All Use CasesPropensity Modeling

Propensity Modeling

Score every customer's likelihood to buy, upgrade, renew, or respond. EkamFlow's propensity models predict the probability of any business action and feed directly into your marketing automation.

In one paragraph

Propensity modeling uses machine learning to score each customer's likelihood of taking a specific action — buying, upgrading, renewing, churning, responding to a specific campaign, redeeming an offer, referring a friend. Instead of running a separate model per action (which most in-house ML teams end up doing, creating a maintenance sprawl), modern propensity systems return a full set of action probabilities in one API call. Marketing and CRM teams use propensity scores to gate campaign audiences, trigger lifecycle motions, and personalize product experiences — replacing the manual scoring rules that CDPs and CRMs traditionally require.

What is Propensity Modeling?

Propensity modeling predicts the likelihood that a customer will take a specific action — purchase a product, upgrade their plan, renew their subscription, respond to a campaign, or churn. Unlike demographic-based targeting, propensity models analyze behavioral signals to produce a probability score for each individual customer.

Companies use propensity models across the entire customer lifecycle: propensity to buy for acquisition targeting, propensity to upgrade for expansion campaigns, propensity to churn for retention triggers, and propensity to respond for campaign optimization. Each traditionally requires its own model.

EkamFlow handles all propensity models within a single private model. One API call returns propensity scores for every action you define — buy, upgrade, renew, respond, churn — trained on your specific customer data and behavioral signals. No separate models to build or maintain.

How it works, step by step

  1. 1

    Connect your warehouse

    Snowflake, BigQuery, Databricks, Redshift, or Postgres. Read-only access to customer, event, and campaign-response tables.

  2. 2

    Auto-detect action tables

    EkamFlow identifies the events that count as 'actions' — purchases, upgrades, sign-ups, referrals, opens, clicks — and treats each as a scorable outcome without manual configuration.

  3. 3

    Private model trains on your response history

    The model learns which customers respond to which actions, and what signal combinations predict each response type — specific to your customer base and campaign catalog.

  4. 4

    Multi-action scores in one API response

    Every request returns propensity scores for buy, upgrade, churn, respond, refer, and other configurable actions — in a single response, not one API call per model.

  5. 5

    Continuous re-scoring

    As new campaigns run and new response data accumulates, propensity scores update automatically. Adding a new scorable action doesn't require training a new model — the platform handles it.

Signals the model uses

  • Historical response rate per action type
  • Recent engagement (opens, clicks, sessions, purchases)
  • Lifecycle stage and cohort behavior
  • Purchase frequency and category affinity
  • Discount elasticity and price sensitivity
  • Cross-channel activity patterns
  • Support-ticket volume and sentiment
  • Product usage or subscription-tier progression
  • Referral and loyalty-program participation
  • Cohort behavior relative to peers

Where it fits in your stack

  • CDPs: Segment, Rudderstack, mParticle, Hightouch
  • Marketing automation: Braze, Iterable, Customer.io, Klaviyo, MailChimp
  • CRM: Salesforce, HubSpot, Zoho
  • Ad platforms: Meta CAPI, Google Enhanced Conversions, TikTok Events API
  • Reverse ETL: Census, Hightouch, Polytomic
  • Data warehouse: Snowflake, BigQuery, Databricks, Redshift
  • BI + analytics: Looker, Tableau, Metabase, Hex

Build in-house vs. EkamFlow

DimensionBuild in-houseEkamFlow
Time to first score3–6 months per action (compounds fast)Days
Team requiredData scientist per model + MLOpsNone
Model architectureOne model per action, maintained separatelyShared backbone + per-action heads, unified training
Adding new action types3+ months per new modelConfiguration, not modeling
Signal consistencyEach model has different featuresAll actions score against the same signal set
Marketing automation integrationCustom per modelREST API + reverse-ETL to CDPs/ESPs
LatencyBatch scoring<60ms real-time
Combined with other predictionsSeparate systemsPropensity + churn + LTV + NBA in one call

What it looks like in different industries

DTC E-commerce marketing

A DTC brand uses propensity-to-buy scores to size Meta lookalike audiences and gate paid retargeting spend to the customers actually likely to convert. Propensity-to-refer scores drive referral-program invitation targeting. Same API, two distinct campaigns — no separate models to maintain.

B2B SaaS growth

A subscription analytics platform runs propensity-to-upgrade scoring across their trial base and paid mid-market accounts. Sales team prioritizes accounts with propensity >0.5 for upgrade outreach; propensity <0.2 accounts get automated feature-education nurture. Sales productivity per account meaningfully improved.

Retail loyalty marketing

A national grocer uses propensity-to-redeem scores to decide which loyalty members get which offers. High-redemption-propensity members get the deeper offers; low-propensity members get engagement-focused content. Offer-redemption rate on issued offers doubled without changing offer creative.

Fintech & neobank

A neobank feeds propensity-to-adopt scores for each product (secured card, buy-now-pay-later, savings goal) into their in-app cross-sell surfaces. Customer sees the specific product they're most likely to adopt; no more one-size-fits-all product tiles. Attach rate per session improved measurably.

Media & subscription video

A streaming platform's retention team uses propensity-to-cancel and propensity-to-upgrade scores in the same API call to route lifecycle motions. High-cancel + low-upgrade subscribers get save-focused content; low-cancel + high-upgrade get annual-plan reveals. No more contradictory campaigns.

Ad tech & performance marketing

A performance-marketing team feeds propensity-to-convert into custom-audience-value on Meta and Google. Bids scale to per-customer conversion propensity instead of blanket audience-level targeting. Blended CAC on retargeting fell without changing creative or budget.

How EkamFlow does it

Multi-action propensity

Score propensity to buy, upgrade, renew, respond, and churn — all from one model. No need to build separate models for each business action.

Feed directly into marketing automation

Propensity scores flow into your CDP, CRM, or marketing platform via API. Trigger campaigns, adjust bidding, and personalize experiences based on real-time propensity.

Continuously learning

Propensity models retrain on new behavioral data automatically. Scores stay accurate as customer behavior and market conditions evolve.

GET /v1/predict
{
"customer_id": "cust_29841",
"propensity_to_buy": 0.74,
"propensity_to_upgrade": 0.61,
"propensity_to_churn": 0.12,
"propensity_to_respond": 0.88,
"latency_ms": 10
}
one API · sub-60ms · all predictions

Frequently asked about propensity modeling

CDP-native propensity features usually score one or two actions (churn, high-value customer) using a limited feature set from CDP-collected events. EkamFlow scores across every action you configure, uses your full warehouse signal (which usually far exceeds what your CDP sees), and returns multiple propensity scores in a single response. Companies commonly run both — CDP scoring for CDP-native workflows, EkamFlow for the deeper multi-action scoring that drives marketing automation and personalization.

Practically unlimited — the model architecture (shared backbone + per-action heads) scales to dozens of actions without linear cost or maintenance growth. Most customers score 5–15 actions initially and add more as marketing motions require them. Adding an action is configuration, not a new model.

Rare-action scoring uses the shared backbone's signal representations plus a small amount of task-specific data, so it works better than pure per-action modeling on limited data. For extremely rare actions with fewer than a few hundred positive outcomes, scores are provided with wider confidence bands and cohort-fallback defaults.

Yes. Propensity-to-buy, propensity-to-upgrade, and propensity-to-respond are commonly pushed to Meta CAPI, Google Enhanced Conversions, and TikTok Events API as per-customer values that scale bids to predicted response probability. This is one of the highest-ROI single applications of propensity scoring.

Cold-start propensity for a specific action uses cohort behavior — customers with similar profiles who have taken the action. Personal-signal weight increases as the customer produces action-specific response data (either positive or negative), typically within 3–10 exposures per action.

Yes — that's the primary usage pattern. One API call returns propensity to buy, propensity to churn, propensity to upgrade, plus predicted LTV and next best action. Marketing automation platforms consume the whole response and use each score for a different piece of the lifecycle motion.

Ready to add propensity modeling to your stack?

Connect your warehouse. Get live predictions in weeks, not months.

Book a demo