Data-anchored lead scoring (L4)

The steps

1. 01

   Define the success target

   Start by opening your CRM (Salesforce or HubSpot) and pulling a report of all opportunities tagged as 'Closed Won' over the previous 12-18 months. You must define a 'Success Event' that is binary and objective,avoid 'Sales Qualified Lead' as it is subjective. Instead, filter for deals where the Close Date is within 90 days of creation and the Amount/ACV exceeds your historical median (e.g., >$10,000). • Who owns it: RevOps Manager or Head of Sales. • Time: 3-5 hours. • Tools: Salesforce Report Builder or HubSpot Custom Report Builder. • Pitfall: Including 'legacy' accounts or accidental duplicates which inflate success rates. Always filter for 'New Business' record types only. • Definition of Done: A CSV export or a SQL table containing two columns: 'Lead/Account ID' and 'Is_Success' (1 for won meeting criteria, 0 for everything else). This is your 'Ground Truth' dataset.

2. 02

   Build the multi-source feature set

   Now, gather the characteristics (features) that might predict that success. You need four distinct categories: Firmographic (from ZoomInfo or Clearbit), Product Usage (from Mixpanel, Pendo, or your Snowflake/BigQuery warehouse), Intent (from G2 or 6sense), and Engagement (from Marketo/HubSpot). • Actions: Map specific fields like 'Monthly Active Users', 'Visit to Pricing Page', 'Company Revenue', and 'Total Email Clicks'. • Example SQL: SELECT account_id, count(login_event) as login_count FROM product_logs WHERE timestamp > now() - interval '30 days'. • Ownership: Data Engineer or RevOps Specialist. • Time: 10-15 hours. • Pitfall: Using 'leaky' features,data that only exists *after* a deal is already won (like 'Seat Provisioned Count'). Only use data available at the moment the lead was first created or assigned. • Definition of Done: A master 'Feature Catalog' spreadsheet documenting every data point, its source, and its refresh frequency.

3. 03

   Backtest and validate lift

   Before deploying the model, you must prove it works on past data. Use a tool like Census, Hightouch, or a Python notebook (Random Forest or Logistic Regression) to apply your scoring logic to a 'Holdout Set',a group of leads from 6 months ago that the model hasn't 'seen' yet. • Actions: Divide your historical leads into 10 groups (deciles) based on their predicted score. Calculate the actual win rate for each decile. • Goal: The top 10% (Decile 1) should have a win rate at least 3x higher than the bottom 50%. • Who owns it: Data Scientist or RevOps Analyst. • Time: 8-12 hours. • Pitfall: Overfitting. If your model is 99% accurate on historical data, it’s likely 'memorizing' names rather than learning patterns. • Definition of Done: A 'Lift Chart' showing a clear correlation between high scores and high win rates. If the bottom decile wins more than the top, stop and rebuild.

4. 04

   Run a champion/challenger pilot

   Now, move the model into your CRM. Create two 'Score' fields: 'Current_Score' and 'Challenger_Score'. Use a routing tool like LeanData or a simple Salesforce Flow to assign leads to two different groups. • Config: Route 80% of leads to the 'Control' group (your old scoring) and 20% to the 'Challenger' group (the new AI-anchored model). • Instructions: Tell the BDR team that for the 20% group, they must follow the AI's prioritization strictly. • Who owns it: Sales Ops and BDR Manager. • Time: 4 weeks (for observation). • Pitfall: Sales reps cherry-picking leads regardless of the score, which ruins the experiment. Use a 'Blind' test if possible where the rep doesn't know which model scored the lead. • Definition of Done: A dashboard comparing 'Conversion to Opportunity' and 'Win Rate' between the 80% and 20% groups.

5. 05

   Execute the quarterly drift check

   AI models 'decay' as market conditions change. Every 90 days, you must run a 'Drift Report' to see if the features that used to predict success still do. • Actions: Re-run your backtest with the most recent 3 months of data. Check if the 'Importance' of certain features has shifted (e.g., maybe 'Webinar Attendance' matters less now than 'G2 Intent'). • Prompt for Analysis: 'Compare the feature weights from Q1 to Q2. Highlight any attribute that dropped more than 15% in predictive power.' • Who owns it: RevOps Manager. • Time: 4 hours quarterly. • Pitfall: 'Set it and forget it' syndrome. If you don't retrain, the model will eventually start surfacing low-quality leads because it's looking for patterns from a year ago. • Definition of Done: A signed-off 'Model Health Report' that either triggers a retraining or confirms the current weights are still valid. No report = the AI is turned off.

Next playbooks

• Voice-of-Customer → Product & Enablement Loop (L2)
• Zapier Agents for sales ops chores (L2)
• Agentic Nightly CRM Hygiene + Ownership Sync (L3)
• AI-driven price-quote acceleration with CPQ (L3)