eHealth puts voice AI at the front line of Medicare enrollment calls
When Medicare open enrollment arrives, eHealth faces a call surge that would strain most contact centers. The company says the most intense stretch lasts about 54 days, during which it can process roughly half a million calls and still fail to reach every customer before the window closes. In that setting, even a six-hour wait to reach a licensed human agent becomes more than an operations problem. It becomes a barrier to enrollment at the exact moment older adults are trying to compare plans, preserve coverage, or make a change that affects their next year of care.
The response was to move voice AI from a limited pilot into the center of the workflow. eHealth now uses Regal’s Alice, an AI voice agent, to prescreen Medicare beneficiaries when they contact the call center. The system answers immediately, handles the first layer of the conversation, and then routes the customer into the right next step. In the data Regal provided to CX Dive, Alice produced 77% “exceptional” satisfaction ratings and a 27% higher purchase rate than eHealth had seen when human screeners worked through an outsourced call center.
Those numbers point to a simple operational truth: in a high-volume, regulation-heavy environment, speed alone is not enough. The call still has to feel respectful, calm, and competent.
A pilot built for after-hours calls became the working model
The deployment began in February 2025 as an after-hours pilot, covering calls that arrived when the eHealth center was closed. That first use case mattered because it limited risk while still exposing the system to real customer behavior. The learnings from that test shaped the full-time screener that followed.
Medicare enrollment calls are not routine retail transactions. They are often emotional, information-dense, and personal, which means a voice system must do more than identify intent. It has to handle silence, confusion, interruptions, and grief without sounding mechanical. Atul Kumar, eHealth’s vice president of product and AI, described one caller who said, “I have the remains of my husband with me.” Alice interpreted the statement as a sign the caller had lost her spouse and responded with empathy rather than requiring a rigid scripted phrase.
Legacy voice automation often depends on explicit word spotting and narrow branching scripts. A language model-based agent can infer meaning from context, which is especially important when older adults speak indirectly, pause mid-thought, or use unusual phrasing. In a Medicare call center, that flexibility is not a cosmetic feature. It is the difference between a system that merely processes speech and one that can sustain a humane interaction long enough to preserve trust.
Compliance rules forced the system to tolerate interruption
Regulated healthcare and insurance conversations bring a second layer of complexity. eHealth must stay within strict rules about what it can say and do during enrollment calls, and one requirement involves uninterrupted reading of a lengthy privacy statement. In practice, callers often interrupt the disclosure because they want to move quickly to the part that matters most: whether they qualify, what they can afford, and how to reach a licensed human who can help them compare plans.
The interruption problem could have broken the workflow. Instead, the team and Regal built a system that could absorb it. Rebecca Greene, Regal’s co-founder and chief technology officer, described the use of a static node — a fixed statement the AI could deliver without relying on live LLM generation and without losing control when the caller cut in. After the disclosure, the agent asks, “Are you still there?” and, regardless of the response, transfers the caller back to the human agent for the substantive enrollment conversation.
That design choice says a lot about where AI fits best in regulated care-adjacent operations. It can accelerate the front end of the interaction, manage repetitive compliance language, and keep the line moving. It cannot be allowed to improvise around policy boundaries.
Two people on one line can complicate enrollment even further
Self-enrollment is not always a one-person event. Some customers call on their own behalf, while others speak with power of attorney. In those cases, the voice AI may need to navigate two speakers, two intentions, and two different roles in the same call. The system has to decide who is speaking, who has authority, and when to hand off to a human who can resolve the enrollment correctly.
Trial and error was unavoidable. Kumar said the team relied on A/B testing and qualitative review, including simply listening to calls to judge whether the interaction sounded right even when the metrics suggested the agent was functioning as intended. That kind of hybrid validation matters in voice care workflows because a technically successful call can still sound cold, confusing, or unnatural. A contact center handling older adults cannot afford a system that is accurate on paper but off-putting in practice.
Human review remained part of the quality bar
The development team had a practical advantage: it knew from direct experience what a good peak-season call should sound like. Kumar assembled an all-hands response to the open-enrollment crunch, and the people tuning Alice had lived through the operational pressure firsthand. Greene said that familiarity helped the agent improve quickly. It is one thing to optimize for abstract call metrics. It is another to tune the agent against the cadence, patience, and reassurance real Medicare callers expect when they are already stressed about coverage choices.
Stakeholder fear centered on hallucination and brand risk
Inside the organization, enthusiasm for AI came with caution. The most obvious worry was whether the system might say something outrageous or hallucinate under pressure. In a Medicare setting, even a single inappropriate phrase could create compliance exposure, undermine trust, or derail a sale. Kumar said the team brought compliance in as a key stakeholder from the start, alongside quality assurance, telephony engineers, customer experience managers, and trainers.
That cross-functional setup was not just about governance. It was a way to make the system legible to the people responsible for protecting the company and the customer at the same time. Kumar also played calls for senior leaders so they could hear how Alice actually behaved with beneficiaries. In a workflow involving older adults, privacy statements, and enrollment decisions, auditory reassurance carries real weight. A dashboard can show success. A call recording shows whether the experience still feels like a conversation instead of a machine dragging a customer through a script.
The result is a voice AI model built less like a fully autonomous agent and more like a carefully bounded intake layer. It answers immediately, keeps callers from waiting for hours, absorbs the most repetitive compliance burden, and passes the rest to a licensed human. The improvement in satisfaction and purchase rate suggests that structure can work when the operational pain point is delay rather than clinical judgment.
What remains unresolved is how far that model can stretch across other regulated health insurance functions, or whether the same performance will hold when call complexity rises beyond Medicare enrollment. The article does not provide a broad cost analysis, long-term retention data, or a breakdown of outcomes by caller type, so the full operational and consumer impact is still only partially visible.