Lifecycle Management

Replacing one global scoring model with portfolio-aware qualification

One global scoring model treated two different buyer portfolios as one. Rebuilt scoring as independent fit-and-behavior models per portfolio, then replaced the monolithic lifecycle workflow with one testable workflow per stage.

Environment

High-growth B2B SaaS (multi-product portfolio)

My role

Owned the scoring architecture and lifecycle redesign end to end — model design, behavioral mechanics, decay and eligibility logic, and the HubSpot lifecycle rebuild.

+25%

increase in qualified MQLs

400K

existing contacts reconsidered, surfacing demand already in the database

The problem

One number was answering two different questions: fit and intent, collapsed into a single score, so an engaged poor-fit lead looked identical to a well-matched buyer. Repeat behavior earned one credit no matter how often it recurred, and with no decay, a two-year-old click still counted as intent.

Context

Marketing qualified leads for two product portfolios through one global scoring model — easy to manage, structurally wrong, and inflating lead volume with inconsistent MQL quality. Lifecycle processing carried the same flaw: one deeply nested workflow handled every stage transition, so scoring couldn't change without risking the machinery it ran on.

Constraints and complexity

The redesign had to be applied retroactively across roughly 400,000 existing contacts without disrupting active routing or in-flight programs.

What I did

How I approached it

01Diagnosed the mismatch as architecturalBusiness complexity had outgrown a model designed for simplicity — no amount of tuning fixes that.
02Analyzed conversion by portfolioRather than in aggregate, where one portfolio's strong fit rate hides the other's weak conversion.
03Separated scoring from lifecycle mechanicsThe model was wrong and the machinery running it was fragile. Treating them as one problem meant fixing neither.
04Grounded fit thresholds in won dealsPulled opportunity data to confirm what the ICP actually looked like per portfolio, rather than setting fit criteria from assumption.

What I built

Fit and behavior scored independently

Each on its own axis with a minimum threshold on both — a high-engagement poor-fit lead no longer qualifies on volume alone, and hand-raisers always qualify regardless of score.

A model per portfolio, interest per product

Each portfolio tunes its own fit and behavior thresholds to its buying motion, while a product-interest property tracks engagement at the product level underneath.

Cumulative scoring, with decay and an ineligible stage

Repeat behavior accumulates instead of earning a single credit, scores decay for inactive contacts, and a dedicated lifecycle stage holds records that should never accrue toward MQL at all.

Modular lifecycle architecture, validated before cutover

One HubSpot workflow per stage replaced the monolith, each independently testable. The new model ran in parallel with the old across all 400,000 contacts before anything switched.

Supporting materials

The same lead scored two ways: under one global score it totals 60 and becomes an MQL; under two axes the same points split into a fit score of 10, below minimum, and a behavior score of 50, above minimum, so it does not qualify — The failure the rebuild fixed — one score can't tell a well-matched buyer from an engaged poor fit.

Matrix mapping fit score bands against marketing engagement score bands, with every cell assigned an action — The handoff contract, written down. Every combination of fit and engagement maps to one action — MQL, hand-raiser only, nurture, or none — so no cell is left to interpretation.

Nine numbered lifecycle workflows, one per stage, running in lifecycle order from Lead through Customer, with one MQL workflow per MQL type — One nested workflow replaced by one numbered workflow per stage — each independently testable, and every lifecycle transition traceable to exactly one place.

Diagrams and figures are representative. Proprietary architecture, internal system names, and confidential customer details are omitted or generalized.

Rebuilding the intake layer so the funnel filtered instead of flooded

Marti — a MarOps AI teammate