as machine-written because it is. Prospects have seen thousands of these since 2024. Personalization has to be consequential, tied to the offer, not decorative flattery.
- Hallucinated claims. Unreviewed AI will invent case studies, stats, and capabilities. Every claim in your sequence should be one you can defend by hand. Never let AI promise outcomes or revenue on your behalf. Promising income you can’t back up is how you lose a prospect’s trust for good, and how honest tools get lumped in with the hype.
- Replacing judgment. AI drafts; you decide the offer, the niche, and what a qualified reply looks like. An autopilot needs a flight plan.
The 4-step sequence template (with copy)
One thread, 4 touches over ~14 days, each email under 90 words. The structure: relevance → value → proof → close the loop. Adapt the copy, keep the shape. The examples assume an agency selling missed-call rescue to home-services businesses. Swap your niche in.
Email 1 (Day 1) — the specific observation + one question
Subject: calls at [Business name]
Hi [Name] — I called [Business] Tuesday at 2pm and got voicemail. That’s normal (a big share of calls to small businesses ring out unanswered), but it’s expensive: most voicemail callers just call the next [plumber/roofer] on the list.
We fix that with an instant text-back plus an AI that answers after hours. Worth a 15-minute look at what you’re missing?
Email 2 (Day 4) — pure value, no ask
Subject: re: calls at [Business name]
No reply needed — here’s the 3-line fix whether or not we ever talk:
- Forward missed calls to any text-back tool, 2) put your booking link in the text, 3) answer within 5 minutes or the lead is 21× less likely to qualify (InsideSales.com research).
Happy to send the full checklist if useful.
Email 3 (Day 9) — proof or mechanism
Subject: what the text-back actually says
[Name] — one example: a caller hangs up, and 10 seconds later gets “Sorry we missed you — are you after a repair or a quote? Book a slot here: [link].” That one message is the difference between a saved job and a competitor’s customer.
Want me to set up a number you can call and test yourself?
Email 4 (Day 14) — the polite close
Subject: closing the file
I’ll stop here — timing is everything with this stuff. If missed calls ever make it onto your list, this takes about a week to fix. Either way: the checklist from my last email stands, no strings.
Good luck out there, [Name].
Why this works: touch 1 earns the right to exist (specific, provable), touch 2 gives before asking again, touch 3 makes the mechanism concrete, and touch 4 uses the mild loss-aversion of a closed loop. That last one consistently pulls replies from people who read everything and answered nothing.
What are the compliance rules? (CAN-SPAM, in plain English)
B2B cold email is legal in the US under CAN-SPAM if you follow the rules, and the fines for not doing so run to over $50,000 per violating email under the FTC’s current adjustments:
- No deceptive subject lines or headers. “Re:” on a thread that doesn’t exist is deception, not a growth hack.
- Identify yourself. Real sender name, real company.
- Include a physical postal address in every email.
- Offer a clear opt-out and honor it within 10 business days, and never email that person again. One-click unsubscribe is now required by Gmail and Yahoo for bulk senders anyway.
- Know your other jurisdictions. Emailing Canada (CASL) or the EU/UK (GDPR/PECR) is far more restrictive, consent-based rather than opt-out. If your list crosses borders, get advice. This isn’t legal counsel.
Compliance is also strategy: every rule above reduces spam complaints, and spam complaints are what kill deliverability.
Where does this fit in your larger system?
Cold email is one lane. The meetings it books still need same-minute follow-up, no-show rescue, and a CRM that remembers everything, which is why the sequence should live inside your platform, not in a disconnected sender. Stack Space keeps the whole outbound loop in one system: the Outreach SDR sends and replies, the AI receptionist catches the calls your emails generate, and Neo, the AI brain, trains all 17 AI employees so nothing falls between them. The Outreach SDR runs sequences like the one above, the Lead Finder builds niche lists (including “hiring” and “no social presence” filters once you connect a data source), and every reply lands in the same inbox as your calls and texts. See the best AI to run your business for how the pieces fit, how AI runs the boring 80% for the automation side, and — if you’re an agency using outbound to land clients — how to get your first 5 agency clients for where cold email does and doesn’t belong at your stage.
FAQ
Is AI cold email legal? In the US, yes. B2B cold email is legal under CAN-SPAM provided you use truthful subject lines, identify yourself, include a postal address, and honor opt-outs within 10 business days. Canada and the EU are consent-based and much stricter. AI-written doesn’t change any of it.
How many cold emails should I send per day? Per inbox: 10–20/day during a 2–4 week warm-up, then hold around 30–50/day. Scale by adding inboxes and secondary domains, never by pushing one inbox harder. Keep bounce rates under ~2–3% and spam complaints under 0.3%.
Can AI really book meetings on autopilot? The honest answer: AI reliably automates the research, the first drafts, the instant replies, and the calendar booking, which is most of the labor. It doesn’t automate having a relevant offer for a well-chosen list. With those, “a meeting booked while you slept” is a normal Tuesday. Without them, AI just automates being ignored.
What reply rate is good for AI cold email? For well-targeted, personalized B2B sequences, a 2–5% positive reply rate is a healthy benchmark, and highly niched campaigns with strong offers can beat that. If you’re under 1%, fix the list and offer before touching the copy.
Build the sending foundation this week, launch the 4-step sequence next week, and let the AI SDR take the night shift on replies.