How CVS Health is exploring digital twin technology

Open original

CVS Health uses over 100,000 AI-driven “agentic twins” to run customer research in 15–30 minutes instead of weeks, achieving 85–95% accuracy and reaching hard-to-access groups. Built with Simile from 3 million consented responses, these simulations remain human-overseen to ensure proper prompting, validation, and governance before live rollouts.

  • CVS runs 100,000+ AI "agentic twins" built from 3 million consented survey responses to simulate customer behavior, cutting research time from 4-6 weeks to 15-30 minutes with 85-95% accuracy versus live studies
  • The twins let CVS probe hard-to-reach populations (like immunocompromised patients) and understand not just what customers do but why, using demographic and behavioral details tied to each twin
  • Human oversight remains mandatory: leads review all prompts/results and investigate any deviation from expected outcomes to guard against bias and privacy risks
  • CVS plans to expand the twin population for "dress rehearsals" of new services before real-world rollout

CVS Health is running more than 100,000 “agentic twins”—AI models built from nearly three million consented survey responses gathered via AI-moderated interviews with over 400,000 people—to simulate real customer behavior. By prompting these digital twins, the customer experience and insights team can run compact studies on messaging, product testing and A/B tests in as little as 15 to 30 minutes, down from the four to six weeks a traditional focus-group approach would take. CVS reports that these simulations hit an 85–95 percent accuracy rate compared with live studies, and they open up research on hard-to-reach groups like immunocompromised patients. Before agentic twins, CVS used standard machine-learning models to predict customer actions. Now it can explore not just what customers will do, but why. The twins were developed in partnership with Simile, a simulation-AI startup partly backed by CVS Health Ventures. Each twin carries demographic details—age, prescription habits, even political affiliation—and answers detailed questions about experience, letting teams tweak wording or features before launching to real customers. CVS also puts strict governance in place. Human leads review all prompts and results to prevent bias or misinterpretation. If a twin’s response deviates from expected outcomes, researchers don’t discard it—they investigate. This hands-on oversight guards against privacy risks, says Hopara co-founder Ricardo Mayerhofer, because agentic twins could otherwise reveal individual preferences or plans. Looking ahead, CVS plans to grow its agentic-twin population and use it for “dress rehearsals” of new services. By testing in a confined virtual setting, product teams refine user flows and identify problems early. That way, when new offerings roll out to customers, CVS runs on data-driven confidence rather than guesswork.

Ask about this article