Healthcare Dive published on October 7, 2026 an account, from talks at Regal's Rise conference in September, of how the US insurance agency eHealth uses Regal's voice agent Alice "to prescreen Medicare beneficiaries when they contact eHealth's call center."
The workload
"That Medicare enrollment period is about 54 days long," said Atul Kumar, eHealth's VP of product and AI. "We go through maybe half a million calls in those 54 days." Although customers can enroll online, "70% to 80% of eHealth customers buy through the telephone channel," and some callers were waiting six hours for a licensed agent.
How the workflow is built
- Pilot first. The deployment "began in February 2025 with a pilot program": an after-hours agent for calls arriving when the center was closed. Its results shaped the full-time screener.
- The AI screens; humans sell. Alice handles intake, then hands the caller to a licensed human agent for the enrollment conversation itself.
- A fixed step for regulated text. One rule requires "the uninterrupted reading of a lengthy privacy statement." Regal CTO Rebecca Greene: "We developed the concept of a static node where the AI agent could say a specific statement, [that wasn't] LLM driven, and not get interrupted." The agent then asks "Are you still there?" and transfers the caller whatever the answer.
- Messy real calls. Some callers have power of attorney, so two people may be talking to the agent at once.
- Why an LLM rather than a script. Older voice automation required "explicit programming to recognize specific words." Kumar described a caller who never said her husband had died, only "I have the remains of my husband with me"; Healthcare Dive reports Alice "understood what the caller meant and responded empathetically." That flexibility is the reason to use a model -- and the reason the regulated parts were taken away from it.
Results and the adoption constraint
Alice now "answers every call immediately," with 77% "exceptional" satisfaction ratings and "a 27% higher purchase rate" than the outsourced human screeners it replaced -- figures that come from "data Regal provided," the vendor, not an independent audit.
The constraint was compliance and trust. "For any business that's regulated it's not a plug-and-play solution," Kumar said. Staff asked, "What if AI says something that is completely outrageous [or] hallucinates?" eHealth's answer was to make compliance "a key stakeholder" alongside quality assurance, telephony engineers and trainers, use A/B tests, and have people listen to test calls even when the metrics said the agent was working.
Enterprise pattern (analysis): the design worth copying is the split between generated and fixed speech. Anything a regulator dictates word for word -- disclosures, consent language -- runs as a deterministic step outside the model, while the model handles the open-ended intake and the handoff. That limits where a hallucination can do legal damage, and it gives compliance teams a part of the system they can sign off on line by line.
eHealth's Medicare call center shows a production voice agent's real shape: an after-hours pilot first, AI intake with human closers, and regulated disclosures run as a fixed, non-LLM step; the headline results are vendor-supplied, and the main constraint was compliance sign-off, not model quality.