Insurance
We helped a global insurer replace weeks of customer research with AI-generated insights grounded in real-world evidence.
A global insurance provider wanted to understand how different customer segments would respond to new financial planning and life insurance concepts, but traditional research was too slow to support rapid product decisions.
ArcticBlue built and validated synthetic customer personas using multiple generations of large language models, calibrated against real customer research. Rather than asking whether AI could replace research, we focused on a more practical question: Could AI become a reliable decision-support tool for product teams?
After multiple rounds of experimentation and validation, the synthetic personas achieved 83% agreement with real customer responses, while significantly reducing prediction error and demographic bias.
The result was a more accurate model along with guidelines on suitable use cases (and critically: unsuitable use cases). The result gave product teams a faster way to test ideas before investing in large-scale research.
The Experiment
Instead of evaluating a single model, we iterated across multiple prompting strategies, model architectures, and demographic weighting techniques. Each version was compared directly against responses from real customers. The objective was to maximize confidence that product teams could trust the output when making business decisions.
Results
83% agreement with real customer responses
29% reduction in prediction error
One-third reduction in demographic bias
Hundreds of product scenarios evaluated in a fraction of the time required for traditional research
Why it mattered
Synthetic research doesn't replace customer insights or traditional market research. It allows organizations to learn dramatically faster between customer interactions, making every future research investment more focused, less expensive, and higher quality.
