Case Study: Fetch Pet Insurance Drives Improved Customer Retention & Growth with Predictive AI

Fetch Pet Insurance is a U.S.-based pet insurance provider covering over 800,000 insured dogs across the United States and Canada.

With a rapidly growing pet care market forecasted to reach USD 68.91 billion by 2032 (1), Fetch recognized an opportunity to improve customer engagement and retention by offering its customers personalised, data‑driven health insights for their beloved pups.

Working closely with Spryfox, Fetch successfully developed and launched the Fetch Health Forecast - an innovative, Al-powered online platform enabling pet owners to predict the future health of their dog, plan the associated care costs and proactively ensure their pet's well-being.

The Opportunity and Challenge

The market growth for this sector reflects not only the rising number of pet owners, but a significant cultural shift: pets are increasingly viewed as family members (1). Consequently, Fetch wanted to differentiate themselves by offering services that go far beyond traditional risk coverage.

The result was an advanced AI solution capable of predicting the likelihood of 48 different disease groups for individual dogs.

1. Data-first approach

A model is only as good as the data that feeds it. Many AI projects fail because of data and infrastructure problems that are often detected too late. Spryfox conducted data and infrastructure audits early in the project which helped build a strong foundation. Multiple data types were consolidated including:

• Insurance and medical history: Prior illnesses, claim frequency, and time since last illness
• Breed-specific factors: Genetic predispositions, size, temperament, and other breed traits
• Environmental factors: Climate, demographics, and socioeconomic data based on location

These data streams were unified through a custom-built medical ontology (a structured graph mapping claims, diseases, symptoms and groups - developed in collaboration with veterinary experts. This ontology makes it possible to organise the data in a way that enables more structured learning and precise predictions.

2. Advanced AI Modelling

A big part of the work to make data AI ready is to translate it to the right features. To ensure scientific correctness and explainability this step was done in collaboration with veterinary epidemiologists. Also, to ensure an objective evaluation of the model, the data was divided into training and test data.

The aim of this separation is to train the model on one part of the data and evaluate its performance on another, unseen part.

Spryfox designed a robust machine learning pipeline:

• Identification of relevant data points that have most predictive power
• Train multiple models and combine them to profit from their individual strengths
• Explainable AI (XAI) techniques to ensure transparency and trust, allowing veterinary experts to validate and refine model outputs

Care was taken to ensure that the distribution of important characteristics such as breed, age and previous illnesses was similar in both data sets in order to avoid bias. An even distribution of these characteristics is crucial to ensure that the model can generalize to different scenarios and is not only adjusted to certain subgroups of the data. 

3. Training & Evaluation

Spryfox implemented an advanced training and evaluation approach:

• Divided data into distinct time windows, separating past health history (e.g. first four years of a dog’s life) from future prediction periods (e.g. the fifth year)
• Generated multiple training scenarios from each dog’s timeline to expand the dataset and improve learning
• Built a simulation engine that predicts a variety of future scenarios
• Evaluated a range of machine learning models. For each of the models, an explainability layer was delivered that allowed veterinarians to understand the decision making of the algorithm
• Benchmarked AUC scores. The final model achieved an AUC of 81% which is a significant performance in the medical field. The approach got recognized by a publication in Nature Scientific Reports. (2)
• Patented the AI algorithm as well as visualization elements

This methodology enabled the model to learn from a larger pool of examples, significantly increased predictive accuracy and built trust with veterinary experts. The explainable approach ensured that AI-driven insights were medically meaningful and scientifically sound.

The best model achieved an AUC of 81%, reflecting strong predictive power and reliability for real-world use.

4. AI Implementation

Spryfox deployed the model for real-world predictions using a set of tools and approaches that helped moving very fast from prototype to production. This included:

• Automated infrastructure for AI models
• Continuous model monitoring, drift and bias detection
• Extensive testing framework to test all corner cases