Article ·
AI Voice Agents in Medical Practices: The 81% Problem
AI voice agents are taking the after-hours calls specialty practices miss, and the transfer to a person is the safety feature nobody publishes numbers for
Contents10 sections
Nobody picks up after five
Specialty practices miss most calls that come in after the front desk closes. AI voice agents are moving in to answer them. A voice agent is an AI system that holds a spoken phone conversation and can act on it, booking or moving an appointment. Banner Health has put one on its phones, and two startups that sell them raised money in October 2026. What makes any of them safe on a medical line is the handoff to a person when the call is urgent. None of the public launches published a number for that handoff.
Where the 81% comes from
Prosper AI, a company that sells AI phone answering to practices, put the number in its 2026 growth report: "18% of calls go unanswered during office hours and 81% after hours." After-hours callers, by that count, are four and a half times as likely to reach nobody.
The method sits in the report's PDF rather than on its web page. Prosper placed 31,712 outbound test calls to 1,994 practices in 29 states and six specialties. That's about 16 calls each, at fixed local times, with nothing booked. It measured how long a live person took to answer. The chart defines a missed call as one "never answered by a person." Voicemail counts as unanswered.
Three cautions follow. Prosper sells the cure for the problem it measured. The report gives no dates for the calls, only that its data was "refreshed Sept 29, 2026." And the hour-by-hour chart stops at 5 p.m.
That last one matters most. The after-hours figure, the one in the headline, has no documented basis of its own. Even the daytime rate drifts: about 17% in the chart, 18% in the text.
So the honest reading is narrower than the headline. A vendor's test calls found that most after-hours calls never reached a person, and the exact share isn't established by anything independent.
What's being deployed
Three deployments carry the story, and all three publish their own figures.
Banner Health and ElevenLabs. On 2 October 2026 ElevenLabs described Banner's rollout of its voice agents, starting with primary-care scheduling. The agents answer at any hour with no hold queue. They book, move and cancel appointments directly in Banner's electronic medical record (EMR). A caller who needs a person gets a warm transfer: a Banner team member picks up with the context of the call already in hand. Banner's chief AI officer described "human escalation built into every conversation, full HIPAA compliance, and real-time oversight." HIPAA is the US law that sets privacy and security rules for health information. The post gives no numbers. There are no call volumes, booking rates or wait times, and nothing on escalation.
Vocca. On 6 October 2026 Vocca announced a $20 million Series A led by Norrsken VC. Its AI phone assistant works across more than 20 specialties in the US and Europe. Vocca reports 15,000 practitioners and more than a million conversations a month. It says 70% of conversations end with no person involved, and over 85% at its most mature customers. Those are Vocca's numbers. One customer, the Fondation Rothschild hospital in Paris, describes urgent calls going straight to its front-desk staff while the agent logs complex, lower-priority requests so that a member of staff can phone the patient back later.
Confido. On 14 October 2026 Confido Health announced a $12.5 million Series A extension led by Z47. It began with voice agents for patient calls and now automates scheduling, insurance checks, refills and payments. Confido reports 3,200 providers at more than 1,500 sites in 27 states. Its AI handles more than 95% of key workflows, it says, and patients score it 4.55 out of 5 on average. Its executives cite "15x growth since its last round" without naming the measure. Neither Inc42 nor Fierce Healthcare says how a task reaches a person when the agent can't finish it.
Escalation is the safety feature
A voice agent that books well is a convenience. What makes it safe on a medical line is what it does with the call it shouldn't keep — the caller with chest pain, the parent describing a child's breathing, the patient who stopped a medicine yesterday. The design that matters is a short, reliable path to a person.
Three properties separate a real escalation path from a nominal one.
- It fires on the right calls. Spotting urgency is the hardest part. It fails silently: an urgent call the agent reads as routine never reaches anyone.
- The caller can always ask for a person. The request works at any point and ends in a human voice, never in a loop back to the menu.
- The person gets the context. Banner's warm transfer has the right shape: the caller doesn't start again, and the person knows why the call came to them.
The metric vendors lead with pulls the other way. "Resolved without a person" measures efficiency; pushed far enough, it rewards an agent for keeping calls it should have passed on. A practice that buys on that number alone may be paying for the wrong behaviour.
What the launches leave out
The public numbers describe volume and efficiency. The numbers that would show whether escalation works exist inside every deployment, and none of the three launches printed them.
Five questions belong to any vendor, and to any practice running a pilot.
- How many calls were urgent, and how long did each take to reach a person?
- How many urgent calls did the agent keep, as found when someone reviewed the recordings?
- How many callers hung up, gave up or never got the care they called about?
- How often were bookings, messages or details wrong?
- Who reviews a sample of calls, how often, and what happens when they find a miss?
Where MARA sits
MARA, the Medical Agentic Resident Assistant this project is designing, takes Banner's pattern as a rule. Every voice conversation she holds has a path to a person. Whoever picks up sees the transcript and the reason for the transfer. Her escalations are counted and reported, including the ones that should have happened and didn't. MAX4, the desktop resident where her methods are proved first, has to show two things. An observer can stop or take over any conversation. Every handoff lands in the log.
What nobody knows yet
Prosper's after-hours basis is undocumented, and its daytime figure is its own test calls rather than real patients. Banner published no metrics at all. Vocca's and Confido's figures come from the companies; nobody outside has checked them. And no public source yet says how often any of these agents kept an urgent call.
Why it matters
A missed call at nine at night used to end in voicemail, and a person heard it in the morning. An agent that answers instantly replaces that delay with a judgement, made in seconds, about whether the call is urgent. When the judgement is right, the caller is better served than voicemail ever managed. When it's wrong, the call is closed, logged as resolved, and counted toward the very number the vendor leads with. A practice choosing an agent is choosing whose count of its misses it will believe.
Sources
- Prosper AI. Specialty revenue growth report 2026, web page and linked PDF, "refreshed Sept 29, 2026." Opened 21 October 2026. https://www.getprosper.ai/report/specialty-revenue-growth-2026
- ElevenLabs. Banner Health. 2 October 2026. https://elevenlabs.io/blog/banner-health
- Vocca. Vocca raises $20M Series A led by Norrsken VC. 6 October 2026. https://vocca.com/en/resources/blog/vocca-raises-20m-series-a-led-by-norrsken-vc
- Inc42. Confido Health raises $12.5 Mn to automate healthcare administration with AI. 15 October 2026. https://inc42.com/buzz/confido-health-raises-12-5-mn-to-automate-healthcare-administration-with-ai/
- Fierce Healthcare. Confido Health nabs $12.5M to scale agentic AI across PE-backed medical groups. October 2026, opened 21 October 2026. https://www.fiercehealthcare.com/ai-and-machine-learning/confido-health-expands-beyond-voice-ai-agentic-platform-medical-practices
Glossary
| Term | Definition |
|---|---|
| After hours | evenings, nights, weekends and holidays, when a practice's front desk is closed. It's when Prosper's test calls found the fewest people answering. |
| EMR (electronic medical record) | the clinical record system a practice runs on. Banner's agents book straight into it, so a booking error lands in the real schedule. |
| Escalation | passing a conversation from the agent to a person. It's the safety feature this piece measures the launches against. |
| HIPAA | the US law that sets privacy and security rules for health information. Banner says its agents comply with it. |
| MARA (Medical Agentic Resident Assistant) | this project's design for a clinical assistant that recommends while a person decides. Her voice conversations always carry a path to a person. |
| MAX4 | the desktop generation of the MAX AI resident, from October 2026 on, where MARA's methods are proved first. |
| Resolved end to end | a call the agent finished with no person involved. Vendors lead with it, and pushed too far it rewards keeping calls that should have been passed on. |
| Series A | a startup's first large round of venture funding. Vocca and Confido both raised one in October 2026. |
| Voice agent | an AI system that holds a spoken phone conversation and can take actions such as booking an appointment. |
| Warm transfer | a handoff in which the person who picks up already has the call's context, so the caller doesn't start again. |
Keep reading
More field notes
This piece is part of the MAX Research Collective library. Browse the rest, or connect on LinkedIn.