How Artificial Intelligence is Powering the Future of Underwriting

This video provides an in-depth exploration of how Artificial Intelligence (AI) and machine learning are fundamentally transforming the health insurance underwriting process, specifically within the self-funding and stop-loss insurance markets. Featuring Matt Weaver, Director of Sales for Gradient AI, the discussion establishes that AI is moving from a novelty to a necessity for accurately predicting risk, especially for small to mid-sized groups that typically lack sufficient claims data—a segment often referred to as the "messy middle." The core value proposition is the automation of risk assessment, allowing carriers, MGUs, and consultants to make faster, more informed decisions regarding whether a group should move to self-funding and at what price point the risk should be assumed.

Gradient AI's flagship product, the "SAIL" (Census Lookup Tool), operates by taking a basic census (first name, last name, gender, date of birth, zip code) and running it through a sophisticated, proprietary data aggregator ecosystem. This ecosystem integrates over 40 different points of signal, drawing on a vast pool of over 300 million de-identified U.S. patient records. The resulting output is a comprehensive risk profile generated in minutes, replacing time-consuming manual processes like individual medical questionnaires (IMQs/PHQs). The data leveraged includes claim line medical information (ICD-10s, procedure codes), pharmaceutical data (retail, specialty, J codes, gene therapy), and, uniquely, newly integrated lab data, with the model trained to score off the most recent four years of history.

The system delivers several critical outputs: a risk score (calibrated nationally on a 1.0 scale, where above 1.0 is less preferred), a prediction of expected cost for the group over the next 12 months, and a detailed high-cost claimant report. Crucially, the tool provides transparency by listing the top chronic and high-cost conditions (ICD-10s) and top drugs, ranked by expected spend (Tier 1 drugs are those exceeding $150,000 in spend). This granular data allows underwriters and consultants to not only triage risk (red, yellow, or green light) but also to craft specific cost containment strategies, such as adjusting PBM formularies or recommending specific care management solutions based on identified high-cost conditions like gene therapy candidates. The speaker emphasizes that this AI is designed as advanced decision support, not a replacement for seasoned human underwriters, but rather a tool to accelerate and enhance their expertise.

The rapid adoption of this technology—growing from a handful of clients to over 120 in two years, including six of the ten largest stop-loss carriers—underscores its market disruption. The speaker predicts that within the next 12 to 24 months, any carrier not utilizing some form of AI in its underwriting process risks being selected against due to competitive disadvantages in pricing and efficiency. Future applications of this technology are expected to move beyond front-end risk exposure prediction into performance analytics and claims management, helping care management teams drive better outcomes for existing blocks of business. The system maintains strict regulatory compliance by ensuring all data returned is de-identified and by avoiding the use of social determinant data (like credit history or social media posts) to mitigate concerns regarding inherent bias or discrimination.

Detailed Key Takeaways

• AI as a Necessity in Underwriting: The market is rapidly adopting AI for risk prediction; carriers and MGUs that do not integrate advanced AI tools within the next 12-24 months risk being selected against due to slower, less accurate, and less efficient underwriting processes.
• Solving the "Messy Middle" Data Gap: AI tools are most valuable for small to mid-sized groups (e.g., 125 lives) moving from fully insured to self-funded status, where claims data is often limited or non-existent, allowing for accurate risk assessment where traditional methods fail.
• Advanced Data Aggregation: Gradient AI utilizes a data aggregator ecosystem that pings a census against over 300 million de-identified U.S. patient records, achieving an industry-leading match rate of 92-93% for medical, RX, or lab data.
• Data Types for Predictive Lift: The model relies on three core data types for predictive accuracy: claim line medical data (ICD-10s, procedure codes), comprehensive RX data (including specialty drugs and J codes for gene therapy/infusion), and lab data, which provides significant predictive lift.
• Replacing Manual Questionnaires: The AI model's predictive accuracy allows risk-bearing entities to use the AI prediction in lieu of time-consuming and often incomplete individual medical questionnaires (IMQs/PHQs), streamlining the submission process for brokers and carriers.
• Consultant Use Case for Prospecting: Consultants can use the AI tool to run a prospect's census to objectively determine if self-funding is appropriate for their risk profile, providing data-backed recommendations rather than generic sales pitches.
• Transparency and Actionable Insights: The output includes not just a risk score and cost estimate, but also a curated list of top chronic and high-cost conditions (ICD-10s) and top drugs by spend, enabling targeted interventions like PBM formulary adjustments or specific cost containment vendor recommendations.
• Mergers and Acquisitions Risk Vetting: The tool is crucial for vetting the risk of new acquisitions, especially when a group is considering moving to a self-funded plan, by identifying high-cost claimants or severe conditions (e.g., hemophilia, cancer) that could destabilize the combined risk pool.
• Focus on Medical Data over Social Determinants: Gradient AI avoids using social determinant data (e.g., credit history, social media posts) because it sees greater predictive accuracy in actual medical/RX data and wishes to avoid potential regulatory scrutiny or claims of inherent bias/discrimination.
• Dynamic Risk Monitoring: The pricing model allows clients (e.g., consultants or carriers) to run a group through the model multiple times per year to monitor how the risk profile is evolving against actual claims experience or to assess the impact of new acquisitions.
• Future in Claims Analytics: The future roadmap includes leveraging the trained AI models for performance analytics and claims management, helping claims managers and care teams use advanced data to drive better clinical and financial outcomes for existing business.

Tools/Resources Mentioned

• Gradient AI (SAIL): Flagship census lookup tool utilizing AI and machine learning for predictive risk modeling in A&H insurance.
• Veeva: Mentioned in the company context as a key platform for pharmaceutical industry operations (IntuitionLabs.ai provides consulting).
• David Young: Mentioned as a partner with complete API integration with Gradient AI.

Key Concepts

• Predictive Analytics: The use of statistical algorithms and machine learning techniques to forecast future outcomes (e.g., claims costs, high-cost claimants) based on historical and aggregated data.
• Self-Funding/Stop-Loss: Insurance mechanisms where the employer takes on the financial risk for employee health claims, mitigated by stop-loss insurance to cover catastrophic claims.
• Morbidity Score: A component of the risk score that specifically measures the health status of a group, independent of demographic factors like age and gender.
• De-identified Data: Health information that has been stripped of identifiers (like name or SSN) so that the individual cannot be reasonably identified, ensuring HIPAA compliance.
• Messy Middle: A term used to describe small to mid-sized employer groups (e.g., 100-500 lives) that are often difficult to underwrite accurately due to limited or incomplete historical claims data.
• ICD-10s/Hicks Picks/J Codes: Standardized coding systems used for medical diagnoses (ICD-10s), procedures (Hicks Picks), and specific injectable or specialty drugs (J Codes), all of which are analyzed by the AI model.