The Portfolio Company Revenue Problem
You’ve acquired a solid business. The product works. The team is competent. But revenue is unpredictable, acquisition costs are creeping up, and the board is asking harder questions about unit economics and growth trajectory.
The core issue isn’t usually the product or the sales team. It’s that the company is still operating on reactive analytics: chasing traffic volume, spreading budget across low-intent prospects, and hoping something converts.
This approach worked when growth was the only metric. It doesn’t work under PE ownership, where capital efficiency, predictable outcomes, and decision-grade visibility are non-negotiable.
This is a Control problem, not a reporting problem.
When numbers exist but are not trusted, the issue is rarely dashboards or tooling. It is lost Control — where metrics no longer support confident decisions early enough to matter.
This pattern sits within how Control governs the system.
Why Human-Powered Analytics Fail Under Pressure
Traditional analytics in portfolio companies follows a predictable pattern:
Report Generation. Hours spent building dashboards showing traffic, conversions, cost-per-acquisition. Metrics that feel important but don’t predict behaviour.
Manual Analysis. Segmenting users by obvious categories—source, device, geography—missing the subtle combinations of behaviour that actually signal intent.
Insight Extraction. Analysts present findings to leadership, but the signal is buried in noise. By the time a decision is made, the moment for intervention has passed.
Latency at the Last Mile. Insights compete with other priorities. Execution stalls. Opportunities are lost to competitors.
This process is fundamentally reactive. It tells you what happened yesterday, not what will happen tomorrow. Under PE ownership, that latency is expensive.
The Propensity Modelling Shift
Propensity modelling answers a different question: What is the exact probability THIS prospect will convert, upgrade, or churn in the next N days?
Instead of spreading resources across all prospects equally, propensity models identify high-intent users in real-time, enabling focused intervention, dynamic offer allocation, and predictable revenue outcomes.
Machine learning algorithms excel at this because they process hundreds of behavioural variables simultaneously—scroll depth, product views, cart additions, time spent, visit frequency, feature engagement, support interactions—and identify non-linear patterns that humans miss.
A prospect who views pricing three times might convert. Or they might be price-sensitive and never close. But a prospect who views pricing, reads case studies, engages with ROI calculators, AND visits the pricing page during business hours? That’s a different signal entirely. ML algorithms catch these interaction effects at scale.
How Propensity Modelling Works in Practice
Step 1: Data Foundation
Propensity models require clean, unified data. This means integrating:
- Website behaviour (page views, scroll depth, time on page, feature interactions)
- Product usage (if applicable—feature adoption, session length, repeat engagement)
- Sales engagement (email opens, demo attendance, proposal views)
- Customer data (company size, industry, geography, firmographic fit)
- Outcome data (conversion, deal size, time-to-close, churn)
For portfolio companies, this often means consolidating data from your CRM, web analytics, product platform, and email system into a single source of truth.
Step 2: Model Selection
Several machine learning approaches work well for propensity modelling:
Gradient Boosted Machines (XGBoost, LightGBM).
Currently the gold standard for tabular data. These algorithms combine many weak predictive models into a highly accurate ensemble. They handle mixed data types, capture non-linear relationships, and provide feature importance rankings—telling you which behaviours most strongly predict conversion.
Random Forests.
Excellent for understanding feature importance. If you need to know “which behaviours most strongly predict conversion,” Random Forests give you interpretable answers alongside predictions.
Neural Networks.
For massive datasets with complex, non-linear relationships, deep learning can uncover patterns other models miss. Less interpretable, but often more accurate at scale.
Survival Analysis.
Predicts not just if, but when a user will convert, enabling perfectly timed interventions. Particularly useful for understanding churn risk and renewal probability.
For most portfolio companies, Gradient Boosted Machines are the practical starting point: accurate, interpretable, and fast to deploy.
Step 3: Real-Time Scoring
Once trained, the model scores every prospect in real-time as they interact with your digital properties. A prospect visiting your site gets a propensity score—a probability (0-100%) of converting in the next 7, 14, or 30 days.
Step 4: Activation
Now the model drives action:
- Dynamic Offer Allocation. High-propensity users see targeted offers, premium content, or expedited sales outreach. Low-propensity users see nurture content or educational resources.
- Bid Strategy Optimisation. Retargeting ad budgets are allocated based on conversion probability, not generic audience segments. You bid aggressively for high-propensity users, conservatively for low-propensity.
- Sales Prioritisation. Your sales team focuses on high-propensity prospects, reducing time wasted on low-intent leads.
- Intervention Timing. Identify “at-risk” users who showed high intent but stalled—and intervene before they convert to a competitor.
Late surprises mean Control is already gone.
When forecasts move late, teams disagree on definitions, or decisions stall due to uncertainty, Control has already failed — even if reporting looks detailed.
At this stage, adding more metrics increases noise, not confidence.
The Commercial Outcomes
Based on real portfolio company implementations:
Conversion Rate Improvement: 35-60%
For targeted segments, conversion rates improve significantly. A portfolio company with a 2% baseline conversion rate might see 2.8-3.2% on high-propensity segments—a material uplift in revenue per visitor.
Acquisition Cost Reduction: 20-35%
By focusing budget on high-probability prospects, customer acquisition costs drop. You’re not chasing volume; you’re chasing intent. Lower spend, higher conversion, better unit economics.
Sales Efficiency Gains
Sales teams spend less time qualifying low-intent leads and more time closing high-intent prospects. Cycle times compress. Win rates improve. Board presentations become more predictable.
Capital Allocation Clarity
For PE investors, propensity modelling delivers something rare: explainable, evidence-based capital allocation. You can justify why you’re spending on specific channels, specific offers, specific timing—because the model shows the probability of return.
Implementation Reality
Propensity modelling isn’t a quick fix. It requires:
Clean Data.
If your CRM is messy, your analytics fragmented, or your product data incomplete, the model will reflect that. Data quality is the foundation.
Technical Capability.
You need someone who can build and maintain ML models, or you need to partner with a vendor. This isn’t a plug-and-play tool.
Commercial Governance.
The model is only useful if leadership understands what it’s telling you and acts on it. A propensity score sitting in a dashboard is worthless. It needs to drive decisions.
Sequencing.
Propensity modelling works best when your core revenue system is stable. If your product-market fit is unclear, if your sales process is chaotic, or if your pricing is unstable, propensity modelling will amplify confusion. Get the fundamentals right first.
When Propensity Modelling Isn’t the Answer
Not every portfolio company needs propensity modelling. Red flags:
- Early-stage product. If you’re still validating product-market fit, focus on that first. Propensity modelling requires stable conversion patterns to learn from.
- Transactional, high-volume model. If you sell thousands of low-value items, propensity modelling overhead may not justify the return.
- Broken sales process. If your sales team isn’t following a consistent process, if deals are unpredictable, or if there’s no clear buying signal, the model won’t help.
- No data integration. If your systems don’t talk to each other, building a propensity model is expensive and fragile.
The Sequencing Question
Here’s where commercial governance matters: propensity modelling only works if it’s addressing the right constraint.
In ATMC terms, propensity modelling is a Control play. It assumes your Attention (demand quality), Trust (buyer confidence), and Movement (deal progression) are reasonably stable. If they’re not, propensity modelling will just help you predict failure more accurately.
Before you invest in propensity modelling, ask:
- Attention: Are you attracting the right prospects? Or are you chasing volume from the wrong sources?
- Trust: Do buyers hesitate at approval? Do they need repeated reassurance? Or do they move forward confidently?
- Movement: Do deals progress predictably? Or do they stall, loop back, or die in the pipeline?
If the answer to any of these is “no,” fix that first. Propensity modelling amplifies a working system; it doesn’t fix a broken one.
Without Control, growth becomes guesswork.
Leadership cannot trust what is happening or what is likely to happen next, every decision carries unnecessary risk.
The Control Focus Package exists to restore decision-grade confidence — not more reporting.
Frequently Asked Questions
The Bottom Line
Propensity modelling isn’t magic. It’s a disciplined application of machine learning to a specific commercial problem: identifying high-intent prospects and allocating resources accordingly.
For PE portfolio companies, it delivers something investors care about: predictable, explainable revenue outcomes. No guesswork. No vanity metrics. Just evidence-based capital allocation.
But it only works if you’re ready for it—if your data is clean, your process is stable, and your leadership understands what the model is telling you and acts on it.
The companies winning under PE ownership aren’t chasing more traffic. They’re chasing better traffic, with discipline, and with evidence
