- Progressive described 14 billion telematics miles and 20+ years of driving data from its Snapshot program, and used H2O for four model families: customer churn, billing, fraud detection and threat analysis.
- The Progressive figures describe its program and its use of H2O, not a published loss-ratio or ROI result.
- H2O Driverless AI and H2O-3 build explainable models from trip data joined to policy and claims history, with reason codes on every prediction.
- The models sit on top of your policy and claims systems and feed your existing fraud screening, so the core you run today stays in place.
Turn driving data into a score your claims and SIU teams can use
Telematics gives you trip-level data on how, when and where each policyholder drives. The volume grows every day, and most of it never reaches the claims desk or the special investigations unit (SIU).
You are measured on claims fraud leakage and on the quality of every referral your SIU works. Fixed screening rules catch the patterns you already know, and each false referral costs an investigator's time.
Anything you build also has to hold up with your actuaries, your compliance team and your regulators. A score nobody can explain will not be used on a claim, and it will not survive a rate review.
How telematics-based insurance models work
Your data science team builds explainable models on driving data, scores claims and policies on a schedule or at first notice of loss, and sends the scores into the systems your adjusters and investigators already use.
Step 1: join trip data to the policy and the claim
Start with trip summaries from your telematics program, such as speed, braking and time of day, joined to policy, billing and claims history. Your consent terms and privacy policies decide which trip data a model may use.
Step 2: build features from how people drive
H2O Driverless AI runs automatic feature engineering on the joined data, for example changes in a driver's own patterns or a gap between the reported loss and the trip record. H2O-3 handles large volumes of data for teams that build in code.
Step 3: train the model families you need
Train classification models for claims fraud, churn and billing risk from your own outcomes. Anomaly detection adds a view for claims whose trip data matches no normal pattern, and each model works beside your actuaries' generalized linear models, not against them.
Step 4: score, explain and route
Each claim or policy gets a score with reason codes, so an adjuster or investigator sees which factors drove it, and your rules decide the routing: fast-track, request more information or refer to the SIU.
Step 5: document, deploy and monitor
Model documentation is written while the model trains, and H2O MLOps deploys the scoring pipeline, tracks model drift as driving patterns change and keeps an audit trail.
Results
Progressive, a current customer, described 20+ years of driving data and 14 billion telematics miles from its Snapshot program, and used H2O to build four model families: customer churn, billing, fraud detection and threat analysis. These figures describe Progressive's program and its scale; no loss-ratio or ROI figure for this work has been published.