How CVS Health is exploring digital twin technology

Twin, where have you been? Apparently at CVS Health, where some consenting individuals are helping its customer experience and insights team make decisions about digital products.

CVS Health is one of the latest companies taking digital twin technology for a test drive. It currently leverages over 100,000 digital twins based on real people, which it refers to as “agentic twins.” The technology is designed to help employees make decisions about the customer experience.

Agentic twin, you say? IT Brew caught up with Sri Narasimhan, VP of enterprise customer experience and insights at CVS Health, to learn about the value of its agentic twins.

Historically, conducting a research study involved finding a pool of people, conducting focus groups, and following up—a lengthy process that could take four to six weeks. Now, Narasimhan said, CVS Health has AI agents representing customers that teams can prompt to understand how services and products are working.

“Now, we’re talking about 15 to 30 minutes to run relatively compact studies on things like messaging or product testing, [and] A/B testing, in terms of how consumers will respond to things that we’re putting in the market,” Narasimhan said. In addition to speed, the agentic twins also give CVS Health the ability to simulate research studies with hard-to-reach populations, such as immunocompromised individuals, and allows the company to conduct research at scale.

“Instead of reacting, we can be really proactive in terms of how we’re offering services,” Narasimhan said, adding that the company’s agentic twin simulations, when compared to traditional research studies, have an 85%–95% accuracy rate.

Double act. While CVS Health has leveraged traditional machine learning models for years to model customer behavior, Narasimhan said the use of agentic twins takes things to the next level: “It’s really getting at not just predicting what they’ll do, but why.”

CVS Health partnered with AI simulation company Simile to craft its agentic twins, which were built on almost 3 million consented responses from more than 400,000 people (CVS Health Ventures, the venture capital arm of CVS Health, is an investor in Simile). Narasimhan told us that the responses were retrieved via AI-moderated interviews.

“So, we’re looking at a particular population of people. Let’s say people that fill prescriptions,” Narasimhan said. “We would ask them detailed questions around their experiences there. We pulled demographic information that they shared, including things like political affiliation and stuff like that.”

Govern up! Governance is crucial when leveraging digital twins that replicate human behaviors, according to Ricardo Mayerhofer, co-founder and VP of engineering at Hopara, a digital twin technology company.

“The challenge becomes what you can infer on the real people based on the agentic digital twin? And that poses a privacy challenge,” Mayerhofer said. “Can you learn about their daily preferences, future plans, [or] weaknesses? And because all this is new, it’s also a challenge for all companies, even the ones on the edge of the technology right now.”

Narasimhan said CVS Health practices good governance by keeping humans in the loop. Part of that means resisting assumptions that the agentic twins may be wrong when results are different than expected.

“That doesn’t mean discount it. What it means is, investigate it,” Narasimhan said. “It’s very human-led in terms of we want to make sure that we have the right data and the right information coming through.”

There’s a focus on how agentic twins are prompted, he added, given they behave in human-like ways: “There is a lot of governance here around my team leading it, because if you just opened it up to a broad set of…people who didn’t understand that, you could make it say something, much like you could with a human population.” That means providing twins with the “right context” around questions they are being asked.

The more, the merrier. CVS Health currently intends to increase its agentic twin population. While the company has used the twins for “research-driven” use cases, Narasimhan said it plans to use them for more “dress rehearsal” purposes, like testing capabilities and offerings in a confined environment before looping in real humans.

“We’re able to do that, get their responses, how these things are effective for them, where we need to improve them,” he said. “And then when we actually go out to live humans, we have way more confidence that we’ve done it correctly.”