Susan: AI-Powered Outbound Voice Agent

Case Study

Introduction

This demonstration validates a crucial insight for healthcare organizations: custom AI voice agents can now be built using commodity tools and modern LLMs, eliminating the need for expensive, specialized vendor platforms.

Susan is an AI-powered outbound voice agent capable of conducting natural conversations with patients to gather health information. The demonstration proves that organizations can build their own solutions with full control, dramatic cost savings, and quality that matches or exceeds commercial platforms.

Understanding Voice Agent Categories

AI-driven voice assistants generally fall into two categories:

  1. Inbound Agents respond to incoming queries and requests, similar to customer support representatives answering calls. Example: An automated hospital phone system handling appointment scheduling and prescription refills.
  2. Outbound Agents proactively reach out to gather information, provide support, and execute tasks. Example: An AI system calling patients for post-discharge follow-ups to monitor recovery and ensure medication adherence.

Susan is an outbound agent designed for patient engagement—demonstrating the capability to initiate calls, conduct structured conversations, and collect actionable health data.

Business Value & Economics

The Traditional Options

Healthcare organizations conducting patient outreach have historically chosen from three expensive options:

1. In-House Human Staffing:

Hiring a full-time nurse or coordinator costs $45K-60K annually including benefits, with a typical capacity of 20-25 productive calls per day, totaling roughly 5,000 calls per year. This translates to a cost per call of $9-12. While in-house staff provides clinical expertise and direct organizational control, the approach comes with significant limitations including fixed capacity that cannot easily scale with demand, scheduling constraints tied to business hours, and quality inconsistency depending on individual staff performance and availability.

2. Outsourced Call Centers:

Call centers offer a cost per call of $2-5, providing scalable capacity and eliminating hiring overhead. Organizations can flex volume up or down based on seasonal or program-specific needs without the commitment of full-time employees. However, this approach introduces substantial challenges including HIPAA compliance concerns when patient data leaves the organization, quality control difficulties with external agents who lack institutional knowledge, high agent turnover that undermines consistency, limited clinical knowledge among general call center staff, and potential brand or reputation risk when patient interactions are handled by third parties.

3. Vendor AI Voice Platforms:

Specialized AI voice platforms charge $3-7 per call when accounting for platform fees, usage charges, and abandoned calls, delivering 40-70% savings compared to human staffing. These platforms offer compelling benefits including scalable infrastructure, 24/7 availability without staffing concerns, and consistent quality across all interactions. Despite these advantages, organizations face significant limitations including costs that remain expensive at scale, vendor lock-in that constrains future flexibility, and limited customization that prevents tailoring to specific clinical workflows or organizational requirements.

The Custom LLM-Based Solution

Cost per call: $0.30-0.75

Real-World Scenario: Post-Discharge Follow-Ups

Consider a hospital conducting 2,000 patient follow-up calls annually:

Approach Annual Cost Cost Per Call
Human FTE (in-house) $45K-60K $22.50-30
Outsourced Call Center $4K-10K $2-5
Vendor AI $6K-14K $3-7
Custom AI $600-1,500 $0.30-0.75

Custom Solution ROI:

The Scaling Advantage

As volume increases, the economic advantage compounds dramatically:

10,000 patient calls annually:

Annual savings with custom solution: $27K-62K vs. vendor AI, $17K-46K vs. outsourced

The Demo: Susan in Action

Demo Scenario

Susan is a virtual assistant at Sunnyside Hospital. Her task is to call patient Clint Eastwood, conduct a health assessment, and record responses in a structured format.

Key Characteristics:

Conversation Script

For this demonstration, Susan follows this structured but flexible conversation flow (provided as a CSV file):

Question nr,Question
1,"Confirm that you are talking to the patient. Ask to get the patient on the phone if it is not him/her"
2,"Ask for a most recent blood pressure measurement, and when it was taken"
3,"Ask for an email address"
4,"Over the last 2 weeks, how often have they been bothered by: Feeling down, depressed, or hopeless"
5,"Over the last 2 weeks, how often have they been bothered by: Trouble falling or staying asleep, or sleeping too much"

Demo Recordings

Three actual conversations between Susan and a patient have been recorded:

  1. Demo 1 (8 minutes) - "Happy path" scenario with screen recording for visual context
  2. Demo 2 (5 minutes) - Audio only with multiple unexpected scenarios
  3. Demo 3 (4 minutes) - Additional edge cases testing adaptive capabilities

Note: Demos 2 and 3 include several curveballs to test Susan's ability to handle unexpected situations. These scenarios were not deterministically programmed—Susan navigates them using general conversational intelligence.

Captured Information

Here's the structured data Susan captured during the first conversation:

Question nr,Answer
1,"Yes"
2,"120/90, taken yesterday"
3,"No email provided"
4,"Several days"
5,"Not at all"
3,"neb@gmail.com"

Notice that question 3 (email address) appears twice—the patient initially declined to provide it due to privacy concerns, then changed their mind later in the conversation. Susan successfully captured both responses and the final answer.

What Susan Can Do: Key Capabilities

The recorded conversations demonstrate several sophisticated capabilities:

1. Dynamic and Context-Aware Conversations

Susan doesn't rigidly follow a script. She generates natural responses based on the script's intent, patient input, and real-time context—while ensuring all required information is gathered.

2. Handling Non-Straightforward Responses

Patient answers are not always clear. Susan successfully manages conversations requiring multiple turns to obtain definitive responses.

3. Understanding Implicit References

When a patient says "same as the previous question," Susan correctly infers and records the appropriate response based on conversation history.

4. Generating Structured Output

Susan interprets and structures patient responses accurately. If a patient says "I've experienced that a few times during the last week," Susan categorizes and records it as "Several days" according to the standardized response format.

5. Managing Unscripted Topics

Patients engage in off-script discussions (expressing privacy concerns about email addresses, asking clarifying questions, going on tangents). Susan navigates these situations effectively without explicit scripting.

6. Detecting Changing Answers

If a patient changes their mind about an answer during the call, Susan recognizes the change and ensures the correct final information is captured.

7. Configuration Flexibility

Susan can be easily configured for different calls:

Strategic Advantages Beyond Cost

Full Control & Customization

Custom solutions deliver control and customization that vendor platforms cannot match. Organizations can connect directly with existing EHR/CRM systems without middleware or API limitations, enabling seamless integration with existing workflows. Conversation flows can be designed specific to clinical protocols and quality standards rather than adapting to vendor templates. Compliance and security implementations follow organization-specific privacy, security, and regulatory requirements rather than one-size-fits-all vendor approaches. Complete data ownership ensures all recordings, transcripts, and analytics remain in-house with no vendor data sharing or external processing. Perhaps most importantly, organizations achieve rapid iteration capability—updating conversation scripts, adding new use cases, or adjusting behavior in days without vendor timelines, approval processes, or additional fees.

Strategic Independence

Building custom solutions establishes true strategic independence. Organizations avoid vendor lock-in and eliminate dependency on vendor pricing, roadmap priorities, or technology choices that may not align with their needs. As AI capabilities evolve rapidly, maintaining flexibility to adopt new models and techniques becomes a strategic asset rather than waiting for vendor upgrades. Proprietary implementations create competitive advantage that competitors using the same vendor platforms cannot replicate. This approach also provides investment protection through the flexibility to pivot or adjust direction without sunk costs in vendor platforms or lengthy contract negotiations.

Healthcare-Specific ROI Opportunities

Post-Discharge Follow-Ups:

Appointment Reminders:

Chronic Disease Management:

Patient Satisfaction Surveys:

Why This Matters: The Market Shift

The Vendor Moat Has Eroded

Until recently, specialized AI voice platforms justified premium pricing through superior technology. Three years ago, vendor platforms commanded premium pricing because they offered superior natural language understanding, more robust commercial conversation engines, and professional voice synthesis that served as a clear differentiator.

Today, as this demonstration validates, the landscape has fundamentally changed. Modern LLMs now provide conversation quality that matches or exceeds vendor solutions. Commodity voice synthesis services like ElevenLabs, Azure, and Google deliver professional-grade audio at a fraction of the cost. Integration has become simpler with modern APIs and frameworks that didn't exist just a few years ago. The moat around vendor solutions has eroded, and with it, the justification for their premium pricing.

Quality Parity Achieved

The critical insight this demonstration validates: the technology gap between custom solutions and specialized vendor platforms has closed.

The demo recordings reveal important context about what's achievable even with minimal infrastructure. The demonstration runs on a local laptop, causing noticeable lag, occasional stuttering, and static noise in the audio. In a production environment with optimized cloud deployment, these issues would be eliminated entirely. Yet even with suboptimal infrastructure, Susan demonstrates capabilities comparable to commercial platforms. Unlike many AI voice assistants that rely on predefined decision trees or rigid scripts, Susan demonstrates advanced adaptability, context awareness, and nuanced conversational capabilities that were previously exclusive to expensive vendor platforms.

Conclusion

This demonstration represents more than a technical proof-of-concept—it validates a fundamental market shift in AI voice technology for healthcare.

The Core Validation: Healthcare organizations can now build custom AI voice agents using commodity tools and modern LLMs, achieving quality that matches or exceeds expensive vendor platforms at a fraction of the cost.

What This Means:

The traditional trade-off between quality and cost no longer exists. Organizations previously faced a difficult choice: invest heavily in vendor platforms or forgo AI-powered patient engagement. This experiment proves there's a third path—custom solutions that deliver economic sustainability through 90%+ cost reduction versus vendor platforms, strategic control with full ownership of technology, data, and roadmap, rapid innovation by testing, iterating, and deploying new use cases in days rather than months, quality parity with conversation capabilities that match or exceed commercial platforms, and measurable ROI through clear paths to value via reduced readmissions, improved adherence, and enhanced patient engagement.

The Market Disruption:

The same commoditization that transformed cloud computing and data analytics is now reshaping conversational AI. Organizations that recognize this shift early position themselves to build proprietary capabilities that create competitive advantage, avoid vendor lock-in and long-term cost escalation, innovate faster than competitors using traditional vendor solutions, and scale patient engagement programs that were previously cost-prohibitive. The window for early-adopter advantage remains open, but as this approach becomes mainstream, the competitive differentiation will diminish.

From Experiment to Production:

The technology is ready. The economics are compelling. The strategic advantages are significant. The question for healthcare leaders is no longer "Can we build this?" but rather "How quickly can we capture this advantage?"

View All Success Stories ->