A B2B SaaS company selling supply chain management software had a pipeline problem. Their marketing team was generating 4,800 leads per quarter across webinars, content downloads, trade shows, and inbound demos. The sales team treated every lead the same — working them in the order they arrived. Average conversion rate: 3.2 percent. The sales team was burning time on thousands of leads that would never convert while high-intent buyers waited in the queue.
We built a propensity-to-buy model that scored every lead based on firmographic data, engagement behavior, technographic signals, and intent data. The model assigned a 0-100 score that predicted the probability of conversion within 90 days. When we ranked leads by propensity score, the top 20 percent converted at 14.1 percent — more than 5x the bottom 80 percent at 2.6 percent. The sales team restructured their workflow to prioritize high-propensity leads, and quarterly bookings increased by 38 percent with no increase in headcount.
Propensity modeling is one of the most practical and highest-ROI deliverables an AI agency can offer B2B companies. The data requirements are modest, the models are relatively straightforward, and the business impact is immediate and measurable. Here is how to deliver these projects.
Why Propensity Modeling Is a High-Leverage Agency Service
Every company with a sales team is trying to answer the same question: which prospects are most likely to buy? Propensity models answer this question with data.
The impact of good lead scoring:
- Sales teams spend 60-70 percent less time on leads that will never convert
- Conversion rates on prioritized leads increase by 2-5x
- Sales cycle length decreases because reps engage high-intent prospects earlier
- Marketing ROI improves because campaigns can be optimized toward propensity, not just volume
- Revenue per sales rep increases by 25-40 percent when working prioritized leads
What clients will pay: Propensity modeling projects range from $40,000 for a basic lead scoring model to $150,000+ for comprehensive buyer intent platforms. Ongoing retainers for model maintenance and optimization run $5,000-15,000 per month.
Client types: B2B SaaS companies, professional services firms, financial services, insurance companies, healthcare companies — any business with a sales team working inbound or outbound leads.
Types of Propensity Models
Propensity modeling is not one model — it is a family of models, each predicting a different behavior.
Propensity to Buy (Lead Scoring)
The most common model. Predicts the probability that a lead will convert to a customer.
Used for: Lead prioritization, sales resource allocation, marketing campaign optimization.
Propensity to Expand (Upsell/Cross-sell)
Predicts the probability that an existing customer will purchase additional products or upgrade their current plan.
Used for: Account prioritization for customer success and account management teams, targeting for upsell campaigns.
Propensity to Churn
Predicts the probability that a customer will cancel or not renew.
Used for: Retention outreach prioritization, customer success resource allocation, churn prevention campaigns.
Propensity to Engage
Predicts the probability that a prospect will respond to a specific outreach (open an email, attend a webinar, take a meeting).
Used for: Campaign targeting, channel optimization, outreach timing.
Propensity to Refer
Predicts the probability that a customer will refer new business.
Used for: Referral program targeting, advocacy program enrollment.
Most engagements start with propensity to buy and expand to other models as the client sees value.
Data Requirements and Feature Engineering
Data Sources
CRM data (Salesforce, HubSpot, etc.):
- Lead and opportunity records
- Conversion history (which leads became customers)
- Sales activities (calls, emails, meetings)
- Deal characteristics (size, stage, duration)
- Account and contact demographics
Marketing automation data:
- Email engagement (opens, clicks, unsubscribes)
- Content consumption (page views, downloads, form submissions)
- Event attendance (webinars, conferences, demos)
- Ad engagement (clicks, impressions)
- Website behavior (pages visited, time on site, return visits)
Firmographic data:
- Company size (employees, revenue)
- Industry and sub-industry
- Location and geography
- Growth rate and funding stage
- Technology stack (from technographic providers)
Intent data:
- Third-party intent signals (topics being researched)
- Review site activity (G2, Capterra, TrustRadius)
- Job posting signals (hiring for roles that indicate need for the product)
- News and trigger events (funding rounds, executive changes, mergers)
Product usage data (for existing customers):
- Feature adoption and usage frequency
- Login frequency and active users
- Support ticket volume and sentiment
- Contract utilization vs entitlement
Feature Engineering
Raw data needs to be transformed into predictive features. The quality of feature engineering determines the quality of the model.
Behavioral features:
- Total number of marketing touches in the last 30, 60, and 90 days
- Recency of last engagement (days since last website visit, email click, etc.)
- Depth of engagement (number of pages viewed per session, time on high-intent pages)
- Engagement velocity (is engagement increasing, decreasing, or flat?)
- Channel diversity (how many different channels has the lead engaged through?)
- Content type affinity (does the lead engage more with product pages, case studies, or educational content?)
- High-intent page visits (pricing page, demo request page, comparison pages)
Firmographic features:
- Company size relative to ideal customer profile
- Industry fit score
- Technology stack compatibility
- Growth indicators (hiring velocity, funding events)
- Geographic match to service area
Temporal features:
- Day of week and time of day engagement patterns
- Seasonal patterns in the client's sales cycle
- Lead age (how long since first touch)
- Time since key events (first form fill, first demo request, first content download)
Interaction features:
- Sales engagement level (has a rep reached out? How many times?)
- Response to sales outreach (replied, booked meeting, no response)
- Multi-stakeholder engagement (are multiple people from the same company engaging?)
Technical Architecture
Model Architecture
Classification approach: Binary classification predicting conversion (yes/no) within a defined time window.
Model selection:
Gradient-boosted trees (LightGBM, XGBoost) are the standard choice for propensity modeling because:
- They handle mixed data types (numerical, categorical) natively
- They are robust to missing values (common in CRM data)
- They capture feature interactions automatically
- They are interpretable through feature importance and SHAP values
- They train quickly and score quickly for real-time applications
Calibration: The raw model output needs to be calibrated so that a score of 0.8 actually means an 80 percent probability of conversion. Use Platt scaling or isotonic regression to calibrate probabilities.
Scoring tiers: Convert continuous probabilities into actionable tiers that sales teams can understand:
- A leads (top 10 percent by score): Highest priority, route to senior reps
- B leads (next 20 percent): High priority, assign to dedicated reps
- C leads (next 30 percent): Standard priority, nurture and monitor
- D leads (bottom 40 percent): Low priority, automated nurture only
Integration Architecture
Propensity scores are useless if they do not reach the people and systems that need them.
CRM integration: Scores must appear in the sales rep's daily workflow — in the CRM lead record, in their prioritized work queue, in their dashboard. Most CRM platforms support custom fields and scoring. Push scores via API.
Marketing automation integration: Scores should inform marketing campaigns — high-propensity leads get aggressive conversion-focused campaigns, low-propensity leads get long-term nurture sequences.
Real-time scoring: For high-velocity sales motions, score leads in real-time as they engage. When a lead visits the pricing page, their score should update immediately and trigger a sales alert.
Batch scoring: For lower-velocity motions, batch score all leads daily or weekly and update CRM records.
Model Monitoring
Propensity models degrade over time as customer behavior and market conditions change. You need monitoring.
Key monitoring metrics:
- Score distribution: Is the distribution of scores shifting over time?
- Conversion rate by score tier: Are A leads still converting at the expected rate?
- Model AUC/precision/recall: Are these metrics declining on recent data?
- Feature drift: Are input feature distributions changing significantly?
- Calibration: Are predicted probabilities still matching actual conversion rates?
Set up automated alerts when any metric drifts beyond acceptable thresholds.
Delivery Framework
Phase 1: Data Audit and Discovery (Weeks 1-2)
Activities:
- Audit CRM data quality (completeness, accuracy, consistency)
- Map the existing lead-to-customer conversion process
- Define the conversion event and time window for the model
- Identify available data sources and features
- Assess data volume (do we have enough historical conversions for modeling?)
- Define success metrics and model performance targets
Minimum data requirements: At least 200-300 positive conversions in the training data. Fewer than that and the model will be unreliable. If the client has fewer, consider a longer time window, a broader definition of conversion, or supplementing with external data.
Phase 2: Feature Engineering and Model Development (Weeks 3-6)
Activities:
- Extract and transform data from CRM, marketing automation, and external sources
- Engineer features from raw data
- Split data into training, validation, and test sets (time-based split, not random)
- Train and evaluate multiple model architectures
- Perform hyperparameter tuning
- Evaluate feature importance and model interpretability
- Validate model fairness (no discrimination based on protected characteristics)
Phase 3: Integration and Deployment (Weeks 7-9)
Activities:
- Build the scoring pipeline (real-time and/or batch)
- Integrate scores with CRM and marketing automation
- Build score-based lead routing and alerting
- Create dashboards for sales managers to monitor lead quality
- User training for sales reps and managers
- Documentation of model methodology and scoring interpretation
Phase 4: Optimization and Validation (Weeks 10-12)
Activities:
- Monitor model performance on live data
- Collect feedback from sales reps on score quality
- A/B test scored vs unscored lead routing
- Measure impact on conversion rates, pipeline velocity, and revenue
- Iterate on features and model based on feedback
- Establish ongoing retraining schedule
Common Delivery Challenges
CRM Data Quality
CRM data is almost always a mess. Leads are not consistently categorized. Conversion timestamps are inaccurate. Opportunity stages are used inconsistently. Duplicate records are rampant.
Mitigation:
- Budget significant time for data cleaning in the first phase
- Work with the sales operations team to understand data entry practices
- Build data quality checks into the pipeline that flag issues
- Be transparent with the client about how data quality affects model performance
- Recommend CRM hygiene improvements as part of the engagement
Defining "Conversion"
What counts as a conversion? Qualified opportunity? Closed-won deal? First meeting booked? The definition matters because it determines what the model learns to predict.
Our recommendation: Define conversion as the earliest reliably recorded event that indicates genuine buying intent. For many B2B companies, this is "qualified opportunity created" rather than "closed-won." This gives the model more positive examples to learn from and provides earlier predictions.
Sales Process Changes
If the client changes their sales process during the modeling period — new qualification criteria, new sales stages, territory changes — historical data may not be representative of the current process.
Handle this:
- Use the most recent stable period for model training
- Exclude periods of significant process change
- Rebuild the model after major process changes settle
- Design the model to be retrained quarterly or semi-annually
Adoption Resistance from Sales Reps
Sales reps trust their instincts and may resist being told which leads to prioritize by an algorithm.
Driving adoption:
- Show reps specific examples where the model identified winners they would have missed
- Start with a "pilot group" of reps who voluntarily use scores and measure their performance improvement
- Present scores as a prioritization tool, not a mandate — reps can still pursue any lead they believe in
- Share weekly "model wins" — cases where high-scored leads converted
- Get sales leadership to champion the tool and set expectations for usage
Pricing Propensity Modeling
Project-based pricing:
- Basic lead scoring model: $40,000-70,000
- Comprehensive propensity platform (buy + expand + churn): $100,000-180,000
- Enterprise propensity system (multiple products, markets, with real-time scoring): $150,000-250,000
Ongoing retainer:
- Model monitoring and retraining: $5,000-10,000 per month
- Feature expansion and optimization: $3,000-8,000 per month
- Reporting and insights: $3,000-5,000 per month
Value justification: A company with $50 million in pipeline that increases conversion rate from 3 percent to 5 percent through better lead prioritization adds $1 million in revenue. A $75,000 propensity model pays for itself in less than a quarter.
Your Next Step
Find a B2B company with at least 200 customer conversions in their CRM history and a sales team that is treating all leads equally. Offer a paid pilot where you build a propensity model on their historical data and backtest it — show them that the top 20 percent of leads by score historically converted at 3-5x the rate of the bottom 80 percent. That retroactive proof is the most powerful sales tool you have because the client can verify it against their own records.