Account Scoring & Propensity Modeling
Static ICP checklists break the moment your market shifts. A propensity model built from your actual wins and losses adapts with your data.The Play
Export wins and losses from your CRM
Pull closed-won and closed-lost opportunities from HubSpot or Salesforce via MCP. Include company attributes: employee count, industry, revenue, funding stage, deal size, sales cycle length.
Enrich the historical dataset
For each company, pull firmographics via Crustdata or Apollo: headcount, tech stack, hiring velocity, funding rounds, revenue estimates. The more signal dimensions, the better the model.
Build a baseline scoring model
Ask Claude to analyze the enriched win/loss dataset and identify which attributes correlate with winning. Weight each dimension by its predictive power. Start simple — logistic regression or a weighted scorecard.
Backtest against held-out deals
Hold out 20-30% of your historical deals. Score them with the model. Check: does the model rank wins higher than losses? What’s the AUC? If it’s below 0.65, the model needs more signal dimensions or the data is too noisy.
Iterate until consistent but not overfit
Tune weights, add or drop dimensions, re-run. Watch for overfitting: if training accuracy is 95% but holdout accuracy is 55%, you’re memorizing noise. Target holdout accuracy within 10% of training accuracy.
Score new target accounts
Apply the model to your pipeline and target account list. Tier them by propensity score. Focus outbound on the top tier.
Tell Claude Code
“Pull all closed-won and closed-lost deals from HubSpot for the last 12 months. Enrich each company with Crustdata firmographics. Build a propensity model that predicts win vs loss based on company attributes. Backtest it on 25% held-out deals. Keep iterating until holdout accuracy is within 10% of training accuracy. Then score the 300 companies in target-accounts.csv and tier them. Find the VP of Sales at every top-tier company and write everything to scored-accounts.csv.”
How the Model Works
Input: Your CRM’s closed-won and closed-lost deals, enriched with firmographic data. What the model learns:| Signal | Why it matters |
|---|---|
| Employee count range | You may win more often at 100-500 than at 5,000+ |
| Industry vertical | SaaS companies may convert 3x vs manufacturing |
| Funding stage | Series B might be your sweet spot |
| Hiring velocity | Companies adding GTM roles signal active buying |
| Tech stack overlap | Using Salesforce + Outreach might correlate with wins |
| Revenue range | 50M ARR might be where you win on value |
| Geography | Regional patterns in close rates |
| Sales cycle length | Short cycles at similar companies predict fit |
Avoiding Overfitting
The goal is a model that generalizes, not one that memorizes your training data.- Split your data — 75% train, 25% test. Never tune on the test set.
- Start with fewer dimensions — 5-8 signals. Add more only if holdout accuracy improves.
- Watch the gap — training accuracy 80%, holdout 72% = fine. Training 95%, holdout 55% = overfit.
- Re-run quarterly — your market changes. The model should too. Pull fresh wins/losses and retrain.
Cost Estimate
- Company enrichment via Crustdata: ~0.4 credits per account
- Contact finding (top tier): ~0.2-0.6 credits per company (via company_to_contact waterfall)
- Email waterfall for found contacts: ~0.3 credits per contact
- Model building and backtesting: no enrichment cost (runs on already-enriched data)
Niche Signal Discovery → | Closed-Lost Recovery → | I Have X, I Want Y →