Lead scoring v2 with first-party + LLM features (L4)

The steps

1. 01

   Aggregate historical conversion data

   Start by exporting your historical CRM data to identify patterns in winning leads. You need a minimum of 18 months of data to ensure seasonal trends don't skew the results. Navigate to your CRM (Salesforce/HubSpot) and create a report of all Leads and Accounts created in the last 18-24 months. You must include fields like Company Domain, Industry, Annual Revenue, and importantly, the 'Converted' or 'Opportunity Created' status. Combine this with Product Usage data from your warehouse (Snowflake/BigQuery) or product analytics tool (Mixpanel/Amplitude). • Technical Action: Export a CSV or sync to a Google Sheet using a connector like Coefficient. Ensure you have the Lead ID, Email Domain, and the outcome variable (e.g., Is_SQL = True/False). • Who Owns it: RevOps Manager or Data Analyst. • Time Estimate: 4-6 hours for data cleaning and joining. • Pitfall: Mixing 'Trial' leads with 'Inbound Content' leads without a source flag; this creates noise because their conversion behaviors differ wildly. • Definition of Done: A master dataset where every lead from the last 18 months has a binary '1' for success (converted to Opp) or '0' for failure, mapped to their company website.

2. 02

   Engineer LLM-derived fit features

   Standard CRM fields like 'Industry' are often garbage. Use an LLM to scrape and categorize leads based on their actual website content and external signals. Use a tool like Clay, Browse.ai, or a Python script calling the OpenAI GPT-4o API. For every lead in your list, fetch the homepage text or recent job postings. • Example Prompt: 'Analyze the following company description and job titles. Return a JSON with three fields: 1. Is_AI_Native (Boolean), 2. Target_Persona_Hiring (High/Med/Low), 3. Value_Prop_Match (1-10 score).' • Logic: Use the 'is hiring AI roles' or 'mentions SOC2' signals from job boards like LinkedIn or Indeed. • Who Owns it: RevOps or Growth Engineer. • Time Estimate: 8-10 hours of tool setup and API execution. • Pitfall: High API costs. Start with a sample of 500 leads to refine the prompt before running it on the full 18-month history. • Definition of Done: A spreadsheet enriched with at least 3 LLM-derived 'Fit' features that didn't exist in your CRM previously.

3. 03

   Build the hybrid scoring model

   Raw 'Fit' scores are half the battle; you must merge them with 'Intent' and 'Product Usage' data. Create a predictive model using a tool like Mutiny, MadKudu, or a simple Logistic Regression in a Jupyter Notebook. You are looking for behaviors that correlate with conversion, such as 'Visited Pricing Page > 3 times' or 'Invited 5 team members to the product.' • Action: Create a weighted formula. For example: (LLM_Fit_Score * 0.4) + (Product_Usage_Score * 0.4) + (Website_Intent_Score * 0.2). • Setting: If using a tool like MadKudu, map your new LLM fields to the 'Customer Fit' dimension and your Segment/Amplitude events to the 'Likelihood to Buy' dimension. • Who Owns it: Data Scientist or RevOps Lead. • Time Estimate: 10-15 hours. • Pitfall: Over-weighting static fit (e.g., VP at a Fortune 500) over actual intent (e.g., Analyst at a Mid-Market firm doing a deep-dive trial). • Definition of Done: A functioning scoring algorithm that ranks your historical leads from 0 to 100, where the top 10% of scores contain at least 60% of the actual historical conversions.

4. 04

   Calibrate thresholds with Sales

   A model built in a vacuum will be rejected by Sales. Schedule a 90-minute workshop with your Sales Development Rep (SDR) Manager and top-performing Account Executives. Present a 'Blind Taste Test': show them 20 leads,10 that the new model scored high, and 10 it scored low,without showing the scores. Ask them to rank which ones they would prioritize calling. • Action: Adjust weights based on feedback. If Sales says, 'We never close people from this specific sub-industry,' go back to your LLM prompt and add a negative weighting for that industry keyword. • Discussion Point: 'If we send you 50 of these leads per week, do you have the capacity to follow up within 4 hours?' • Who Owns it: VP of Sales and RevOps. • Time Estimate: 2-3 hours. • Pitfall: Setting the 'MQL threshold' too low. It's better to send fewer, higher-quality leads than to overwhelm Sales with 'noisy' volume. • Definition of Done: Signed-off threshold (e.g., 'Any lead with a score > 75 is automatically routed to an SDR').

5. 05

   Deploy to CRM and trigger routing

   Now that the logic is finalized, push the score into your CRM so Sales can actually use it. Create a custom field in Salesforce/HubSpot called 'Predictive_Score_v2'. Use a middleware tool like Zapier, Make, or a native integration to sync the score from your data tool back to the Lead/Contact record. • Config: Set up an Automation Rule. If 'Predictive_Score_v2' > 75, change Lead Status to 'MQL' and trigger a Slack notification to the owner. • Visibility: Add the score and the 'Why' (the top 3 reasons the LLM liked this lead) to the CRM Page Layout. Sales needs to see 'High Fit: Hiring 5+ AI Engineers' to believe the score. • Who Owns it: CRM Administrator. • Time Estimate: 3-5 hours. • Pitfall: Forgetting to map the score from the Lead object to the Contact and Account objects. • Definition of Done: Live scores appearing on new leads in the CRM with automated routing active.

6. 06

   Quarterly audit and re-calibration

   Avoid 'whipsawing' your sales team by changing the model too often. Models need time to collect real-world outcome data. Schedule a quarterly review,not monthly,to audit the model's performance. Compare the MQL-to-SQL conversion rate of 'v2' against your old rule-based 'v1'. • Action: Run a 'Precision/Recall' report. Precision = % of high-scored leads that converted. Recall = % of total conversions that the model successfully caught. • SQL Example: SELECT score_bucket, COUNT(*), SUM(converted_flag) FROM leads GROUP BY 1. • Who Owns it: RevOps Manager. • Time Estimate: 4 hours every 90 days. • Pitfall: Changing the model mid-quarter, which makes it impossible to measure the effectiveness of the Sales team's follow-up scripts. • Definition of Done: A quarterly performance deck showing lead-to-opp velocity improvements and a 'v2.1' tweak list for the next iteration.

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)