Insurance
We helped a global insurer accelerate research into hard-to-reach audiences with synthetic personas validated against real customer responses.

A global insurance provider needed to scale customer research across hard-to-reach market segments. Niche audiences in smaller geographies and senior B2B decision makers were difficult to recruit, while four-week research cycles slowed product decisions.
ArcticBlue developed synthetic customer populations grounded in primary research. We created and evaluated more than 60 models representing over 1,500 actual people across international markets. Could these models produce reliable insights quickly enough to inform product decisions?
The validated personas achieved 83% agreement with real customer responses, reduced prediction error by 29%, and cut demographic bias by one-third.
Time to insight fell from four weeks to 45 minutes, giving product teams a faster way to explore concepts and focus subsequent customer research.
The Experiment
We built an iterative framework for training, fine-tuning, and evaluating synthetic personas against primary research. Across more than 60 models, we tested prompting strategies, model architectures, and demographic weighting. Each iteration was compared with real human responses to establish where the approach was reliable and where further research was needed.
Results
83% agreement with real customer responses
29% reduction in prediction error
One-third reduction in demographic bias
Time to insight reduced from four weeks to 45 minutes
Why it mattered
Product teams could explore more ideas between customer conversations and reserve primary research for the decisions that needed it most. Clear guidelines on suitable and unsuitable uses helped teams apply synthetic insights with confidence.
