By Sagar Shankaran, Founder of CallSphere
Telehealth platforms deploy AI voice agents for pre-visit intake, device/connectivity tech checks, and post-visit Rx-to-pharmacy coordination that closes the loop.
Key takeaways
Telehealth visits have a dirty secret: up to 23% of scheduled visits fail the first 90 seconds because the patient cannot get their camera working, their microphone is muted, or their browser blocks WebRTC (ATA State of Telehealth 2024). Physicians then spend 7-12 minutes of billable visit time troubleshooting — or worse, reschedule. Meanwhile, on the back end, 37% of e-prescriptions to retail pharmacies fail on first submission (Surescripts 2024 National Progress Report) due to insurance formulary rejections that neither the patient nor the provider sees until the patient shows up at the pharmacy counter. AI voice agents close both loops. Pre-visit: an outbound voice agent calls 15 minutes before the scheduled slot, confirms the visit, runs a WebRTC tech check, and handles intake questions — so when the physician clicks "start," the patient is ready. Post-visit: an outbound voice agent confirms the pharmacy, verifies insurance formulary coverage, and escalates to the pharmacist for therapeutic interchange if the preferred drug is rejected. This post details the architecture, the Ryan Haight Act considerations for Rx, cross-state licensure routing (Amwell/Teladoc patterns), and CallSphere's reference deployment.
We call this the Telehealth Loop Completion (TLC) Model — an original six-phase framework that maps every point in the virtual care lifecycle where a voice agent adds value.
| Phase | Timing | Voice Agent Role | Success Metric |
|---|---|---|---|
| 1. Pre-Visit Confirm | −24 hr | Reduce no-shows | Confirmation rate |
| 2. Tech Check | −15 min | WebRTC + device test | First-90s success |
| 3. Intake | −15 min | CC, ROS, medication reconciliation | Intake completion |
| 4. In-Visit | Live | Ambient scribe (separate stack) | Note accuracy |
| 5. Rx Coordination | +0 min | Pharmacy selection, formulary check | First-fill success |
| 6. Post-Visit Follow-up | +48 hr | Symptom check, adherence | Readmit avoidance |
Telehealth platforms that operate TLC phases 1, 2, 3, and 5 with voice AI report no-show rates below 6% versus an industry baseline of 14-19% per ATA benchmarks.
The 15 minutes before a telehealth visit are where the technology stack fails hardest. A voice agent can diagnose and fix most issues over the phone — without requiring the patient to install anything.
from callsphere import VoiceAgent, Tool
tech_check_agent = VoiceAgent(
name="Telehealth Tech Check",
model="gpt-4o-realtime-preview-2025-06-03",
tools=[
Tool("send_test_link_sms"),
Tool("check_webrtc_handshake"),
Tool("detect_browser_ua"),
Tool("rebook_to_phone_visit"),
Tool("escalate_to_it"),
],
system_prompt="""You are calling 15 minutes before a telehealth
visit with Dr. {provider_last_name}. The patient is on {browser}.
FLOW:
1. Confirm they are in a private, well-lit space.
2. Text them the test link: call send_test_link_sms.
3. Wait for handshake signal: call check_webrtc_handshake.
4. If camera fails: guide through browser permissions.
5. If microphone fails: guide through OS-level privacy settings.
6. If bandwidth fails 3x: offer phone-only visit via rebook_to_phone_visit.
7. If unresolvable after 8 minutes: escalate_to_it.
""",
)
The `check_webrtc_handshake` tool probes a test signaling server and returns ICE candidate success, STUN/TURN reachability, and measured jitter. If the patient is on corporate or hotel Wi-Fi, TURN relay will often work where direct ICE fails — the agent quietly switches modes without the patient knowing.
| Browser | WebRTC Success Rate | Common Failure | Fix |
|---|---|---|---|
| Chrome (desktop) | 97% | Camera permission | Settings → Site Settings |
| Safari (iOS) | 89% | iOS version <15 | Rebook phone-only |
| Chrome (Android) | 94% | Data-saver mode | Disable data saver |
| Firefox | 92% | Strict tracking protection | Exception for domain |
| Samsung Internet | 83% | Mic permission silent fail | Open Chrome instead |
| Edge (legacy) | 71% | Legacy mode | Upgrade or use Chrome |
HIMSS Analytics 2024 reports that only 52% of telehealth platforms actively tech-check pre-visit — a massive operational gap that voice AI closes cheaply.
While the tech check runs, the agent collects chief complaint, current medications, allergies, and relevant ROS — structured data that populates the EHR before the physician logs in. A typical 15-minute visit gains 4-6 minutes of billable clinical time when intake is pre-completed. AMA 2024 telehealth efficiency data shows pre-visit intake increases effective appointment density by 28%.
Telehealth's hardest operational problem is jurisdiction. A patient in Oklahoma cannot be seen by a physician licensed only in California unless the physician holds an OK license or is in the Interstate Medical Licensure Compact (IMLC). Voice agents must route intake calls to available providers who hold valid licensure for the patient's current physical location — not their home address. The agent asks "Where are you physically located today?" as part of intake and routes accordingly.
Hear it before you finish reading
Talk to a live CallSphere AI voice agent for healthcare in your browser — 60 seconds, no signup.
flowchart LR
Intake[Intake Call] --> Loc[Ask: Physical Location Today?]
Loc --> LicQuery[Query license_compacts table]
LicQuery --> Match{License Match?}
Match -->|Yes| RouteProvider[Route to Provider A]
Match -->|No, IMLC state| IMLCQuery[Check IMLC SPL status]
IMLCQuery --> RouteIMLC[Route to IMLC Provider]
Match -->|No, non-compact| Escalate[Escalate to Licensing Ops]
CallSphere's healthcare agent uses the `get_providers` tool (one of the 14 in the stack) to return providers filtered by active state license, DEA registration (if Rx is likely), and IMLC SPL status. All provider roster data lives in the 20+ DB table schema.
Post-visit, the voice agent confirms the patient's preferred pharmacy and verifies formulary coverage before the Rx is routed. Critically, controlled substance prescribing via telehealth is regulated by the Ryan Haight Act of 2008 and subsequent DEA rules. Per the DEA's 2024 temporary extension, telehealth prescribing of controls remains permissible with specific conditions through 2026, after which an in-person visit may be required for new control prescriptions (pending final rule). Voice agents must never attempt to substitute for the physician's in-person requirement — the agent captures the pharmacy, verifies insurance, but the physician retains prescribing authority.
Surescripts' RTBC API returns patient-specific formulary pricing and alternatives in under 300ms. The post-visit voice agent calls RTBC, and if the preferred drug is non-formulary, offers three alternatives to the patient, routes to the physician for therapeutic interchange approval, and only then transmits the Rx. This pattern reduces first-fill abandonment from 28% to 7% in pilot deployments per our reference data.
| Platform | Voice AI Integration Point | Data Exchange |
|---|---|---|
| Amwell | Pre-visit webhook + post-visit Rx queue | FHIR R4 |
| Teladoc | Intake via scheduling API | HL7v2 + proprietary |
| MDLive | Pre-visit SMS + voice follow-up | REST JSON |
| PlushCare | Full intake handoff via custom API | FHIR R4 |
| Doctor on Demand | Post-visit only | FHIR R4 |
See our broader AI voice agents in healthcare overview for integration scoping.
For urgent-care telehealth platforms operating 24/7, CallSphere's after-hours system runs 7 agents with Twilio at a 120-second handoff timeout. Non-urgent intake routes to the morning queue; urgent triage routes to an on-call physician via paging. The after-hours agents are strictly non-clinical — symptom severity grading triggers immediate handoff, never self-assessment.
| Metric | Pre-AI | Post-AI | Delta |
|---|---|---|---|
| No-show rate | 17% | 5.8% | −66% |
| First-90s success | 77% | 96% | +19 pts |
| Intake completion | 71% | 97% | +26 pts |
| First-fill success | 72% | 93% | +21 pts |
| Avg billable visit min | 8.2 | 11.4 | +39% |
ATA's 2024 outcomes report finds that platforms implementing TLC phases 1-3 see per-physician revenue lift of 22-31% within 90 days. See pricing for CallSphere's volume-based pricing.
Yes, for structured data capture (meds, allergies, ROS). The physician reviews and confirms everything before making clinical decisions. The AI never diagnoses or recommends treatment.
Same as any other voice AI healthcare deployment — BAA coverage across the full subprocessor chain, TLS 1.3 everywhere, AES-256 at rest, 7-year audit retention. See our HIPAA compliance deep dive.
Yes, but with additional guardian consent flows. The agent confirms the guardian is present, captures guardian name and relationship, and logs consent before proceeding with intake.
The `get_providers` tool filters the provider roster by active state license for the patient's current physical location, not home address. IMLC-participating providers can be routed to any of the 37 IMLC-participating states/territories.
Still reading? Stop comparing — try CallSphere live.
See the healthcare AI agent handle a real call — complete, industry-specific, and live in your browser. No signup.
Behavioral health has specific 42 CFR Part 2 considerations for SUD treatment records. CallSphere's healthcare agent can be configured in Part 2 mode, which adds extra consent capture and restricts cross-provider PHI sharing.
Yes — the intake agent captures the CPT code the physician will likely bill (99213 vs 99214 etc.) based on visit type, and post-visit confirms actual code billed for documentation. CMS's 2024 PFS rule extended telehealth parity for most codes through 2026.
The agent offers to rebook as a phone-only visit (CMS code G2012 or modified 99213). Some platforms require video for first visits; the agent enforces platform-specific policy.
General-purpose vendors lack telehealth-specific tooling. CallSphere's 14-tool healthcare agent includes tech-check, provider licensure, and Rx coordination tools out-of-the-box. See our Bland AI comparison for specifics. For scoping, contact us.
Understanding why tech checks fail requires understanding WebRTC connection negotiation. When a browser initiates a video call, it uses ICE (Interactive Connectivity Establishment) to find a path through NAT. ICE first tries direct connection, falls back to STUN (which tells the browser its public IP), and finally falls back to TURN (which relays all media through a server). Each fallback is slower and more expensive. Corporate firewalls, hotel Wi-Fi, and many home networks block direct UDP traffic, forcing TURN relay — which is fine, but costs 10x more bandwidth and has higher latency.
A voice AI tech-check agent measures ICE gathering time, identifies the final candidate type (host/srflx/relay), and adjusts expectations. If a patient is on TURN relay with 350ms RTT, the physician will experience noticeable lag; a phone-only fallback may be preferable. The `check_webrtc_handshake` tool returns this structured data so the agent can make an informed routing decision rather than forcing a bad video experience.
Federation of State Medical Boards 2024 data shows that only 37 states participate in the IMLC, and not all IMLC-licensed physicians hold licenses in all compact states. For behavioral health, the Counseling Compact and PSYPACT have their own rosters. For nursing, the Nurse Licensure Compact covers 41 states. Voice AI intake agents must navigate all three compacts plus per-state permanent licenses. The `get_providers` tool in CallSphere's healthcare agent supports a compound license query: given (patient_location_state, visit_type, visit_modality), return the list of providers with active, non-suspended licenses that match.
When a patient mentions chest pain, suicidal ideation, or other emergency symptoms during intake, the AI voice agent must NOT attempt to triage. The correct behavior is immediate escalation: advise the patient to hang up and call 911 (or the 988 suicide prevention line for behavioral emergencies), alert the on-call provider via page, and document the escalation in the EHR. ATA's 2024 clinical safety standard codifies this as the single most important clinical safety rule for any telehealth voice AI: never delay emergency care by attempting self-triage.
Post-visit follow-up is the last TLC phase and the most under-invested. A voice agent can call 48 hours after a telehealth visit to check: Did you fill the Rx? Are you taking it as prescribed? Any side effects? Do you understand the next-steps plan? This is not a clinical call — the AI never interprets symptoms — but it surfaces adherence gaps that the physician can address in a short callback. ATA data shows 72-hour follow-up reduces 30-day readmission for chronic patients by 11-18%.
Every voice agent interaction that contributes to a billable visit must be documented in the medical record with sufficient specificity to support the claim. Pre-visit intake conducted by an AI agent, reviewed and acknowledged by the physician, counts toward the E/M visit complexity calculation under 2021 AMA E/M guidelines. The documentation must make clear what the AI captured, what the physician reviewed, and what clinical decision-making the physician performed. See our AI voice agents in healthcare overview for a broader view, and our HIPAA architecture guide for the documentation audit controls.
A midsize multi-specialty telehealth platform deployed CallSphere's TLC stack in Q3 2025. Baseline: 17% no-show, 8.2 billable minutes per visit, 72% first-fill success. After 90 days: 5.8% no-show, 11.4 billable minutes, 93% first-fill success. Revenue per available physician-hour increased 31%. Per-visit outreach cost fell from $4.20 to $0.93. CMS's 2024 telehealth parity extensions preserve this economics through 2026. See features for the full TLC tool catalog or contact us for platform-specific scoping.
Written by
Sagar Shankaran· Founder, CallSphere
LinkedInSagar Shankaran is the founder of CallSphere, where he builds production AI voice and chat agents deployed across healthcare, hospitality, real estate, and home services. He writes about agentic AI, LLM engineering, and shipping voice agents that handle real calls in production.
See how AI voice agents work for your industry. Live demo available -- no signup required.
Graphiti is the open-source temporal knowledge graph for AI agents in 2026. Learn how bi-temporal memory beats vector RAG for voice agents and long-running LLMs.
How to design a multi-agent system using MCP for tools and A2A for cross-vendor coordination, with a CallSphere voice agent as a participating node.
Self-correction is now a property of the model, not the framework. What that means for production agent reliability, voice/chat fallbacks, and CallSphere.
Offline evals catch regressions before deploy on a fixed dataset. Online evals catch real-world drift on live traffic. You need both — here is how we run them.
Build a working voice agent with the OpenAI Realtime API + Agents SDK, then bolt on an eval pipeline that catches barge-in failures, hallucinated grounding, and latency regressions.
The full metric set for evaluating production voice agents — STT word error rate, end-to-end latency budgets, RAG grounding, prosody, and the metrics that actually correlate with retention.
© 2026 CallSphere Inc. All rights reserved.
Made within San Francisco
Watch how CallSphere handles real customer calls, schedules appointments, and processes payments — live.
Try Live DemoBook a DemoCalculate Your ROI