Prologue: Redefining the patient support experience
When a patient dials a healthcare support line with questions about a medication or program eligibility, they aren’t looking to interact with a rigid database or navigate an endless automated menu. They want help, reassurance, and a listening ear, often during moments of deep vulnerability or distress.
This means that designing an AI agent for healthcare is far more than building something that answers routine questions. It is about building a digital “persona” that is reliable, deeply empathetic, and clinically safe. In fact, we were recently trusted by a biopharmaceutical company to create their very first patient-facing agent, in great part because they trusted our expertise and collaborative approach.
With that in mind, here is the story of how we at Infinitus craft an AI experience for handling inbound patient support calls that listens, recovers from real-world chaos, and protects patient safety above all else – while creating a positive experience that patients approve of.
Chapter 1: The human element in conversation
At Infinitus, we believe AI that engages in patient conversations should speak like a human, not read like a manual. To bring warmth and natural rhythm to every conversation, we established foundational design principles that shape every interaction:
- One turn, one idea: Rather than overwhelming callers with exhaustive options, the agent focuses on one question at a time, leading with clear answers and waiting patiently for responses.
- Confirm before acting: Early missteps can ruin trust. Thus, the agent always verifies crucial details before taking action – details such as names, numbers, or choices.
- Expect interruption: Real conversations aren’t scripted. When callers interrupt or pivot, the agent seamlessly addresses the new context before gently guiding them back.
- Acknowledge, then solve: True empathy comes first. If a patient shares a difficult situation, the agent acknowledges their pain with genuine warmth before moving toward a solution.
Chapter 2: Balancing warmth with regulatory rigor
In healthcare, answering patient questions requires strict adherence to regulatory guidelines. Balancing this rigor with a warm experience is one of our most important design challenges. We achieve it through a dynamic conversational model: for routine inquiries, the agent draws on flexible, natural language to build rapport. Yet, the moment a privacy or safety disclosure is required, the agent seamlessly locks into verbatim constraints, delivering the exact legal phrasing our customers require without a single deviation.
Chapter 3: Safety above all – a guardian escalation framework
Patient safety must never be compromised. To ensure rapid and appropriate responses to adverse events (AEs) or emergencies, we classify clinical acuity by leveraging the Schmitt-Thompson Triage Acuity Framework. We then categorize the level of risk across distinct acuity tiers, layer on adverse event detection, and can tailor adaptable client protocols for system actions, scripting, and escalations. A sample breakdown is shown below:
| Clinical Acuity Level | Triggers and scenarios | Protocol and response |
| Level 1: Life-threatening Immediate emergency | Life-threatening physical emergencies (e.g., anaphylaxis, severe respiratory distress) or active mental health crises. | Immediate warm transfer to a live specialist. If unavailable, normal call flow halts and the agent directly instructs the caller to dial 9-1-1 or 9-8-8 immediately. |
| Level 2: High riskPotential morbidity/mortality | Acute high-risk symptoms (e.g., stroke indicators, accidental overdose) | Immediate warm transfer to a live specialist. If unavailable, normal call flow halts and the agent directly instructs the caller to dial 9-1-1 immediately. |
| Level 3: Moderate risk Requires urgent evaluation | Active symptoms that require urgent evaluation but do not pose potential mortality risks. | Agent captures callback details and attempts a warm transfer. If a live agent is reached, the call transitions; otherwise, the agent gently collects adverse event logging details. |
| Level 4: Low riskMild or expected behavior | Active symptoms but do not require any urgent evaluation (minor pain or injuries or fever like symptoms). | No transfer required. The agent seamlessly transitions into structured intake questions to log the incident accurately. |
| Level 5: Info only/ passive eventsNo active symptoms | Reports of past/resolved symptoms or clinical events or general informational inquiries | No transfer required. Passive reports will be handled similar to low risk and document the historical incident. Any other general inquiries will be handled accordingly by the agent. |
Chapter 4: The art of self-recovery
Real phone calls take place in noisy cars, over muffled speakerphones, or across weak cellular connections. Our agents are specifically built to gracefully handle these real-world acoustic hurdles through advanced self-recovery:
- Intelligent re-asking: When static masks key information, the agent politely acknowledges the noise (“Sorry, I didn’t quite catch that…”) and asks for clarification without frustration.
- Resilient recovery: Under strict testing, our self-recovery system achieved remarkable resilience, successfully resolving misheard inputs on the very next attempt in over 90% of instances.
- Outcome-driven care: A single misheard word doesn’t define the call. What matters is ensuring every patient finishes their call having received the exact support they needed.
Chapter 5: Forging trust through rigorous evaluation
Earning trust in healthcare demands relentless testing. Before any agent update touches a live line, it undergoes intensive validation combining cutting-edge synthetic simulations and human evaluation:
Synthetic simulations and multi-dimensional testing
Instead of simple test prompts, we simulate real-world diversity through structured, multi-dimensional scenario design:
- Orthogonal scenarios: We vary caller intents, medical acuity levels, line noise, and emotional states to build thousands of unique test cases.
- Dynamic personas: Simulated callers operate under dynamic briefs with context-driven goals, testing the agent’s ability to manage interruptions and unscripted conversation flows.
- SOP assertion graders: Automated grading tools check every session for compliance disclosures, correct routing, and adherence to Standard Operating Procedures (SOPs).
Human evaluation and optional stress testing
Automated testing is paired with human-in-the-loop review to evaluate emotional resonance and delivery:
- Ops team pressure testing: Operational specialists conduct live manual test calls to stress-test the AI under complex persona scenarios.
- Hold-out set testing: Human testers employ varied phrasing and novel scripts to prevent model overfitting to synthetic test data.
- Quality & empathy rubrics: Evaluators grade each call against comprehensive rubrics covering empathy, professionalism, clinical accuracy, and safety.
Epilogue: Transforming the patient journey
At the end of the day, our goal is for technology to serve a lofty human goal: supporting patients during their time of need. When people reach out for support, they are often navigating new treatments, seeking clarity on symptoms, or coping with unexpected health challenges.
Because of this patient-centric framework, patients who engage with Infinitus AI agents never feel stranded in automated loops during critical health events. Life-threatening emergencies receive instant escalation to emergency services or specialists, while active symptoms are logged with empathy and care. By combining human-centered conversation design with unwavering clinical safety protocols, we empower every patient to feel truly heard, valued, and safely supported.
