ElevenLabs Voice Agents Reach Patients in 3.5 Minutes Instead of 2 Hours
ElevenLabs has launched ElevenAgents for Healthcare, a set of voice agents for clinics and insurers. At Cadence, a service for chronically ill patients, an agent calls a patient when the monitoring system flags an alarming reading. The median time to reach the patient dropped from nearly 2 hours to 3.5 minutes. Patients answer these calls 47% more often than calls from nurses. When a patient asks about symptoms, the agent hands the call to a clinician along with the conversation context.
AI-processed from ElevenLabs Blog; edited by Hamidun News
Imagine a service that monitors people with chronic conditions. As soon as the monitoring system flags an alarming reading, such as a spike in blood pressure, someone needs to reach the patient quickly and find out what is going on. Nurses used to do this, and reaching a person took a median of nearly two hours.
Now at Cadence, an ElevenLabs voice agent handles it: the median time to reach a patient has dropped to 3.5 minutes, and patients answer its calls 47% more often than calls from nurses. These are the cases ElevenLabs highlights in its announcement of ElevenAgents for Healthcare, a package of voice agents for hospitals, insurers, and pharma companies.
For developers and product managers, it is an example of how to build a voice agent where mistakes are costly: it needs accurate recognition of terminology and strict limits on behavior.
Why an Agent Loses Trust Easily in Healthcare
A worried person calls a clinic's line and has to read out an insurance policy number or the complicated name of a drug. If the agent asks them to repeat it, interrupts, or misunderstands, the person simply hangs up. So ElevenLabs built the product around two goals: the caller should trust the agent, and the clinic's team should be able to control what the agent says and does, and when it hands the call to a human.
The main new feature is Scribe v3 Live Medical. It is the company's first real-time transcription of medical speech for agents. Previously, ElevenLabs could transcribe clinical conversations only in finished recordings. The new model is built on the general Scribe v3 Live model and tuned for drug names and clinical terms.
Here are the numbers. On doctor-patient conversations from the PriMock57 set, it has the lowest error rate among the medical models tested: 17% lower than the closest one. On the Eka dataset (more than 3600 English medical recordings), it gets 11% of words wrong versus 11.54% for the general Scribe v3 Live model. That is 4.7% fewer errors, so the gain is modest. In addition, all measurements are internal, and the comparison was not against all medical models but against selected ones, each in its best configuration.
Two features are useful in practice. First, you can load up to 10 000 custom terms into an agent: names of insurance plans, therapies, devices, and clinics. Second, if a caller gives a policy number in English but asks the question in Spanish, the model follows the language switch.
Speed, Pauses, and System Integration
ElevenLabs builds speech recognition, the turn-taking model, and voice synthesis itself, so it tunes them as a single system. According to the company, in a study of ElevenAgents deployments, an average 16% reduction in response latency coincided with a 26% increase in the share of successful conversations. This is a correlation, not a proven cause.
The turn-taking model distinguishes the end of a sentence from a pause while the person looks for an insurance card, and it does not cut in. The wait time before a response is configurable: a clinic with elderly patients can make it longer. The Eleven v4 Turbo voice can speak evenly with an irritated customer and more slowly with someone who is anxious about a procedure.
The agent's usefulness depends on access to systems. ElevenAgents connects to scheduling, electronic health records (EHR), CRM, and telephony. At Banner Health (one of the largest nonprofit health systems in the US, with 33 hospitals and more than 400 clinics), agents answer inbound calls around the clock and book, reschedule, and cancel primary care appointments directly in the EHR, with no wait on the line. For legacy and homegrown systems, ElevenLabs offers engineers who work alongside the customer's team. At Angle Health, which administers health plans for small and mid-sized businesses, agents resolve more than 90% of provider calls without handing them to a human. The company also names Pfizer and AstraZeneca among its customers.
Limits of Responsibility and Security
If a patient who has booked a mammogram asks whether they should worry about a symptom, that is a question for a doctor. The agent passes the call to clinical staff together with the content of the conversation, so no one has to start over. Built-in guardrails keep the agent on topic, and custom rules are written in plain language, for example "send any question about symptoms to a human." Before a sensitive action, you can require a step such as verifying the caller's identity. Before launch, the team runs simulated calls and checks that the agent follows the rules.
For data protection, HIPAA-compatible deployment is supported, but only for Enterprise customers who have signed a BAA and enabled Zero Retention Mode: audio and transcripts are not stored on ElevenLabs infrastructure. Call recordings can be sent to the customer's own systems.
For a quick start, there are templates, such as Healthcare Receptionist, which already includes a system prompt, a call flow, and integration settings. The built-in Architect assistant takes a plain-language description ("offer a telehealth visit if the nearest in-person slot is more than a week away"), modifies the agent, and tests the change before launch. A new agent, for example one for billing questions, reuses the integrations that have already been approved, so there is no need for another vendor with a separate security review and BAA.
What to Do With This
The announcement gives no prices or per-minute costs, so you will have to work out the economics with the sales team. Most of the figures come from the vendor's own case studies and internal tests. But the approach applies to any agent in a regulated industry: a custom term dictionary, clear limits of responsibility, handoff to a human with context, testing on simulations before launch, and a single set of integrations for all scenarios.
- Nov 18, 2026AWS shows how to route Bedrock agents to private APIs through a central firewall
- Nov 18, 2026ffmpeg-skill gives Claude Code and Cursor 42 local video-editing tools
- Nov 18, 2026Huang steps down as NVIDIA CEO, company loses more than $1 trillion in market cap
- Nov 18, 2026Airbnb switched coding-agent models — heavy users' spending dropped 7x
Key Questions Answered
Why an Agent Loses Trust Easily in Healthcare?
A worried person calls a clinic's line and has to read out an insurance policy number or the complicated name of a drug. If the agent asks them to repeat it, interrupts, or misunderstands, the person simply hangs up. So ElevenLabs built the product around two goals: the caller should trust the agent, and the clinic's team should be able to control what the agent says and does, and when it hands the call to a human. The main new feature is Scribe v3 Live Medical.
What are the key details on Speed, Pauses, and System Integration?
ElevenLabs builds speech recognition, the turn-taking model, and voice synthesis itself, so it tunes them as a single system. According to the company, in a study of ElevenAgents deployments, an average 16% reduction in response latency coincided with a 26% increase in the share of successful conversations. This is a correlation, not a proven cause. The turn-taking model distinguishes the end of a sentence from a pause while the person looks for an insurance card, and it does not cut in.
What are the key details on Limits of Responsibility and Security?
If a patient who has booked a mammogram asks whether they should worry about a symptom, that is a question for a doctor. The agent passes the call to clinical staff together with the content of the conversation, so no one has to start over. Built-in guardrails keep the agent on topic, and custom rules are written in plain language, for example "send any question about symptoms to a human." Before a sensitive action, you can require a step such as verifying the caller's identity.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.