The Anti-Playbook, Part 3: The AI-Powered Outbound Stack
This is where most outbound frameworks wave their hands.
They mention “AI-powered personalization” and “intent signals” and move on, without ever telling you what the actual stack looks like, what each tool does, what the workflow is, or what to expect when the deliverability gods turn against you.
That hand-waving is exactly what gets demand gen leaders fired in month four — when the manual motion that worked beautifully in week two scales to 500 sends a day and the reply rate collapses because nobody set up DMARC correctly.
So this post is the part nobody explains. The actual stack, what each layer does, and the operational mechanics that make of break it.
The Two Phases of the Stack
The outbound motion has to operational phases, and they require fundamentally different infrastructure.
Phase 1 (Days 1-28): Manual sends from real inboxes. Volume is low, personalization is high, infrastructure is whatever the company already has. The marketing leader and the SDR each send 20 messages a day from their own work email. No automation, no sequencing tool, no domain warming. The constraint is human time, not deliverability.
Phase 2 (Days 28 onward): Scaled sends from warmed secondary domains. Volume scales to 200-500 sends per day. Sequencing is automated. Personalization is generated at scale using AI. Deliverability infrastructure becomes the foundation everything else sits on.
The mistake most demand gen leaders make is trying to skip Phase 1 and jump directly to Phase 2. The mistake most consultants make is treating Phase 1 as “not real outbound.”
Both are wrong. Phase 1 is the learning phase. Phase 2 is the scaling phase. Skip the learning and you scale the wrong thing — at 500 sends a day.
Phase 1: The Manual Stack
The Phase 1 stack is deliberately minimal. The whole point is to start producing conversations before infrastructure exists.
List building: Clay or Apollo, plus LinkedIn Sales Navigator. Clay is the better tool — it lets you enrich at scale with conditional waterfalls (try Apollo first, then Hunter, then ZoomInfo, fall back to manual). But for the first 200 accounts, even a free Sales Navigator search exported to a sheet is enough.
The enrichment data that actually matters: Recent funding round, recent leadership hire, tech stack signal, recent content the prospect published, mutual connection. The goal is having one specific, verifiable, recent fact about each account that the message can reference. If you can’t find that fact, the account isn’t ready for outreach yet. Leave it for later.
Message authoring: Claude of ChatGPT for variation, not generation. This distinction matters. The spine of every message is written by a human, because the spine requires judgement about what to say. AI rewrites the same message for 10 different prospect contexts in the time it takes to write one. The prompt that work: “Here’s the base message. Here’s prospect A’s specific context. Rewrite the first two sentences to reference that context naturally. Keep the rest identical.”
Sending: From real inboxes, no tooling. Marketing leader sends from their email, SDR from theirs. Volume per inbox stays under 30/day to avoid deliverability flags. Replies route back to the human who sent them — which means the person who wrote the message handles the reply. This is a feature, not a bug. You’ll learn ten time more from handling your own reply triage than from reading a report about it later.
Tracking: A shared Google Sheet, not a CRM. One row per send. Account, contact, message version, send date, reply status, notes. Setting up CRM tracking in week one is exactly the kind of premature infrastructure that delays the actual motion.
Phase 2: The Scaled Stack
This is where the operational complexity lives. Get this wrong and the whole motion collapses.
Domain and Inbox Infrastructure
Buy 2-3 secondary domains. Variants of the company name work. Never send cold outbound from the primary domain. If it gets flagged, the company’s regular email stops working — and now you’ve created a company-wide crisis instead of a marketing problem.
On each secondary domain, set up SPF, DKIM, and DMARC correctly. This is non-negotiable. A misconfigured DMARC record sends everything to spam regardless of how good your messages are. The single technical detail is the most common reason scaled outbound fails.
Create 3-5 inboxes per domain. Each inbox gets it own warming schedule and its own daily send limit.
The volume math: 3 domains x 4 inboxes x 30 sends/day = 360 sends/day at full scale. Need more capacity? Add domains, not more sends per inbox. Pushing existing inboxes harder is how you burn them.
Domain Warming — The Long Pole
Warming takes 4-6 weeks per domain to do properly. Tools like Instantly, Smartlead, or Mailwarm handle the mechanics — sending small volumes to a network of inboxes to generate engagement signals that train Google and Microsoft to trust the domain.
Here’s the uncomfortable 2026 reality: warming tools are degrading. Google and Microsoft have gotten significantly better at detecting reciprocal warming networks, and pre-warmed domains for sale are mostly worthless because they’re warmed via the same detectable patterns.
The best version of warming in 2026 is real engagement: send the new inbox’s first 50-100 emails to actual humans you know — investors, advisors, team members, friendly customers — and get them to reply naturally. It’s slow, but it produces a domain that actually performs.
Plan accordingly. If scaled outbound needs to be live by day 30, domain purchase and warming has to start on day 1, in parallel with the manual motion. Warming is the long pole in the tent.
Sequencing
Instantly or Smartlead are the two production-grade options at this volume. Smartlead has better infrastructure for serious senders; Instantly has the better UX for getting started. At early-stage scale, either works.
The sequence structure that performs: 3-4 messages over 12-14 days.
Message 1: the personalized opener
Message 2: soft follow-up referencing the original — short, no new ask
Message 3: a value drop — a relevant resource, a teardown, an insight specific to their situation
Message 4: the breakup
Counterintuitively, the breakup message has the highest reply rate of the entire sequence. People who ignored your four times respond when you signal you’re about to stop.
Cadence: weekday sends only, randomized timing within business hours, randomized intervals between messages. Anything that looks robotic is a deliverability flag.
The AI Personalization Layer
This is the part that earns the “AI-powered” label — and it’s doing exactly one job: generating human-quality personalization at scale.
At 360 sends a day, you can’t hand-write every opener. But a templated mail merge with {{firstname}} gets you a 0.5% reply rate and a burned domain. The AI layer is what sits between those two outcomes — taking the enriched data on each prospect and generating an opener that references something real and specific about them, at volume.
The spine of the message is still human-written. AI handles the variation. Same principle as Phase 1, just automated and scaled.
What It Costs
Here’s the full Phase 2 stack cost:
| Tool | Monthly Cost |
|---|---|
| Secondary domains | ~$50/year, one-time |
| Instantly or Smartlead | $97–$500/month |
| Clay (enrichment) | $150–$500/month |
| Sales Navigator (1–2 seats) | $100–$200/month |
| AI API (Claude or GPT) | $50–$200/month |
| Deliverability monitoring | $60–$100/month |
| Total | ~$500–$1,500/month |
Compare that to the cost of a single SDR — $80K-120K loaded — and the economics speak for themselves. The point of the AI stack isn’t that it replaces the SDR. It’s that one SDR plus the AI stack does the work of three SDRs without it.
Why “AI-Powered” Means Something Specific Here
A lot of demand gen frameworks use “AI-powered” as a buzzword. In this stack, AI is doing one specific job: generating human-quality personalization at scale. That’s it. The rest of the stack — list building, sequencing, sending, reply handling, deliverability — is conventional outbound infrastructure that’s existed for years.
The only durable competitive advantage in could outbound today is whether your personalization at scale is good enough to feel hand-written. Companies that nail it get reply rates of 5-10%. Companies that don’t get 0.5% and burn their domains in three months.
AI is the difference between those two outcomes. Used correctly, it’s the leverage point for the entire motion. Used as a buzzword, it’s the thing that destroys it.
What’s Next
The outbound stack creates conversations. But conversations create leads, and leads create the next question: how do you decide which ones sales actually talk to?
Part 4: Kill the MQL — why the lead scoring model most teams use is the wrong way to answer that question, and what to do instead
If you’ve ever watched a beautifully personalized manual motion collapse the moment it scaled, this post was for you. Tell me what I got wrong.
