Liberty Mutual

In commercial insurance, claims decisions have long depended on adjuster intuition and standardized checklists rather than systematic data analysis. For a carrier of Liberty Mutual's scale — processing millions of claims annually — this approach created significant blind spots. Unstructured data sources including images, field notes, recorded calls, medical records, and investigative reports were largely ignored, despite containing signals critical to detecting fraud and flagging complex claims early. Without models to surface elevated-risk claims in the first weeks of a claim's lifecycle, high-cost cases were often misclassified at intake, delaying intervention and inflating reserves. The gap between available data and actionable insight was costing the business in recoveries, cycle efficiency, and fraud exposure.

Liberty Mutual deployed natural language processing and predictive machine learning across a dataset of 200 million data points drawn from five million claims — one of the largest internal claims corpora in the industry. The system ingests unstructured sources (adjustor notes, images, call transcripts, third-party records) alongside structured claim data to power a suite of next-generation models covering claims complexity segmentation, fraud scoring, and compensability prediction. Rather than replacing adjusters, the models function as an augmentation layer, surfacing actionable signals so specialists can direct their judgment where it matters most. To extend these capabilities into emerging AI disciplines — computer vision, language understanding, and risk-aware decision making — Liberty Mutual formalized a $25 million, five-year research partnership with MIT, ensuring continued access to frontier research and talent.

The predictive modelling programme delivered measurable improvement across every targeted outcome:

• Doubled identification of elevated-risk (>$100k) claims by day 30 of the claim lifecycle
• 4x increase in annual recoveries with no corresponding rise in cycle time or operational expense
• Automated fraud model performs 20x better than random chance at flagging fraudulent activity
• Claims with potential compensability issues predicted at a 60% accuracy rate
• A dual-strategy model yielded a $20k reserve reduction on a single claim by prompting specialists to evaluate alternative resolution paths at outset

Qualitatively, the programme shifted claims management from checklist-driven assessment toward an evidence-based workflow, freeing adjusters to focus on customer empathy and complex judgment calls.

• Unstructured data is an untapped asset: Notes, images, and call records contain fraud and complexity signals that structured data alone misses — NLP processing at scale is what unlocks them.
• Early identification is where value concentrates: Doubling high-risk claim detection by day 30 shows that intervening earlier in the lifecycle drives outsized recoveries and reserve accuracy.
• Augmentation outperforms replacement: Presenting model outputs as decision support — not automated decisions — accelerates adjuster adoption and preserves judgment quality on edge cases.
• Academic partnerships fill capability gaps: Proprietary data science has limits; a structured research partnership (as with MIT) provides access to frontier methods that internal teams cannot develop alone.
• Model portfolio beats single-model thinking: Deploying separate models for fraud, complexity, and compensability targets each problem with the right feature set, yielding better precision across the board.