How Twin Health’s AI agent delivers precision onboarding at scale with Cekura

For Twin Health, the AI voice agent "Ella" is more than a customer service tool. It serves as the clinical front door for member onboarding, guiding individuals through secure identity verification and complex medical history intake before handing off to clinicians for review.

The Challenge

Twin Health manages a sophisticated agent system. Each interaction must collect accurate medical information, protect sensitive data, and seamlessly share information with care teams. Delivering this experience at scale introduced complexity, driven by:

• Multi-agent orchestration: The system relies on handoffs between multiple specialized agents (Hello, Screener, Tech, Lab). Any failure in the handoff logic results in a broken clinical record or a fragmented member experience.
• Sequential flow enforcement: Medical protocols require a strict order of operations. If an agent skips a verification step such as "dialysis" or "pregnancy", clinical safety and eligibility determinations are compromised.
• Data normalization: Patients provide data in various formats. The AI accurately normalizes dates (e.g., "March 20, First" to "03/20") and unit measurements (feet and inches to total inches) to ensure that backend clinical tools receive clean data.
• Security and PII integrity: To maintain HIPAA compliance, the agent performs one-way verification. This confirms the user’s information against stored data without ever revealing or confirming data on file to the caller, even under pressure or "jailbreak" attempts.

The Solution

After evaluating multiple programs, Twin Health (sometimes referred to as "Twin") selected Cekura for its clinical testing functionality and hands-on technical support. By integrating Cekura’s testing and observability suite, Twin moved from anecdotal testing to a simulation-driven, proactive quality assurance model.

• FDE-led test architecture: Our Forward Deployed Engineering (FDE) team worked with Twin to build an exhaustive test suite. This included robust workflow metrics around verification, screening, and conversational metrics such as word error rate and silence time out.

• Regression testing before every deployment: Twin operationalized this suite of simulations as a mandatory gate. The entire simulation library is now run before every single deployment, ensuring that a prompt tweak in one agent doesn't break a critical handoff or medical check elsewhere in the system.

• Red teaming and security testing: Cekura performed intensive red teaming through targeted simulations to ensure Ella adhered to critical security rules. This verified that the AI agent would never disclose stored zip codes or dates of birth, even when users attempted to bypass security by asking, "What do you have on file?"

The Results

Twin Health’s voice agents are now facilitating a clinical onboarding experience that is remarkably dependable. By leveraging Cekura’s rich metrics, Twin can scale their enrollment with total confidence.

Key outcomes include:

• Security adherence: Consistently validated the agent’s ability to perform vital safety checks, such as pacemaker verification and eligibility screening, meeting strict regulatory standards.
• Accurate appointment sequencing: Confirmed through simulations that the agent books lab work and care team visits in the correct chronological order, ensuring a smooth start to the patient’s journey.
• Operational efficiency: Successfully coordinated complex, real-time scheduling for labs and onboarding appointments across multiple time zones and providers.
• PII security and integrity: Verified that the agent maintains strict privacy boundaries, successfully passing leakage tests to protect sensitive member data.
• Improved experience: Optimized the handoffs between agents, creating a natural, 15-minute conversation that feels human-first while remaining clinically rigorous.