AI Outreach Sequence in 4 Emails: Prompts, Timing, and Target Benchmarks

Cold B2B email averages 1–3% reply rate. A four-email outreach sequence fully generated by an LLM, with quality personalization, can reach a double-digit reply rate and demo-call conversion in the 3-5% range. This article covers the framework: prompts for generating each email, send timing, an approach to A/B testing subject lines, and target benchmarks for each step.

Why Standard Outreach Templates Don’t Work

A typical sales team template: recipient’s name, company name, one sentence about their “pain” from the playbook, a paragraph about the product, a CTA for a call. This format goes out to thousands of people with minimal variation.

Three reasons it fails:

The template is obvious. The recipient opens the email and recognizes a mass blast in 3 seconds. “I noticed that [Company] is growing fast” doesn’t feel personal — it feels like a merge tag.

No research context. The template ignores recent company news, the recipient’s LinkedIn posts, their tech stack, recent job postings. Everything that distinguishes one recipient from another collapses into {first_name} and {company}.

Identical follow-ups. The second and third emails repeat the first with “just checking in” or “wanted to bump this to the top of your inbox” tacked on. No new information. No reason to reply.

LLMs change the approach: each email is generated individually from data about the recipient. But “individually” doesn’t mean “chaotically.” You still need a fixed sequence structure — AI personalizes the content within it.

The 4-Email Outreach Sequence Architecture

Four emails is the sweet spot — both three and seven underperform it. In practice, after the fourth email, reply rates on subsequent messages drop below 0.5% while unsubscribe rates climb; four emails typically capture the large majority of all replies in a sequence.

#	Day	Email Type	Goal	Target Open Rate	Target Reply Rate
1	0	Research-Based Opener	Show the sender did their homework	50–60%	3–5%
2	3	Value-First Follow-up	Deliver useful content without asking for anything	40–50%	3–4%
3	7	Social Proof + Specifics	Case study with numbers from a similar industry	35–45%	2–3%
4	12	Breakup + Low-Friction CTA	Final email with a low-barrier ask	45–55%	3–5%

Combined: the target is 11–17% cumulative reply rate with personalization at the research level, not merge-tags.

Personalization Data: What to Collect

Before generating emails, the AI needs structured data about each recipient. Minimum required:

- Name, title, company
- Company size (headcount)
- Industry / vertical
- Recipient's last 2–3 LinkedIn posts
- Most recent company press release or news item (within 90 days)
- Tech stack (from BuiltWith, Wappalyzer, job postings)
- 1–2 competitors of the recipient's company

Data collection is automated via Apollo, Clay, or PhantomBuster. The LLM gets a JSON object with these fields and generates the email. Without recipient data, generation is pointless — AI without context writes the same templates, just longer.

Email 1: Research-Based Opener (Day 0)

The first email has one job: get read to the end, not sell or book a call outright. Showing that the sender did their homework is enough for a first touch.

Generation Prompt

Role: You are a senior sales development representative with 8+ years of experience
in B2B outreach. You are writing a cold first email.

Recipient data:
{recipient_data_json}

Rules:
1. First sentence — a specific fact about the recipient or their company.
   NOT a compliment. A fact from the data.
2. Second sentence connects this fact to the problem
   the product solves. No product name.
3. Third sentence — one specific number (metric, result,
   cost saving) relevant to their industry.
4. CTA: a question answerable with "yes" or "no."
   NOT a request for a call.
5. Maximum 65 words in the email body.
6. Tone: direct, respectful, no exclamation marks.
7. No "Hope this finds you well" or equivalents.
8. Language: English.

Output format:
Subject: [subject line]
Body: [email body]

Generated Email Example

Subject: re: your Kubernetes migration

Body:

Your team posted 3 SRE positions last month, all mentioning Kubernetes migration from on-prem. Companies at this stage typically see 40% of incidents from misconfigured resource limits in the first 6 months.

We reduced post-migration incidents by 62% for a Series B fintech with a similar stack.

Worth a 2-minute look at how?

A/B Subject Line Test (Email 1): patterns worth comparing

Four subject line patterns for a research-based opener worth running through an A/B test at the start of a campaign:

Pattern	Example	Logic	Usually stronger on
re: prefix	re: your Kubernetes migration	Reads as a continuation of an existing thread	Open rate
Topical question	Question about your SRE hiring	Sparks curiosity, invites a reply	Reply rate
{company} + metric	{company} + Kubernetes incident reduction	Recognizable but reads close to a merge-tag	Neither — usually middling
Result number in subject	62% fewer incidents post-migration	Reads like a marketing blast	Usually the weakest

In practice, the re: variant tends to lead on open rate specifically because of the “continuing thread” effect, while the question variant tends to lead on reply rate because it directly invites a response. A reasonable approach: split the two stronger patterns 50/50 on the main send and drop the two weaker ones.

Ethics note: “re:” in the subject of a first email is contested. Some recipients find it deceptive. In regulated industries — finance, healthcare — avoid it.

Target Benchmarks for Email 1

• Open rate: 50–60%
• Reply rate: 3–5%
• Positive reply rate: 2–3%
• Bounce rate: up to 3–4% (higher signals a list quality problem)
• Unsubscribe rate: up to 0.5%

Email 2: Value-First Follow-up (Day 3)

The recipient opened the first email and didn’t reply. Or didn’t open it at all. The second email doesn’t reference the first — it delivers value with no expectation of a response.

Generation Prompt

Role: You are the same SDR. You are writing the second email in the outreach sequence.
The recipient did NOT reply to the first email.

Recipient data:
{recipient_data_json}

First email context:
{email_1_body}

Rules:
1. Do NOT mention the first email. No "following up" or
   "I wrote to you last week."
2. Share one specific insight, framework, or checklist
   that's useful to the recipient. It should connect to the problem
   from the first email.
3. The insight must be actionable — the recipient can apply it
   today, without the product.
4. Last sentence: a soft CTA in the format
   "If useful, I can send [something specific]."
5. Maximum 80 words.
6. Tone: helpful colleague, not a salesperson.

Output format:
Subject: [subject line]
Body: [email body]

Generated Email Example

Subject: K8s resource limits checklist

Body:

After working with 30+ teams on Kubernetes migrations, one pattern repeats: resource limits set during staging never get updated for production traffic. The result is OOMKill cascades under load.

Three-point check that catches 80% of these cases:

1. Compare P95 memory usage in staging vs actual production for each service
2. Set requests at P50, limits at P99 + 20% buffer
3. Add VPA recommendations as a CI gate before deploy

Want the full 12-point checklist as a PDF?

A/B Subject Line Test (Email 2): what usually loses

Pattern	Example	Typical result
Specific technical term	K8s resource limits checklist	Strongest — reads as substantive, not a blast
Problem description	Preventing OOMKill after migration	Middling result
Merge-tag in the subject	Quick checklist for {first_name}	Consistently weaker — recipients read {first_name} in a subject line as an automation signal

Personalizing directly in the subject line with a merge tag is a pattern worth excluding from tests — recipients have long since learned to spot curly braces in a subject line as an automated-outreach signal. A specific technical term nearly always performs better.

Target Benchmarks for Email 2

• Open rate: 40–50%
• Reply rate: 3–4%
• Positive reply rate: ~2%
• Unsubscribe rate: up to 0.6%

Email 3: Social Proof + Specifics (Day 7)

A week has passed. Two emails, no reply. The third email does what people trust more than any promise: shows results at a company like theirs.

Generation Prompt

Role: Same SDR. Third email in the sequence.

Recipient data:
{recipient_data_json}

Previous email context:
{email_1_body}
{email_2_body}

Case study bank:
{case_studies_json}

Rules:
1. Select from the case study bank the one that most closely matches
   the recipient's industry and company size.
2. Structure: one sentence about the company from the case study (no name,
   only a description: "Series B fintech, 120 engineers"),
   one sentence about the problem, one sentence about the outcome
   with specific numbers.
3. Last sentence: a question — "Worth seeing exactly how they did it?"
4. Maximum 70 words.
5. No opening phrases. Start directly with the case.

Output format:
Subject: [subject line]
Body: [email body]

Generated Email Example

Subject: 62% fewer incidents, 120-engineer team

Body:

A Series B fintech with 120 engineers migrated 40 microservices to Kubernetes over 4 months. Post-migration incident volume spiked 3x in the first quarter. After implementing automated resource tuning and pre-deploy validation, incidents dropped 62% and mean time to recovery went from 47 minutes to 11.

Their stack is similar to what your job posts describe.

Worth 15 minutes to walk through what they did?

A/B Subject Line Test (Email 3): avoid the words “case study”

Pattern	Example	Typical result
Specific result number	62% fewer incidents, 120-engineer team	Strong open rate
Storytelling format	How a fintech fixed post-migration chaos	Often the best reply rate — reads as a story, not a pitch
Explicit “case study”	Case study: K8s migration done right	Consistently the weakest

“Case study” in the subject line tends to underperform noticeably on open rate: recipients read it as marketing content and skip it. A specific number or a storytelling format both outperform it.

Target Benchmarks for Email 3

• Open rate: 35–45%
• Reply rate: 2–3%
• Positive reply rate: ~1.5%
• Unsubscribe rate: up to 0.8%

Email 4: Breakup + Low-Friction CTA (Day 12)

The final email. Two principles: be upfront that this is the last message, and drop the barrier to responding as low as possible.

Generation Prompt

Role: Same SDR. Fourth and final email.

Recipient data:
{recipient_data_json}

Previous email context:
{email_1_body}
{email_2_body}
{email_3_body}

Rules:
1. First sentence: "Last note on this."
2. One sentence: a brief summary of what was offered
   (NOT a recap of all emails — the essence in one phrase).
3. CTA: three reply options, one word each. Example:
   "1 — interested, let's talk"
   "2 — not now, circle back next quarter"
   "3 — not relevant"
4. Maximum 50 words.
5. Tone: respectful, no pressure, no guilt-tripping.

Output format:
Subject: [subject line]
Body: [email body]

Generated Email Example

Subject: closing the loop

Body:

Last note on this. We help engineering teams reduce post-migration Kubernetes incidents through automated resource tuning. Worked with 30+ teams in the last year.

If it helps to keep this short: 1 — worth a quick call 2 — not now, circle back in Q3 3 — not relevant

Either way, good luck with the migration.

A/B Subject Line Test (Email 4)

Pattern	Example	Typical result
Neutral closure	closing the loop	Good balance of open rate and reply rate
Personal note	Last note from me	Middling result
Provocative question	Should I stop reaching out?	High open rate from curiosity, but often a weaker reply rate

The provocative question usually opens well out of curiosity but underperforms on reply rate — the phrasing gets opens but creates awkwardness that discourages a response. A neutral tone that implies closure without manipulation tends to hold up better.

The breakup email tends to deliver one of the highest reply rates in the sequence overall — the deadline effect and the low-friction reply format (a single character instead of a full response) lower the barrier for the recipient. Among those who do reply to a breakup email, a meaningful share typically pick the call option rather than just closing the loop.

Target Benchmarks for Email 4

• Open rate: 45–55%
• Reply rate: 3–5%
• Positive reply rate (call option): ~1.5–2%
• Deferred interest (later option): ~1.5–2%
• Unsubscribe rate: up to 0.3%

Target Sequence Funnel

To sanity-check the economics before launch, it helps to run the funnel math on a round number of contacts using the target benchmarks from the sections above. Example at 1,000 contacts:

Metric	Target
Contacts in campaign	1,000
Valid emails (after bounces, typically 3-4%)	~960-970
Cumulative unique open rate	65-75%
Cumulative reply rate	10-15%
Contact-to-demo conversion	3-5%
Demo calls per 1,000 contacts	roughly 30-50

For context: a template-based sequence on the same audience typically delivers a cumulative reply rate of 2-3% and contact-to-demo conversion under 1% — a multiple-fold gap, not a few percentage points.

Typical Reply Distribution Across Emails

Email 1 (Day 0):  ████████████████░░░░░  largest share of replies
Email 2 (Day 3):  ██████████░░░░░░░░░░░  dip — warm-up email
Email 3 (Day 7):  ██████░░░░░░░░░░░░░░░  lowest — warm-up email
Email 4 (Day 12): ███████████████░░░░░░  rises again — breakup effect

This illustrates the typical shape of the distribution, not a specific campaign’s measured percentages: email 1 captures the most replies from fresh contact, email 4 from the breakup effect, and emails 2 and 3 warm the relationship in between.

What Makes AI Generation Better Than Templates

The difference isn’t “prettier writing.” It comes down to three things:

Personalization at the research level, not merge-tag level. AI analyzes LinkedIn posts, company news, and tech stack, then writes an opening sentence that couldn’t go to any other recipient. Not {company} is growing — a specific reference to their Kubernetes migration, visible in job postings.

Sequence consistency. Each email accounts for the content of all previous ones. No repetition, no contradictions. The AI gets the full prior context and generates a follow-up that adds something new.

Scalable variation. Every contact gets a unique first email rather than one of a handful of templates, while every email still follows the same A/B-tested structure. An SDR can’t write thousands of unique research-based intros by hand — an LLM generates them in seconds per contact.

Send Timing: When to Send Each Email

Three interval configurations worth keeping in mind when choosing timing:

Configuration	Intervals (days)	Trade-off
Aggressive	0, 1, 3, 6	Gets replies faster, but higher unsubscribe rate — recipients get annoyed by frequency
Balanced	0, 3, 7, 12	Usually the best balance of reply rate and unsubscribe rate
Relaxed	0, 5, 12, 21	Less annoyance, but recipients tend to forget the context of earlier emails

The Balanced configuration is a reasonable starting point for most B2B audiences; more aggressive intervals are worth testing only on audiences with a short decision cycle.

Time of day: Tuesday–Thursday, 8:00–9:30 AM recipient local time, typically produces a noticeably higher open rate than sends after lunch.

Generation Cost and ROI

Cost of generating 4 emails per recipient via Claude Sonnet 4.6 ($3/$15 per 1M tokens):

Component	Tokens	Cost
Input (prompt + data)	~1,200 per email	$0.0036 / email
Output (generated text)	~150 per email	$0.0023 / email
4 emails per contact	~5,400	~$0.024
Per 1,000 contacts	~5.4M	~$24

Generation costs roughly two orders of magnitude less than the equivalent volume of manual SDR personalization. Data collection via Apollo/Clay adds another rough $30–50 per 1,000 contacts. Even including data collection, generation stays a minor line item in the campaign budget — the real cost of cold outreach has always been a rep’s time, not the words in the email.

Limitations and What Doesn’t Work

AI doesn’t replace strategy. If the ICP is wrong, AI generation just speeds up sending emails to the wrong people. Garbage in, garbage out. List quality matters more than copy quality.

Hallucinations in personalization. In 3–4% of cases, AI misreads the recipient data. Example: it references a “Kubernetes migration” based on an SRE job posting when the company is actually moving to ECS. Fix: human review of a 10–15% sample before sending.

Spam filters. Mass-sending unique emails from one domain looks suspicious. Domain warm-up is required (minimum 2 weeks), account rotation, and a cap of 50–80 emails per day per sending account.

Doesn’t work without data. A prompt without {recipient_data_json} generates the same boilerplate. The difference between AI outreach and template outreach lives in the data, not the model.

More on personalizing cold emails with AI is in the AI cold outreach personalization article.

AI Outreach Sequence Launch Checklist

1. Define ICP and build a list of 500+ contacts with enriched data
2. Prepare prompts for 4 emails (adapt the examples from this article for your product)
3. Generate emails for the first 100 contacts; human-review 15–20 of them
4. Set up A/B subject line tests (minimum 2 variants per email, 150+ contacts per variant)
5. Warm up the domain and configure account rotation
6. Launch to 150 contacts, collect metrics over 2 weeks
7. Scale to the full list using the winning subject lines
8. Monitor unsubscribe rate: if it exceeds 2% on any email, revisit the copy or frequency

---

Need help building AI-personalized outreach sequences? I help startups build marketing systems at belov.marketing and develop AI solutions at belov.works.

FAQ

How do you handle recipients who reply with “not interested” after the first email, without burning the relationship for future outreach?

Acknowledge, remove from the current sequence, and log the reason. Reply with a single sentence: “Understood — I’ll take you off this list. If the timing ever changes, feel free to reach out.” Do not attempt to reframe the offer or ask qualifying questions. Set a 90-day suppression flag in your CRM, not a permanent “do not contact” unless they explicitly requested that. Markets shift and roles change — a contact who is not the right person today may be the decision-maker in two quarters. The goal is to leave a professional impression, not to extract a meeting from someone who has clearly declined.

What recipient data has the highest impact on first-email open rate versus reply rate?

Open rate is driven by subject line, and the subject line data that performs best is role-specific or company-specific information that makes the recipient curious about what you know — a job posting reference, a recent product launch, a specific technology in their stack. Reply rate is driven by email body personalization, where the most impactful data is a recent content signal (LinkedIn post, podcast appearance, conference talk) that lets you reference something the person cares about in the first sentence. Prioritize collecting content signals for reply optimization; prioritize collecting specific triggering events (hiring spree, funding round, tool migration) for open-rate optimization.

How do you calibrate the sequence for an audience with long decision-making cycles, such as enterprise procurement?

Extend the timing while compressing the ask. For enterprise contacts, shift the intervals to 0, 7, 18, 30 days rather than 0, 3, 7, 12. Replace the direct demo ask in email 1 with a thought-leadership piece relevant to their function. Shift email 3’s social proof to a named enterprise customer story rather than a size-matched example. In email 4, the breakup CTA should offer a “share with your team” or “send to the right person” option alongside the standard call and “not now” options — enterprise deals rarely die; they stall waiting to reach the right stakeholder.