Complete Blueprint — AI Pipeline Engine

What B2B Founders Are Using Instead of Hiring More SDRs

A new AI-driven outbound system (powered by Claude Sonnet 4.6) that turns cold prospects into booked conversations, without the usual cost, complexity, or guesswork.

4 Workflows
5 Prompt Kits
30 Days Rollout Plan
▶ Watch the AI Pipeline Engine — Free

See the system in action before implementing.

Intro / What This Is

This blueprint is a machine,
not a document.

Most outbound playbooks tell you what to do. This one shows you how it runs — automatically, at scale, without a rep touching every step. What you're holding is a complete operating system for B2B outbound, built on Claude Sonnet 4.6.

The Core Idea
Your pipeline shouldn't depend on rep bandwidth. It should depend on a system. Four tightly integrated workflows handle research, personalisation, routing, and pipeline movement — whether your team is online or not.
What's inside
  • Part 1: The Architecture — the 4-workflow system and how they connect
  • Part 2: The Prompt Library — exact prompt skeletons for each workflow
  • Part 3: Multi-Channel Orchestration — one state machine across all channels
  • Part 4: The 30-Day Rollout — day-by-day implementation checklist
  • Part 5: Governance and QA — controls that let you scale without risk
Who this is for

B2B founders and revenue leaders at $500K–$5M doing feast-or-famine revenue. If your pipeline depends on referrals, rep effort, and luck — this is your fix.

Part 1 / The Architecture

Four workflows.
One integrated system.

The system runs as four interdependent layers. Each feeds the next. Together they compress 30–60 minutes of rep research per account into under 5 minutes — and keep running whether your team is online or not.

WF 1
Research
Account + contact snapshots
→
WF 2
Personalisation
Hypothesis-driven copy
→
WF 3
Intent + Routing
Classify and route replies
→
WF 4
Pipeline Engine
CRM, tasks, sequences
Layer What It Does Trigger Owner
Research Builds structured account and contact snapshots from raw data New account; weekly rescan Ops / RevOps
Personalisation Generates hypothesis-driven copy: email, LinkedIn, calls New contact; sequence refresh Marketing / SDR
Intent and Routing Classifies every reply; routes to right rep with SLA Every inbound reply or DM Sales / Ops
Pipeline Engine Updates CRM, creates tasks, moves leads on signals alone State changes, timeouts RevOps

Workflow Detail

WF 1
Deep Account and Contact Research
TriggerNew account enters Tier 1 list, or new contact added
InputsCompany URLs, LinkedIn profiles, enrichment data, job postings, funding signals
OutputsJSON: company_summary, likely_initiatives (with why_now), risks, persona_context, personalisation_hooks (3–10)
Runs as batch job. Rescans Tier 1 accounts weekly to catch new trigger signals.
WF 2
Personalised Messaging and Adaptive Sequences
TriggerNew account researched; sequence refresh requested
InputsResearch snapshot + offer library + persona playbook
OutputsMessage Strategy object, email sequence (4–6 touches), LinkedIn DM flow, call opener + discovery questions
Sequences are state machines, not static cadences. Sonnet updates next touch based on engagement signals.
WF 3
Intent Detection and Reply Routing
TriggerAny inbound email reply or LinkedIn DM
InputsFull thread text + last outbound message + lead metadata (persona, stage, segment)
Outputsintent_label, confidence 0–1, urgency, recommended_next_action, owner_recommendation
Auto-acts only above confidence threshold. Below threshold: creates human task with suggested reply.
WF 4
Pipeline Movement and CRM Automation
TriggerState transitions, engagement signals, timeouts, performance thresholds
InputsLead state machine + engagement history + classifier output
OutputsCRM state updates, task creation, sequence pauses/advances, nurture enrollment, performance alerts
Meeting booked or HIGH_POSITIVE reply: all outbound on that contact pauses immediately.
Part 2 / The Prompt Library

Exact prompt skeletons.
Replace the brackets. Run it.

These are the exact prompt skeletons powering each workflow. Replace {{bracketed variables}} with your own data. Each prompt instructs Sonnet to return strict JSON — feed the output directly into the next layer.

Prompt 1 — Research Agent WF 1
System Prompt
You are a senior GTM research analyst. Turn noisy company and contact data into structured insight an SDR or AE can use immediately for outbound. Only write what a seller would reference on a call or email. Respond ONLY with valid JSON matching the schema below. No commentary.
Variables
{{raw_company_text}} website / careers / blog / press extracts
{{raw_contact_text}} LinkedIn summary, experience, recent posts
{{solution_capabilities}} short description of your product/service
{{icp_config}} vertical, size, core pains you solve
Output Schema
"company_summary": ["bullet 1", "bullet 2"], // max 5 "likely_initiatives": [{"initiative": "...", "why_now": "..."}], "risks_or_constraints": ["..."], "persona_context": "...", "personalization_hooks": ["hook 1", "hook 2"] // 3–10 hooks
Prompt 2 — Message Strategy WF 2
System Prompt
You are an outbound strategist designing hypothesis-driven campaigns to start high-quality sales conversations. Your outputs feed separate agents that render email, LinkedIn, and call scripts. Focus on clear hypotheses — not fluffy marketing language. Return valid JSON only.
Output Schema
"core_hypothesis": "...", "angles_to_test": ["efficiency", "revenue", "risk"], "objections_to_preempt": [{"objection": "...", "response": "..."}], "cta_ladder": ["soft CTA", "medium CTA", "meeting ask"]
Prompt 3 — Copy Renderer (Email) WF 2
System Prompt
You are a senior SDR writing concise, conversational outreach. Always reference at least one specific detail from the research snapshot. Avoid clichés. Sound like the same person across all channels. Keep email body under 120 words. Return valid JSON only.
Output Schema
"subject_lines": ["variant A", "variant B", "variant C"], "body": "...", // 50–120 words, one clear ask "ps": "...", // optional social proof line "angle_tag": "..." // which angle this email tests
Prompt 4 — Reply Classifier and Router WF 3
System Prompt
You classify and route inbound replies to outbound campaigns. Be exceptionally strict. Never hallucinate intent. If intent is ambiguous, mark as UNCERTAIN and escalate. Return valid JSON only. No commentary.
Intent Labels
HIGH_POSITIVE
Strong buying signal, ready to engage
MID_POSITIVE
Interested but needs nurture
NEUTRAL
No clear signal either way
UNCERTAIN
Ambiguous — escalate to human
SOFT_NEGATIVE
Not now, but not closed
HARD_NEGATIVE
Firm no — suppress
COMPLIANCE
Legal/unsubscribe — suppress permanently
SOFT_NEGATIVE
30-day nurture, lower frequency
Output Schema
"intent_label": "HIGH_POSITIVE", "confidence": 0.94, "urgency": "high", "recommended_next_action": "schedule_demo", "owner_recommendation": "AE_team"
Prompt 5 — Sequencing Orchestrator WF 4
System Prompt
You are a sequencing orchestrator. Given current prospect state, recent engagement events, and campaign goals, choose the next best touch — or decide to stop. Return valid JSON only.
Output Schema
"next_touch": { "channel": "email", // email | linkedin | call "angle": "risk", "timing_days": 3, "template_id": "break_up_v2" }, "reasoning_summary": "...", "should_continue": true
Pro Tip
Run Prompts 1 → 2 → 3 in sequence for each new account. Store output JSON from each step and pass it as input to the next. Prompts 4 and 5 run reactively on every inbound reply and engagement event.
Part 3 / Multi-Channel Orchestration

One state machine.
All channels.

All channels share one global lead state machine. One record per contact, regardless of how many channels are active. This prevents conflicting messages and ensures every action responds to what actually happened.

Default Sequence Structure

Day Email LinkedIn Trigger
Day 0 Email 1, Hypothesis A Profile visit + connection note Account enters sequence
Day 2–4 Email 2, Hypothesis B View or comment on post No reply to Email 1
Day 5–7 Light bump email DM if connected No reply; LI accepted
Day 10–14 Breakup / value email Final DM Still no engagement
Day 15+ Nurture drip (monthly) — Shift to slow-burn track

Channel Priority Rules

HIGH_POS
Meeting booked or HIGH_POSITIVE reply: Pause all channels immediately. Route to AE within SLA.
MID_POS
MID_POSITIVE reply: Pause email. Switch to LinkedIn DM. AE reviews within 4 hours.
SOFT_NEG
SOFT_NEGATIVE: Move to 30-day nurture. Lower frequency, don't hard-close the contact.
HARD_NEG
HARD_NEGATIVE / COMPLIANCE: Suppress across all channels. Log permanently. Never contact again.
NO ENG
No engagement after 5 touches: Enroll in slow-burn content drip. Sonnet drafts monthly value email.
Part 4 / The 30-Day Implementation Rollout

Follow this in order.
Each day builds on the last.

Don't skip to Week 3 without completing Week 1. The system requires foundations before automation.

Week 1 — Strategy, Scope and Foundations Days 1–7
Day
1–2
Define target state
Lock ICP, segments, personas. Start with 1 ICP, 1 geo, 2 personas max.
Day
3
Audit current stack
Document where leads live today. Identify enrichment gaps and CRM field needs.
Day
4
Define objects and states
Design state machine: New → Engaged → Qualified → Opportunity → Won/Lost/Recycle.
Day
5
Select pilot segment
50–200 accounts, 200–500 contacts. Lock success metrics before touching anything.
Day
6–7
Governance and guardrails
Set tone rules, compliance windows, human-approval triggers, prohibited phrases.
Week 2 — Research and Personalisation Engine Days 8–14
Day
8–9
Build Research workflow
Wire inputs: website URLs, LinkedIn, enrichment. Configure Research Agent prompt.
Day
10
Run research batch
Execute on pilot list. Spot-check 20–30 outputs with reps. Iterate prompt once.
Day
11–12
Build messaging prompts
Create Strategy prompt + channel renderers (email, LinkedIn, call script).
Day
13
Generate first sequences
4–6 touch email sequences + LinkedIn flows for each pilot persona.
Day
14
Freeze v1 message kits
Lock approved prompts into central Prompt Library. No changes without QA.
Week 3 — Intent, Routing and Controlled Go-Live Days 15–21
Day
15–16
Build classifier
Finalise intent taxonomy. Build classifier prompt. Map labels to routing rules.
Day
17
Wire CRM updates
State transitions, task creation, sequence pause logic. Set up full logging.
Day
18–19
Soft pilot launch
Go live for 20–50 accounts. Require human review for all first touches.
Day
20
QA and prompt tuning
Collect great/bad examples. Tweak prompts. Adjust human-approval thresholds.
Day
21
Expand pilot volume
Scale to full cohort. Keep approval rules active for 3–5 more days.
Week 4 — Scale, Optimisation and Handoff Days 22–30
Day
22–23
Multi-channel rules
Define email cadence, LinkedIn caps, channel-switching triggers.
Day
24
Roll out to more reps
Train team on reading snapshots, overriding copy, flagging bad outputs.
Day
25–26
Instrument performance
Dashboard: reply rate, positive-intent rate, meetings booked, time saved.
Day
27
Optimise sequences
Clone winning angles. Rewrite underperforming steps. Expand to adjacent segments.
Day
28–29
Cement SOPs
Define fully automated vs human-approved. Write onboarding runbook.
Day
30
Executive review
Present baseline vs results. Decide next segments. Roadmap deeper automation.
Part 5 / Governance, QA and Risk Controls

Scale fast.
Stay in control.

Scale without governance creates brand damage and compliance risk. These controls let you move fast while keeping a human in the loop where it matters.

Human-in-the-Loop
  • Review first 10 messages per rep per day for the first 2 weeks, then reduce
  • Auto-surface 5% of all sends weekly via random sampling for QA
  • Classifier confidence below threshold: create task with suggested reply, not auto-send
Brand Guardrails
  • Global style guide prompt: append to every Copy Renderer call
  • Prohibited phrases list baked into every prompt system instruction
  • Email body max 120 words, LinkedIn DM max 80 words, define 3–5 tone descriptors
Compliance
  • COMPLIANCE intent label: immediate cross-channel suppression, permanent log
  • Region-aware sending: CAN-SPAM for US, GDPR opt-out handling for EU
  • Complaint or unsubscribe rate exceeds 0.3%: auto-pause, require manual review
Performance Monitoring
  • Track: reply rate, positive-intent rate, meetings booked, cost per booked call
  • Track per rep: accounts researched per week, high-quality touches per day
  • Weekly Sonnet review: feed aggregate performance data, receive suggested optimisations
Automate what's repetitive. Keep humans on what's irreversible. A misclassified reply costs you 2 minutes. A suppressed compliance contact saves you a crisis. When in doubt, escalate.
Next Step

See exactly how this
works in your business.

The AI Pipeline Engine builds this acquisition architecture for B2B service businesses ready to stop running manual outreach and start running a system. Signal-based targeting. AI-personalised outreach. Automated follow-up sequences.

Watch the free breakdown first. Then, if it looks like a fit, book a 30-minute strategy call. On the call we'll identify exactly where your pipeline is breaking, show you how the system applies to your business, and give you a clear path to predictable client acquisition.

▶ Watch the AI Pipeline Engine — Free

No pitch. No pressure. Just a clear plan for what to do next.

© 2026 Dylan Power. All rights reserved. Confidential — Internal Distribution Only.