Thesis
Intelligence is no longer scarce. Calibrated probability is. Most enterprises drown in dashboards that describe the past. Predictive Intelligence exists to answer a narrower, harder question: given what we know about this actor, how likely is a defined outcome within a defined window?
Heuman Labs treats predictive work as propensity modeling — supervised learning systems that emit a score, typically on [0, 1] or 0–100%, representing the mathematical probability that a user, account, or entity takes a named action. The name of the model is the name of the behavior it forecasts.
What a propensity model is
A propensity model is trained on historical trajectories: who acted, who did not, under what conditions, across which channels. The objective is not narrative insight — it is a rank-ordered probability that lets operators prioritize.
Technically, the stack is familiar but the discipline is not: feature stores that encode behavioral state, careful label definition (the outcome must be unambiguous), temporal leakage controls, and calibration so that a score of 0.72 means roughly 72% realized frequency in a held-out cohort — not merely a relative rank.
Common estimators include gradient-boosted trees (XGBoost, LightGBM, CatBoost), regularized logistic regression for interpretable baselines, and — when sequence and timing matter — temporal models over event streams. Architecture is secondary to label honesty and calibration under shift.
Commercial & marketing models
These are the propensity families enterprises ask for most often. Each answers a different capital-allocation question.
Purchase Propensity
Predicts — Likelihood a lead or visitor buys within a defined horizon.
Best at — Media and sales prioritization when CAC is rising and volume is noise.
Churn / Attrition Propensity
Predicts — Risk of cancel, uninstall, or competitor switch.
Best at — Retention budgets aimed at accounts where intervention still moves the curve.
Upsell & Cross-Sell Propensity
Predicts — Probability an existing customer expands tier or adds adjacency.
Best at — Account growth without blanket discounting every cohort.
Engagement Propensity
Predicts — Likelihood of open, click, visit, or meaningful interaction.
Best at — Thinning contact lists so attention goes to receptive surfaces.
Unsubscribe / Opt-Out Propensity
Predicts — Proximity to fatigue and channel exit.
Best at — Protecting deliverability by lowering frequency before the unsubscribe.
Conversion / Lead Scoring
Predicts — Probability of a funnel milestone — trial, demo, download.
Best at — Routing scarce human follow-up to prospects with real completion odds.
Specialized & domain models
Beyond commercial defaults, domain physics changes the label and the features. We research and deploy specialized propensity when the business outcome is not a generic click.
Event Propensity
Predicts — Timing of life or buying milestones (move, wedding, vehicle).
Best at — Periodic markets where readiness arrives in windows, not continuously.
Donation / Contribution Propensity
Predicts — Likelihood to donate, volunteer, or contribute.
Best at — Non-profits and campaigns optimizing scarce outreach without alienating bases.
Channel-Preference Propensity
Predicts — Which channel (SMS, email, push, voice) yields response.
Best at — Omnichannel stacks where the wrong surface destroys ROI.
How Heuman Labs builds them
Our method is research-first and deployment-honest. Forward Deployed Engineers sit with the decision owners; AI/ML experts own the statistical contract.
- 01Frame the decision the score will change — not a vanity metric.
- 02Define the label with temporal boundaries and exclusion rules that survive legal and product review.
- 03Engineer features from behavioral state without leaking the future into the past.
- 04Train competing estimators; select on discrimination and calibration, not leaderboard vanity.
- 05Stress-test under cohort shift, seasonality, and cold-start segments.
- 06Ship monitoring: score drift, label delay, and business KPI coupling.
Where accountable decisions demand it, we retrain on a cadence tied to the half-life of the behavior — weekly for fast funnels, monthly or quarterly for slower B2B cycles.
Enterprise engagement
For enterprises, we propose a bounded research engagement that produces a production-ready propensity surface — not a slide deck.
Typical shape: map the decision and data reality → stand up one flagship propensity (often churn or conversion) → validate lift in a controlled holdout or shadow mode → wire scores into CRM, CDP, or Decision Intelligence policies → expand the model catalog once the first score earns trust.
From score to system
A propensity score is an input. The enterprise still must decide: who gets the retention offer, which lead gets the call, which channel fires tonight. That is where Predictive Intelligence meets Decision Intelligence— and, when the chosen action must run across systems without human babysitting, Agentic Business Optimisation.
Heuman Labs designs the stack so the probability is not orphaned in a warehouse — it becomes the prior for policies that act under constraint, cost, and uncertainty, then for agents that close the loop.
