By Sagar Shankaran, Founder of CallSphere
CallSphere ships 14 healthcare voice tools, schemas, and execution logic. Vapi gives you Function Calling and 100+ engineering hours of work to match.
Key takeaways
CallSphere's Healthcare voice agent ships 14 production function-calling tools with finished JSON schemas, validated argument types, server-side execution, error handling, and Prisma-backed persistence. To match the surface area on Vapi.ai, an engineering team would write the schemas, the executors, the auth layer, the audit trail, the test harness, and the data model from scratch. At an industry-average 8 engineering hours per tool for the schema layer alone (and double that with full integration testing), the build-it-yourself path on Vapi is 100 to 200 engineering hours before the first appointment is booked. CallSphere ships those tools turn-key.
Modern voice AI does not run on prompts alone. The interesting behavior — booking an appointment, looking up a patient, confirming insurance, escalating to a human — happens through tool calls, also known as function calls. The LLM picks a tool from a registered list, fills in the arguments, and the platform executes the tool and returns the result for the next turn.
This pattern is a feature of every modern model: OpenAI's Realtime API, Anthropic's Claude, and Google's Gemini all support it. What differs across platforms is how much of the glue they ship. The glue is what turns a tool-call signal into a real database write, an SMS send, a calendar update, or a clinician page.
Vapi.ai's answer is Function Calling: you describe a tool, give it a webhook URL, and Vapi calls your URL when the model requests the tool. CallSphere's answer is a typed tool catalog with the executors already implemented inside the FastAPI backend.
Healthcare Voice (powering Marengo Asia Hospitals) runs a single Head Agent with 14 function-calling tools. Here is the catalog:
| # | Tool | Purpose | Persistence |
|---|---|---|---|
| 1 | search_provider | Find clinician by name, specialty, location | Prisma read |
| 2 | get_provider_availability | Open slots for a provider over a date range | Prisma read |
| 3 | book_appointment | Create a confirmed appointment | Prisma write + SMS |
| 4 | reschedule_appointment | Move existing appointment | Prisma write + SMS |
| 5 | cancel_appointment | Cancel and free the slot | Prisma write + SMS |
| 6 | lookup_patient | Patient record by phone or ID | Prisma read |
| 7 | create_patient | Register a new patient | Prisma write |
| 8 | check_insurance | Verify insurance plan against accepted list | Prisma read + rules |
| 9 | get_clinic_hours | Retrieve hours per location | Prisma read |
| 10 | escalate_to_human | Page on-call coordinator | Twilio voice |
| 11 | send_pre_visit_sms | Reminder + paperwork link | Twilio SMS |
| 12 | record_symptoms | Append structured symptom data | Prisma write |
| 13 | get_referral_status | Status of a pending referral | Prisma read |
| 14 | log_call_outcome | Final disposition for analytics | Prisma write + analytics |
Every tool has a strongly typed JSON schema, a Python executor in FastAPI, an idempotency check where relevant, structured logging, and a return shape that the model knows how to consume.
Hear it before you finish reading
Talk to a live CallSphere AI voice agent for healthcare in your browser — 60 seconds, no signup.
{
"name": "book_appointment",
"description": "Book a new appointment for an existing patient.",
"parameters": {
"type": "object",
"properties": {
"patient_id": { "type": "string" },
"provider_id": { "type": "string" },
"start_time": { "type": "string", "format": "date-time" },
"duration_minutes": { "type": "integer", "minimum": 5, "maximum": 120 },
"visit_type": { "type": "string", "enum": ["new", "follow_up", "telehealth"] },
"notes": { "type": "string" }
},
"required": ["patient_id", "provider_id", "start_time", "visit_type"]
}
}
The executor validates the slot is still free, writes to the appointments table, fires an SMS via Twilio, and returns a confirmation number to the model. All of this is shipped.
Vapi's Function Calling primitive does the right thing as a primitive: it lets you define a function with a JSON schema and a webhook URL. When the model calls the function, Vapi posts the arguments to your URL and waits for a response.
That is roughly 5 percent of what is required to run a healthcare voice line. The remaining 95 percent — the executor itself, the database schema, the SMS integration, the audit trail, the failure modes, the retry policy, the idempotency layer, the analytics plumbing — is all on you.
For a single tool, that is a few days of engineering. For 14 tools across a regulated vertical, the math gets ugly fast.
Industry averages from voice AI engineering postmortems put the cost of a production-ready function-calling tool at 8 hours minimum for the schema, the executor, and basic happy-path testing. With error handling, observability, idempotency, and a real test suite, the per-tool cost rises to 15 hours.
| Task | Hours per tool |
|---|---|
| Schema design + review | 1 |
| Executor implementation | 3 |
| Database schema migration | 1 |
| Auth + tenant scoping | 1 |
| Logging + analytics | 1 |
| Webhook contract + retries | 1 |
| Happy-path tests | 2 |
| Edge case + failure tests | 3 |
| QA pass with voice loop | 2 |
| Subtotal | 15 |
For 14 tools, that is 210 engineering hours — roughly 5 weeks of one full-time engineer, before any compliance review. Most healthcare deployments require a HIPAA Business Associate Agreement, a security review, and an audit log certification on top of the build. CallSphere ships those workflows.
graph LR
A[OpenAI Realtime API] -->|tool_call event| B[FastAPI Dispatcher]
B --> C{Tool Name}
C -->|book_appointment| D[Prisma + Twilio]
C -->|lookup_patient| E[Prisma read]
C -->|escalate_to_human| F[Twilio Voice]
C -->|send_pre_visit_sms| G[Twilio SMS]
D --> H[Tool Result]
E --> H
F --> H
G --> H
H -->|tool_result event| A
| Capability | CallSphere Healthcare | Vapi Function Calling |
|---|---|---|
| Pre-built tool count | 14 healthcare tools | 0 |
| JSON schemas authored | Yes | Build yourself |
| Database schema | Prisma migrations included | Build yourself |
| SMS integration | Twilio wired | Build yourself |
| Idempotency layer | Yes | Build yourself |
| Audit logging | Yes | Build yourself |
| Tenant scoping | Multi-tenant ready | Build yourself |
| HIPAA readiness | BAA + audit hooks | Your responsibility |
| Hot-reload of tool code | k3s + hostPath | Cloud redeploy |
| Test fixtures | Shipped | Write yourself |
The reason CallSphere invested in turn-key tool catalogs for healthcare, real estate, salon, IT, and after-hours is that regulated industries cannot afford a half-built integration layer. A missed SMS confirmation costs revenue. A wrongly written appointment is a compliance event. An audit log gap during a chart review can void a deal.
Vapi's Function Calling primitive is honest about what it is — a primitive. CallSphere bundles primitives plus operators plus regulated-vertical defaults so the customer signs an agreement and books their first call within a week, not a quarter.
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.
The tool catalog scales with verticals:
Building this surface area on Vapi is the equivalent of writing a vertical SaaS company on top of a voice API. Some teams will choose to do that. Most teams discover halfway through that the platform decision was the most expensive one they made.
You can. CallSphere's tool layer is extensible — you can register new tools with custom schemas and executors via the FastAPI backend. The 14 healthcare tools are the defaults, not a ceiling.
Yes. The voice agent supports 57+ languages, and the tool layer is language-agnostic. The model invokes tools regardless of conversation language.
Tool inputs and outputs flow through the FastAPI backend on infrastructure governed by a BAA. PHI is never sent to third-party services without explicit configuration, and the audit log captures every tool invocation with a tenant scope. See the healthcare industry page for the full compliance model.
No. The model selects the right tool per turn. A typical healthcare call uses 2 to 4 tool calls — book + lookup + send_sms is a common path.
Yes. Schedule a live demo and we will run a real booking call where you watch the tool calls hit the dispatcher in real time.
Each executor returns a structured error shape that the model interprets and turns into natural conversation ("I'm sorry, that slot was taken — would you like the next available one?"). The retry policy is configurable per tool.
Talk to CallSphere about deploying the healthcare voice stack and ship in days instead of months. Book a demo to see all 14 tools dispatch live.

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.
AWS HealthScribe became the open scribe layer EHR vendors built on top of in 2026. Here's the API surface, the per-encounter pricing, the BAA terms.
Apollo, Manipal, and Narayana scaled AI agents across Bangalore in 2026. Here's the deployments across radiology, intake, and follow-up, the costs.
Notable's AI agents now handle scheduling, intake, and revenue cycle for 6,000+ clinics in 2026. Here's the multi-agent architecture, the per-clinic pricing.
Abridge raised $250M in April 2026 at a $2.7B valuation. We break down the deployment numbers, the EHR integrations across Epic and Cerner. The Q2 2026 buyer briefing.
Enterprise CIO Guide perspective on Hippocratic AI's deployment numbers show healthcare voice agents are moving from pilot to production across major US health systems.
Home health agencies use AI voice agents for daily check-ins on 2.5M+ patients in 2026. Here's the deployments at major agencies, the per-patient pricing.
© 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