Candid Health's AI Agents Save 850 Hours a Month and Fully Resolve 59% of Insurer Calls
When an insurer denies a claim for treatment, a clinic employee has to call and find out why. Each call takes anywhere from 10 minutes to an hour. The Candid Health platform has handed these calls to ElevenAgents voice AI agents. In a pilot of 12 860 calls, the agents saved 850+ hours a month and carried 59% of calls through to the end: the reason for the denial and the next step were identified. Takeaway: an agent works well where the conversation follows a clear list of questions.
AI-processed from ElevenLabs Blog; edited by Hamidun News
Imagine a clinic bills an insurer for a patient's treatment, and the insurer denies the claim. To find out why, a staff member has to call the insurer, push through the voice menu, wait for an operator, and ask the same questions every time. One such call takes from 10 minutes to an hour.
The Candid Health platform manages the revenue cycle (RCM, meaning the whole path of money from a submitted bill to payment) for large US healthcare organizations. It has handed these calls to ElevenAgents voice AI agents from ElevenLabs. In the first pilot, the agents made 12 860 calls and saved more than 850 hours of staff work per month.
59% of calls were counted as fully resolved: in them, the agent found out the reason for the denial and worked out what to do next. For anyone planning to build voice agents, this is a rare case with a strict success criterion and a clear division of work.
What exactly was automated, and how success was measured
Candid's agents do not appeal denials; they investigate them. Each agent follows a detailed protocol of questions: why the claim was denied and what step is needed to fix it. The agents also check whether the patient's insurance is active (eligibility). This work frees staff for more complex tasks.
The "resolved" criterion is deliberately strict. A call counts only if the reason for the denial is established and the next step is defined. According to co-founder and COO Doug Proctor, a partial answer does little to help billers, so such calls were not counted. The average length of an agent call is 8 minutes 46 seconds.
There were also side effects. Sometimes the agent's questions led the insurer's operator to notice the error and reprocess the claim. In another case, it turned out that a doctor had not been credentialed with the insurer on the date of service. By the company's estimate, fixing a root cause like this can prevent dozens, or even hundreds, of future denials for a single client.
Who builds the agents and how they are tested
The most practical part of the case is organizational. The agents are built, tested, and launched not by engineers but by Customer Operations and RCM staff. Every insurer has its own phone menu, and the staff tailor each agent's route to it by repeating the steps billers take.
The cycle looks like this:
- the team sets up the agent's path in ElevenAgents; if the agent gets stuck, they ask the built-in assistant Architect how to fix it;
- before launch, the agent calls a live biller who plays the insurer's operator;
- every call is scored against set criteria;
- when the biller hears a failure, they edit the prompts and configuration themselves.
Meanwhile, engineers handle scaling, stability, and monitoring. Transcripts, extracted data, and scoring results arrive directly at Candid through post-call webhooks (automatic delivery of data after a call). The company uses them to prepare client reports, measure agent quality, and refine prompts.
Integration, privacy, and choosing a vendor
Candid's engineers connected ElevenAgents in 48 hours, including telephony, API, and webhooks. Earlier integrations with other providers, they say, sometimes took weeks. Privacy matters in healthcare: in Zero Retention mode, ElevenLabs does not store call data on its side.
Before choosing, Candid compared self-hosted open-weight speech models and other voice vendors. After security, the second requirement was voice quality: if a conversation sounds mechanical, the insurer's operator loses interest, and the call can stall. In side-by-side tests, ElevenLabs voices looked better. Because the company has its own speech and turn-taking models (detecting when to speak), Candid was able to tune tone and pauses. The platform also lets users plug in their own LLM.
What to keep in mind
This is a case study published by the vendor itself. It gives no cost per call, does not name the LLM under the hood, and does not make clear over what period the 12 860 calls were collected or how the 850 hours were calculated. It does not say what happens with the remaining 41% of calls. The agents work with "dozens" of the largest insurers out of more than 1 000 in the country, so coverage is still partial.
Practical takeaways for teams. Define success strictly: 59% on a "reason plus next step" criterion is more convincing than abstract "automation." Give configuration to the people who know the process: billers understand the nuances of these conversations better than developers do. Run the agent through a role-play test with a live employee before launch. And count not only hours saved but also discoveries: an uncovered credentialing problem can be worth more than hundreds of successful calls.
Next, Candid plans to cover more insurers and add new call types, for example credentialing calls.
- Nov 19, 2026OpenCode получил мобильное приложение: пульт к агенту на вашем компьютере
- Nov 19, 2026Skyflow in AgentCore Gateway hides personal data from AI agents' tools
- Nov 19, 2026Accenture Cuts Clients' Bedrock Token Costs by 40–85%
- Nov 19, 2026A Lawyer Taught an AI Agent to Make Surgical Edits in Word with Track Changes
Key Questions Answered
What exactly was automated, and how success was measured?
Candid's agents do not appeal denials; they investigate them. Each agent follows a detailed protocol of questions: why the claim was denied and what step is needed to fix it. The agents also check whether the patient's insurance is active (eligibility). This work frees staff for more complex tasks. The "resolved" criterion is deliberately strict. A call counts only if the reason for the denial is established and the next step is defined.
Who builds the agents and how they are tested?
The most practical part of the case is organizational. The agents are built, tested, and launched not by engineers but by Customer Operations and RCM staff. Every insurer has its own phone menu, and the staff tailor each agent's route to it by repeating the steps billers take.
What are the key details on Integration, privacy, and choosing a vendor?
Candid's engineers connected ElevenAgents in 48 hours, including telephony, API, and webhooks. Earlier integrations with other providers, they say, sometimes took weeks. Privacy matters in healthcare: in Zero Retention mode, ElevenLabs does not store call data on its side. Before choosing, Candid compared self-hosted open-weight speech models and other voice vendors.
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.