← Cases
United StatesFinance and insurance2025 – ongoing

eHealth

Alice, a voice AI agent that screens Medicare callers before a licensed agent

eHealth moved a voice AI agent from an after-hours pilot to full-time Medicare call screening

WorkingVendor case study

Problem

Most eHealth customers buy Medicare plans by phone, and the open enrollment period lasts about 54 days. Call volume peaked at roughly half a million calls. Callers had waited up to six hours for a licensed agent, and staff worked 14 to 16 hour days, including weekends. Screening had been handled by an outsourced human call center.

Approach

eHealth started in February 2025 with an after-hours voice agent built with Regal, a vendor of AI agent software, to take calls when the call center was closed. Lessons from that pilot shaped Alice, an LLM-based agent that now screens Medicare beneficiaries and hands them to a licensed human agent. The team tuned it through trial and error and A/B tests. They also listened to test calls, because quality problems showed up there even when the numbers looked fine. Regal added a fixed, non-LLM step so the required privacy statement is read without interruption.

Results

77%share of callers who gave Alice an exceptional rating, per data from RegalVendor case study[1]
27% higherpurchase rate than under the outsourced human screening call center, per data from RegalVendor case study[1]

Takeaway

Alice's first versions screened callers efficiently but sounded short on empathy, so eHealth retrained them. Listening to calls caught a problem that the metrics did not.

Categories

Innovation type
Process innovation
Evidence
Vendor case study

Sources

Source summaries

[1]Healthcare Dive7 Oct 2026

The article follows eHealth from a February 2025 after-hours voice AI pilot to a full-time Medicare screening agent built with Regal. It gives call volume and working conditions from eHealth's VP of product and AI. The caller rating and purchase-rate figures come from data Regal gave to CX Dive. It also describes the tuning work, including A/B tests and a fixed privacy-statement step.

Full details are in the original sources.

Added 3 Nov 2026 · Updated 3 Nov 2026
Open case page ↗