How to Write AI Cold Emails That Get Replies (2026)

The failure is usually visible before the email is opened. Instantly’s benchmark roundup, updated January 2026, reports that 69% of recipients mark an email as spam based on the subject line alone, and that personalized subject lines lift open rates by about 50% (Instantly cold email statistics). Most AI-generated outreach fails both tests at once: a subject line like “quick question about your stack” plus an opener like “I came across your brand and love what you are doing.” Nothing in that email is false — and nothing in it is about the recipient, so nothing earns the open or the reply.

That points to a division of labor that works. Large language models like Claude and ChatGPT are excellent at the research-and-variation half of cold email: they can read a prospect’s public activity or a company’s funding news and turn it into three candidate angles in seconds, and they never tire of rewriting. They are unreliable at the other half: knowing what is true about your prospect, choosing the one proof point that will land, and writing a final line that asks for something specific. When a model does not know something, it tends to fill the gap with a plausible-sounding company fact — the exact failure that makes recipients treat an email as mass-produced.

So set the boundary before you write anything: the AI researches and drafts candidates; a human verifies every fact, keeps or cuts the angle, and owns the CTA. The workflow in the next section is built on that split, and every prompt below forces the model to mark its inferences instead of smuggling them in as facts.

The workflow has six steps and is designed for one person with a spreadsheet plus a sending tool. Each step feeds the next, and the two places where quality is decided — the research input and the human edit — are deliberately not automated.

Here is the workflow applied to a fictional prospect so you can see where quality enters and where it leaks. (example) Maya Chen, Head of E-commerce at Loopstock, a 200-person DTC apparel brand that launched a subscription box two months ago and is hiring for three retention roles. Your offer: an AI-powered email flow that recovers subscription cancellations before the billing date.

Draft A — what the model writes with no research input: “Hi Maya, I came across Loopstock and love what you are doing. We help DTC brands reduce churn with AI-powered email. Would you be open to a quick call this week?” This is the email every spam filter has already seen fifty times: a generic compliment, a vague capability, and a CTA that asks for a meeting instead of promising an outcome.

Draft B — what the research pass changes. Subject: “your 3 retention hires + subscription churn.” Body: “Maya — you posted three retention roles two weeks after the Loopstock Box launch, which says churn is the number on your mind. We built a flow that intercepts cancel-intent signals before the billing date and recovered 18% of would-be cancellations for [name your customer] last quarter — case study attached. Worth 20 minutes to see if it maps to your numbers?” The signal is specific, the capability maps to it, and the CTA names an outcome. Two rules applied: the 18% figure is a placeholder you must replace with your own verifiable result, and the case study link must exist before send.

Draft C — the same prospect on a bad day: the model found no signal, so it invented one (“saw your recent expansion”), which a five-second check would have caught. This is the draft you delete. If the research pass turns up nothing usable, the prospect leaves the list — that is the 5.8%-versus-2.1% list discipline from Section 1 working as intended.

Writing is half the reply; the other half is showing up again. Woodpecker’s roundup reports that 58% of all replies come from the first email, 42% come from follow-ups, and yet 48% of sales reps never send a follow-up at all. Adding a single follow-up lifts total replies by about 66%, and the first follow-up is often the highest-performing step in the sequence at an 8.4% reply rate. Campaigns with three to five follow-up steps average 8.3% replies versus 4.1% for sequences with none, and the optimal sequence length is four to seven touchpoints. The one phrase that measurably hurts: “just checking in,” which the same data links to 14% fewer meetings booked.

Sequencing rules we use: every follow-up adds information or proof instead of repeating the ask; touchpoint three offers something concrete (a teardown, a template, a benchmark relevant to their signal); and the final touchpoint is an explicit close — “if this is not a priority, say so and I will stop writing” — because unsubscribes and spam complaints are the metric that actually hurts your domain.

Deliverability is where AI-written campaigns die as a group. Google’s bulk-sender requirements, in force since February 2024, apply once you send more than 5,000 messages a day to Gmail addresses: authenticate with SPF, DKIM, and DMARC, offer one-click unsubscribe, and keep the spam-complaint rate below 0.3% — Google recommends under 0.1% — or messages start getting rejected (Google bulk sender requirements). Recipients marking an identical AI template as spam is precisely what trips that threshold.

The practical implications are boring and mandatory: warm new domains slowly, verify addresses before send to keep bounce rates low, and use a platform that handles the infrastructure. Reply.io’s pricing page states that email warmup is included with every mailbox on its $69/mo per-account plan with unlimited sends to active contacts; Instantly’s pricing page lists unlimited email accounts and warmup from its $97/mo Hypergrowth tier (125,000 emails monthly) up to its $358/mo Light Speed tier at 500,000 emails monthly — both fetched 7 September 2026, and entry tiers start lower. The warmup and mailbox limits are what you are actually paying for. More context in our best AI email marketing tools ranking.

Q: Does Google penalize AI-written cold email? A: Google’s bulk-sender rules do not scan for AI authorship. They enforce authentication (SPF/DKIM/DMARC), one-click unsubscribe, and a spam-complaint rate below 0.3% for senders above 5,000 messages a day (Google bulk sender requirements). The indirect risk is real, though: generic AI copy earns low engagement and more “spam” clicks, which is exactly what pushes a domain over the complaint threshold. Write for relevance first; deliverability follows engagement.

Q: Which tool should I use to send AI cold emails? A: Most teams choose between Instantly (sequence automation plus an AI reply agent in its Unibox), Smartlead (tiers from $59/mo with clear send limits), and Reply.io (unlimited sends with warmup on each mailbox, plus a built-in contact database). Our Instantly vs Smartlead vs Lemlist comparison walks through the trade-offs with verified pricing. All three automate SPF/DKIM/DMARC setup and warmup, so pick on sending volume and CRM integration rather than headline features.

Q: How many AI cold emails should I send per day? A: Start smaller than feels productive: 10–30 per mailbox per day on a newly warmed domain, then scale as replies and complaints stabilize. This is operational advice, not a benchmark — the real guardrails are Google’s: below 5,000 messages a day you avoid bulk-sender classification, and your Postmaster Tools spam rate should stay well under 0.3%. Watch that rate weekly; it moves before your reply rate does.

Q: Can I personalize every AI cold email without it taking all day? A: Yes, if you keep lists small and research templated. The per-prospect research pass in Section 3 takes two to three minutes per contact, and enrichment tools (like Reply.io’s live database) supply the signals to feed it. Depth is the highest-leverage factor in the data — 17–18% replies with advanced personalization versus 7–9% basic — but only truthful signals count. A model inventing a “signal” is worse than none, so spot-check roughly one in ten drafts against the source.

Three actions, in order of impact. First, cut your next campaign to 100–200 verified prospects and spend the saved time on the per-prospect research pass — list quality is the biggest multiplier in the data above. Second, run the drafting prompts in Section 3 and enforce the 80-word, one-CTA edit; if a draft survives without a specific signal in the first line, delete it. Third, put the sequence in place before you send: four to seven touches, a “just checking in” ban, and an explicit close at the end. Then watch the two numbers that decide everything — reply rate and the spam-complaint rate in Postmaster — and let them tell you whether the problem is your copy or your list.