The AI Marketing Playbook

Seven systems we actually deploy. Not opinions. Not "trends." The exact stacks, prompts, and SOPs we use for paying clients — written down here so you can build them yourself before we ever talk.

Most "AI marketing" content is one of two things: a vendor pitch dressed up as advice, or a 30-thousand-foot overview that doesn't tell you what to actually do on Monday morning. This is the opposite. Each chapter is a system you can build in 1–3 weeks with off-the-shelf tools and modest budget. If you build them all, you have an AI marketing engine that compounds.

A note before we start: tools are 10% of these systems. The other 90% is the messaging matrix, the data pipeline, the eval loop, the reply handling, and the CRM integration. We've watched companies spend $80K on Clay, Smartlead, and Apollo and produce zero pipeline because they skipped the 90%. Read the whole playbook before you buy anything.

---

System 1: The Cold Outbound Stack

Cold outbound, done well, is still the single highest-leverage growth motion for B2B companies under $50M ARR. Done poorly, it's also the fastest way to torch your domain reputation. Here's the stack that works.

Architecture

SOURCING (Apollo, LinkedIn Sales Nav, Clay)
    ↓
ENRICHMENT (Clay → 8–12 data providers → AI fallback)
    ↓
ICP SCORING (LLM-judged based on firmographic + behavioral signals)
    ↓
PERSONALIZATION (Claude, with 3-tier personalization budget)
    ↓
MULTI-CHANNEL SEND (Smartlead, Instantly, LinkedIn auto-tools)
    ↓
REPLY HANDLING (AI triage → human review → CRM sync)

The non-obvious choices

1. Sourcing should be intent-led, not just firmographic. Apollo and LinkedIn Sales Nav are tablestakes. What separates good outbound from great is layering intent signals on top: G2 visitor data, Bombora topic surges, Vector funded-rounds data, scraped job-posting signals from your competitors' customer pages. This is where the leverage compounds — sending to a CFO is fine; sending to a CFO whose company just posted three open Data Engineer roles last week is gold.

2. Enrichment should be a waterfall, not a single provider. Run waterfalls in Clay or Hooks: try Apollo first, then Clearbit, then Hunter, then Anymail Finder, then AI inference as last resort. Each step costs pennies but raises your match rate from ~60% to ~92%. The hit-rate difference is the difference between an outbound system that breaks even and one that 4×'s your spend.

3. ICP scoring should be done by an LLM, not by a rule engine. Rule engines say "VP+ at companies with 50–500 employees." LLMs read the company website and say "this is a Series B post-PMF B2B SaaS with a sales-led GTM and they just expanded into Europe" — which is what you actually wanted the rule to express. A simple ICP scoring prompt that returns {score: 0-1, reasoning: string, signals: string[]} works better than every static rule we've ever shipped.

4. Personalization should have a budget, not be uniform. Tier your list. The top 50 prospects per week get hand-written-feeling 5-line opens that reference something specific from their 10-K, their podcast, or their hiring page. The next 200 get LLM-personalized opens based on 2-3 enrichment signals. The next 2,000 get templated. Personalization budget should scale inversely with list quality — top-of-funnel quality plus top-of-personalization equals reply rates above 30%.

5. Use dedicated sending domains, always. A separate domain (like tryworkmail.com if your main is workmail.com) absorbs all the deliverability risk. Warm up over 4–6 weeks. Use Smartlead or Instantly's deliverability network. Sign up for Google Postmaster Tools. The day your primary domain shows up on a blocklist is the day every customer email stops landing — don't risk it.

Sample prompt: ICP scoring

You are an expert B2B sales analyst. Given the following information about a company,
score it 0.0–1.0 on fit for our ICP and explain your reasoning.

OUR ICP:
- B2B SaaS, $5M–$50M ARR
- Sales-led GTM with at least 5 SDRs
- Currently using HubSpot or Salesforce
- Marketing leader hired in last 18 months (indicates investment in growth)

COMPANY:
{company_description}
{recent_news}
{employee_count}
{tech_stack_signals}

Return JSON: {score: number, fit_reasons: string[], misfit_reasons: string[], confidence: "high"|"medium"|"low"}

What we expect at 90 days

A well-built outbound stack on a tier-1 ICP should produce:

• 25–40% reply rate (cold, not warm)
• 8–18% positive-reply rate
• 2–5% meeting-booked rate
• Pipeline coverage 2–3× your sourced-pipeline goal

If your numbers are below this range after 90 days, the messaging matrix is wrong — not the tooling.

---

System 2: Inbound Qualification Agent

You have a contact form. Someone submits it. What happens next?

For 90% of companies the answer is: an SDR sees it eight hours later, sends a "let's grab time" email, the lead has already booked with a competitor, end of story. You can do better in under two weeks of work.

The system

Every inbound submission triggers an AI agent that:

1. Enriches the lead in 4 seconds (Clay, Apollo, LinkedIn).
2. Reads the form fill + enriched data.
3. Scores the lead's ICP fit (see System 1's scoring prompt).
4. Drafts a personalized 3-sentence reply offering a specific time, a relevant case study, or a routing to a self-serve resource — depending on the score.
5. Routes to a human (if hot) or self-serves (if cold) automatically.

The whole loop runs in 30–60 seconds. The lead gets a thoughtful, in-context reply while still on your site.

The numbers

We've shipped this in 11 companies. Average results:

• 64% reduction in lead response time
• 2.3× lift in meetings booked from inbound
• 41% reduction in time SDRs spend on inbound triage

The SDR savings alone usually pay for the build within 6 weeks.

Common mistake

Don't auto-send the AI reply unsupervised in week one. Run shadow mode: the agent drafts, a human approves in Slack, the email sends. After two weeks of approving 95%+ of drafts, flip the switch and run autonomous. This single discipline prevents 100% of the embarrassing-AI-email stories you've seen on LinkedIn.

---

System 3: Programmatic SEO

Pages-at-scale. Done well, this is the highest-leverage SEO investment a B2B company can make. Done lazily, it's how you get a manual action from Google.

When this works

• You have two or more dimensions of variance that produce useful pages.
  Example: [Tool A] vs [Tool B] × hundreds of tool pairs. Or Best [Software Category] for [Industry]. Or How to do [Job] in [Tool].
• You have real underlying data or expertise to feed each page — not just templated puff.
• You're willing to prune ruthlessly when pages don't earn their seat.

How we build it

1. Build the data layer first. Either scrape, license, or assemble structured data (tool capabilities, industry attributes, integration matrices). This is the hard part. Skip it and you get garbage pages.
2. Write a template that uses Claude for prose and your data layer for facts. A typical template: H1 + intro from prose model, comparison table from data layer, FAQ section, CTA. The model writes; the data layer constrains.
3. Use an editor pass. A human reads every page before publish. With a good template, you can review 50 pages per hour.
4. Watch Search Console weekly. Pages that don't earn impressions in 60 days get deindexed (noindex) or removed. Don't let dead pages drag your domain.

What's different about this in 2026

Google has gotten better at detecting AI slop. The companies that win are the ones who use AI to scale the prose, not the fact-gathering — and who layer in unique data, screenshots, expert quotes, or genuinely useful comparisons that no other site has. Programmatic SEO + real-data layer + human editing = pages that rank. Programmatic SEO without those = a manual action.

---

System 4: Ad Creative Pipeline

Performance advertising has become a creative-throughput game. Your CPM and CTR aren't the bottleneck; how many concept variations you can ship per week is.

The pipeline

HOOKS LIBRARY (Notion) — 100+ tested headline/angle/promise variants
        ↓
STORYBOARD GENERATOR (Claude → static + video script outputs)
        ↓
PRODUCTION (Canva, Figma, Runway, UGC creators)
        ↓
LAUNCH (Meta, LinkedIn, Reddit, X)
        ↓
PERFORMANCE FEEDBACK (Meta API → Notion → next round)

The leverage

Most teams test 1–3 ad concepts per week. With this pipeline, you can test 12–20 — at a cost-per-creative that's a fraction of agency rates. The point isn't to replace your creative team; it's to feed them with structured raw material so the senior people stay on the high-leverage work.

Sample workflow

1. Marketing lead picks a winning hook from last week's data ("Our CFO finally gets it").
2. Claude takes the hook + product context + audience persona, generates 6 storyboard variations (3 static, 3 video).
3. Storyboard goes to Canva or a UGC creator network.
4. Variants ship same week. Performance data flows back to Notion. Next week, the system recommends which hook to evolve.

In 60 days you'll have a structured library of what works on each audience — which becomes a competitive moat your competition can't easily catch.

---

System 5: Lifecycle & Nurture

Cold outbound and inbound get most of the attention. Lifecycle is where the easy money is.

The frame

Every active relationship in your CRM is a stream of intent moments — opens, clicks, content downloads, demo bookings, account expansions, integration usage spikes, support tickets, plan changes. Each one is a chance to act. Most companies act on three of them. AI-powered lifecycle systems act on twenty.

What we build

• Behavioral triggers wired into your CRM or product analytics: trial day 3, integration installed, feature first-used, plan downgraded, NPS detractor, etc.
• AI-drafted, human-approved emails per trigger. Not "Hi {first_name}" templates — actual contextual messages that reference what the recipient just did.
• LinkedIn DM mirrors for high-value triggers, where appropriate.
• A reporting layer that shows which triggers convert and which to retire.

What changes

Lifecycle revenue moves from 5–10% of marketing-attributed revenue to 25–35%. That's not a typo. The reason is that most companies are running 6 generic lifecycle emails when they should be running 60 specific ones, and the AI editor is what makes 60 maintainable.

---

System 6: Content Engine

Most "AI content" looks like AI content. Here's how to build a content engine that doesn't.

The mistake everyone makes

They prompt Claude or GPT, paste the output into WordPress, hit publish. The result reads like AI, ranks for nothing, converts no one.

What works instead

TOPIC RESEARCH (Ahrefs, ClickFlow, ChatGPT search trends)
    ↓
BRIEF GENERATION (Claude — competitive analysis, outline, key claims)
    ↓
EXPERT INTERVIEW (15 min with internal SME — recorded)
    ↓
DRAFT (Claude with transcript + brief + previous-post style examples)
    ↓
HUMAN EDIT PASS (always — by a senior writer or marketer)
    ↓
PUBLISH + DISTRIBUTE (CMS + LinkedIn + Twitter + newsletter)

The recipe: AI handles the volume; humans handle the voice and the facts. The expert interview is the secret — it makes the post unique, since no other site has that specific 15-minute conversation.

The throughput

A 3-person team running this pipeline can ship 8–12 high-quality long-form posts per month with this approach. Without AI, it's 2–4. With only AI and no human in the loop, it's 30 posts that nobody reads.

---

System 7: Attribution & Reporting

You can't optimize what you can't measure. Most marketing teams are running on vibes. AI changes that.

What we build

• Pipeline-back attribution (not last-touch, not first-touch — weighted multi-touch with revenue tied back to each touch).
• Channel-level ROI dashboards that surface what's working, weekly.
• AI-summarized exec briefings — every Monday morning, an AI-written paragraph that summarizes last week's marketing performance and flags anomalies.
• Anomaly detection — sudden drop in form fills? Spike in a specific landing page bounce rate? The system pages you before you'd notice.

The dashboard you actually need

Stop building 47-tile dashboards nobody reads. You need 4 numbers:

1. Sourced pipeline this period (vs. target, vs. last period).
2. Influenced pipeline this period (where marketing touched but didn't source).
3. Channel mix this period (which channels produced).
4. Forward-looking signal (1 leading indicator for next period's pipeline).

If your dashboard answers those four questions in 10 seconds, it's good. If not, throw it out.

---

The compounding effect

Each of these systems is useful alone. Built together, they compound. Outbound (1) generates inbound flow (2) which feeds SEO pages (3) which produce ad-tested hooks (4) which inform lifecycle (5) which produces content topics (6) which feed attribution (7) which optimizes outbound. Six months in, the engine is running 24/7. Your team is operating the engine, not doing the work.

That's the goal. That's why we wrote this down.

---

Where to start

If you have to pick one: start with System 2 (Inbound Qualification Agent). It ships in 2 weeks, the ROI is immediate, and it builds your team's confidence to ship the bigger ones.

If you're outbound-led: start with System 1. Just budget for the messaging matrix — it's the part that ruins outbound projects when teams skip it.

If you have a content team already: start with System 6. The marginal cost of more good content is the lowest in your stack.

---

Want this built for you?

You can run this playbook yourself. Most of our paying clients did at some point — that's how they found us. They got 60% of the way, then realized it was someone's full-time job to maintain. That's where we come in.

Book a 30-minute strategy call. We'll tell you which system to build first, whether your team can run it without us, and whether Byte Dreams is even the right partner for you. We turn down about 1 in 3 conversations because the fit isn't right. That's a feature, not a bug.

— Drew Chimento, Founder, Byte Dreams