Build a Lead Scoring Model That Understands Your Customers

Small sales teams often start with their CRM's built-in scoring. HubSpot's lead scoring is a rigid point system. A form submission gets +5 points, whether it is from a high-fit prospect or a student downloading a whitepaper. It cannot differentiate intent. Salesforce's Einstein Scoring is more advanced but requires the expensive Enterprise edition and a minimum of 1,000 converted leads to activate, a threshold most small teams have not reached.

Consider a 6-person sales team at a logistics software company. Their best leads come from demo requests that mention specific shipping lanes. Their CRM scores this the same as a generic contact us form fill. A sales rep has to manually read every form submission to find the high-value leads. This manual process means a 4-hour delay in response time, during which a competitor often engages the prospect first. The team is missing its best opportunities because its tools treat all leads equally.

These systems fail because they only process structured data like form fields, page views, and email opens. The real buying intent is in unstructured text: the specific questions in a demo request, the pain points mentioned in an initial email, or the job title in an email signature. Off-the-shelf tools cannot read and interpret this free-text data, leaving the most predictive signals on the table.

Syntora's engagement to develop a custom lead scoring system would typically begin with a comprehensive data discovery and extraction phase. We would work with your team to pull 12-24 months of historical lead, contact, and deal data from your CRM via its API. Unstructured text data, such as call notes within Salesforce fields, Intercom chat transcripts, or email threads integrated via services like Nylas, would also be extracted. Leveraging Python with the Pandas library, Syntora would consolidate this diverse information into a structured dataset for model training, focusing on a robust collection of closed deals.

The core of the system would involve a prompt-driven classification model powered by the Claude API. Syntora would custom-engineer a system prompt to analyze the combined text data for each lead, designed to identify specific signals of purchase intent, budget authority, and timeline relevant to your sales process. This prompt would be optimized to produce a structured JSON output, including a score and a concise rationale, which is critical for reliable downstream parsing and integration into your CRM. Our approach to prompt engineering draws on experience building sophisticated classification systems, such as the product matching algorithm developed for Open Decision.

The scoring logic would be developed as a Python application using FastAPI, providing a dedicated API endpoint that accepts lead identifiers. For deployment, Syntora frequently utilizes serverless platforms like AWS Lambda due to their scalability, cost efficiency, and ease of integration. A typical implementation involves configuring webhooks in your CRM to trigger the Lambda function upon new lead creation or updates, initiating the real-time scoring process.

To optimize performance and manage API costs, Syntora would propose integrating a caching layer, potentially leveraging Supabase. This ensures that if a lead's textual data remains unchanged, a previously generated score can be retrieved quickly without re-running the full AI analysis. For ongoing operational insight, performance tracking and structured logging would be integrated, often using tools such as Grafana for dashboards and Datadog for detailed logs and proactive alerting, providing full visibility into the system's behavior.