Part 3 — Turning Precision into Pipeline: The Build, The Message, The Lift – Step by Step

By /

Story + step-by-step manual. Continued from Parts [1] and [2].

Act I — Monday: Give the model a brain (RAG), not a hunch

Story: We started the week with a CRM full of names and no direction. The promise to the team: “By Friday, you’ll know exactly who to call and what to say.”

What you do (how-to):

  1. Create your rulebook doc (the “brain”).
    Put this in a single document you control (Google Doc/Notion/MD file):
    • ICP tiers: HIGH (post-sale owners), MEDIUM (adjacent influencers), LOW (upstream/plant/dealer).
    • Scoring rubric: +100 Golden Account, 100/50/0 for title/function, +intent boosts.
    • 20 labeled examples (title → score + 1-line rationale).
  2. Embed it for Retrieval-Augmented Generation (RAG).
    • Chunk the doc (≈500–1,000 tokens).
    • Store embeddings in a vector DB (Supabase/Pinecone/Weaviate).
    • Retrieval params: top_k = 3–5, cosine similarity.
  3. Define the classifier contract.
    The model must return JSON only:

Why this matters: RAG stops guesswork. Every decision pulls the same rulebook before the model scores a contact—consistent, updatable, explainable.

Act II — Tuesday: Wire the loop (HubSpot → RAG → ChatGPT → HubSpot)

Story: By day two, we connected the plumbing so AEs would stop guessing and start dialing.

What you do (how-to):

  1. Add three HubSpot properties
    • is_golden_account (Company · Checkbox)
    • pqd_job_title_score (Contact · Number)
    • pqd_lead_score (Contact · Number)
  2. Create a Contact-based workflow (triggers):
    • On Contact Created OR Job Title changes OR Company association changes.
  3. Custom code action (Node/Python) — skeleton

4. Auto-action by score

  • If pqd_lead_score ≥ 150 → auto-enroll in Outreach sequence.
  • If 100–149 → add to nurture task list.
  • Else → park with a research task.

Why this matters: The loop makes the decision in real time and hands reps a prioritized list. No pep talk required.

Act III — Wednesday: Arm the spear (the message)

Story: Mid-week, the first calls sounded different because the message fit the persona.

What you do (how-to):

  1. Fast persona matrix → openers PersonaOne-liner openerAsset/CTAWarranty/Field Quality (HIGH)“We flag in-fleet issues in days, not months, shrink campaign scope, and verify fixes VIN-by-VIN.”After-Sales Quality Report + 20-min benchmark inviteOTA/Connected-Vehicle (HIGH)“Catch post-release regressions and prove post-fix stability within days.”Benchmark their last OTA rolloutRegulatory/PRO (HIGH)“Gain lead-time and precision to make faster recall vs. campaign calls.”Lead-time/campaign-size case studySDV/Software Quality (MEDIUM)“Spot <0.01% silent failures after firmware changes.”Short tech deep-diveSupplier/Launch/Plant (LOW)Nurture onlyTopical newsletter
  2. Email + LinkedIn templates (drop-in)
    • Benchmark invite (build-vs-buy safe):
      “Bring one recent field issue or OTA release. In 20 minutes we’ll show how our anomaly signatures would have surfaced it earlier and reduced the campaign footprint.”
    • Connect (≤300 chars):
      “We analyze ~5k recalls & 30k complaints to find post-sale issues in days, not months. Worth a 15-min walkthrough on shrinking campaign scope?”
  3. Make the report your spear tip
    Use your After-Sales Quality study as proof (grounded in recall/complaint datasets). Always link it in High-fit outreach.

Why this matters: When title and talk track align, responses jump—and “we build in-house” becomes a comparison, not a brush-off.

Act IV — Thursday/Friday: Go-live, measure, tighten

Story: By Friday, they booked the first cold meetings ever. Not luck—tight feedback.

What you do (how-to):

  1. Pilot with 25–50 contacts
    • Manually spot-check score + rationale.
    • Fix obvious misses (e.g., “Quality Engineer” → plant vs. field disambiguation) by adding examples to the rulebook.
  2. Instrument the lift
    • Leading: % High-fit in active queue, connect rate, reply rate, booked meetings.
    • Lagging: win rate by fit tier, cycle time to Stage 2, pilot rate.
  3. Close the loop (weekly)
    • Export overrides (“rep disagreed”) → add good ones to few-shot examples.
    • Re-embed the rulebook (RAG) — no retraining required.

Why this matters: Tiny weekly updates keep the classifier aligned with reality, so precision compounds.

System prompt (classifier):

RAG retrieval tip:
Query with: "{Title}" + "{Company}" + ("warranty" OR "field quality" OR "OTA" OR "connected vehicle") to bias toward the right snippets.

HubSpot properties to create:

  • is_golden_account (Company · Checkbox)
  • pqd_job_title_score (Contact · Number)
  • pqd_lead_score (Contact · Number)
  • (optional) pqd_rationale (Contact · Single-line text)

Troubleshooting (fast)

  • Too many “Quality Engineer = High” → Add examples that mark Supplier/Launch/Plant QE = 50 or 0.
  • Hallucinated rationales → Lower temperature to 0.1 and ensure only top-K snippets are injected.
  • AEs still not prospecting → Auto-enroll High-fit in Outreach, and put a daily Orum call block on the calendar.
  • Bounce/relevance issues → Stop raw dumps; enrich with intent (report downloads, webinar attendees) before scoring.

The Lift (what success looked like)

  • Week 1: cleaner lists, High-fit personas at the top, AEs actually prospecting.
  • Weeks 2–3: first outbound meetings ever with the right OEM stakeholders.
  • Ongoing: “We’re building in-house” → benchmark invitation instead of a dead end.

That’s the whole arc: codify your ICP → retrieve it on every decision → return score + why → act automatically.
When the brain is right, the pipeline follows.

Related Posts

Scroll to Top