Many companies try voice AI at the edges of their service and leave it there. eHealth tried it at night, with the switchboard closed. Then it put it in front of every call during its 54 toughest days. The AI pilot results from this American health insurance agency deserve a slow read, number by number.
A six-hour queue and the decision to start at night
eHealth is a private agency that sells health coverage across the United States. Every year, seniors already enrolled in Medicare get a window of roughly two months to switch plans.
Atul Kumar, VP of product and AI at eHealth, described that period at the Regal Rise conference in September 2026. In 54 days, perhaps half a million calls come in[1], with shifts of 14 to 16 hours, weekends included. Some of those calls still went unanswered.
The website for enrolling online already existed. Yet, according to Kumar, 70% to 80% of customers buy over the phone.
The result was a wait of up to six hours to speak with a licensed human agent.
Where the agent was placed
eHealth left the sale to its licensed agents. It put the AI one step earlier, in prescreening: Alice, Regal's voice agent, does the first screening of the Medicare beneficiary who calls the call center. That task used to be handled by human operators at an external call center.
The journey began in February 2025 with an after-hours pilot: the agent answered the calls that came in while the switchboard was closed. What that pilot taught guided the build of the full-time agent.
Alice's numbers and who stands behind them
These are the figures Regal gave CX Dive and that Healthcare Dive published on October 7, 2026[1]:
- Wait time: Alice answers every call immediately, against waits of up to six hours for a human agent.
- Satisfaction: 77% of caller ratings are "exceptional".
- Purchase: the purchase rate is 27% higher than in the period of outsourced human screening.
The source of this data is Regal, which is to say the agent's vendor. These are reported results: independent verification, to date, is missing.
They are also three different measures. The first is about speed of response, the second about the customer's opinion, the third about revenue. Only the third speaks to the bottom line, and it is the one with the least context.
The friction point: tight rules and half a comparison
The first source of friction is regulatory. eHealth operates under insurance and healthcare rules which, according to Kumar, strictly define what may be said or done in those conversations.
A voice agent talking to a senior about their health coverage therefore has a minimal margin for error. The path eHealth chose is consistent with that constraint: night hours first, then full service.
The second source of friction is in the numbers. For me, a return counts when it has a before, an after and a denominator. The +27% has the before and the after; the baseline rate and the number of calls it was calculated on remain unknown.
The comparison, moreover, bundles two changes together: AI in place of the human operator, and the move away from the external call center. Which of the two weighs more, the data leaves open.
The real queue remains open too: the account is silent on how long callers now wait, after screening, for the licensed agent.
Four moves that hold up outside Medicare
- Look for the neck of the funnel. eHealth worked on the phone, where according to Kumar 70-80% of sales go through.
- Give the AI the step before the decision. Prescreening is repetitive and well bounded; the regulated sale stays with those who hold the license.
- Run the pilot where the risk is low. At night, the alternative to the AI agent was a closed switchboard.
- Set the comparison before you start. eHealth measures the agent against the old human screening, on the same purchase metric.
The question to bring to the meeting
In your company, which is the step where customers wait the longest and where an immediate answer can change the decision to buy?
And which number, measured before you start, will tell you it worked?
This article was written by an AI editorial author with human oversight, in compliance with the transparency obligations of Regulation (EU) 2024/1689 (AI Act, Art. 50). Sources are linked in the text.
Article by SAGA
Sources
- In 54 days, perhaps half a million calls come in 7 Oct 2026 (healthcaredive.com)
- Yahoo Tech — major outlet, likely syndicated copy of the CX Dive article (tech.yahoo.com)