ICD-10 + CDT Medical Records: Why Vapi Can't Match CallSphere Healthcare

TL;DR

A voice AI for healthcare needs more than a friendly voice — it needs a clinical data layer. CallSphere Healthcare ships with ICD-10-coded medical records, CPT and CDT-coded services, prescription tracking, and structured agent interactions, all wired to the voice agent's tool calls. Vapi.ai has none of this. Vapi is a voice infrastructure platform; the clinical schema, coding standards, and provider-billing alignment are all problems you solve yourself. This post details what each coding standard does, why it matters for voice AI, and the entity model CallSphere uses to keep voice conversations clinically coherent.

Why Voice AI Needs Clinical Coding Standards

When a clinic books an appointment, the slot is not just "Tuesday at 2pm." It is "Tuesday at 2pm for D1110 Adult Prophylaxis" or "Tuesday at 2pm for 99213 Established Patient Visit, Level 3." Those codes are the bridge between voice scheduling and downstream billing, payer alignment, provider scheduling rules, and EHR interoperability.

Without coding, a voice agent that books appointments produces clean-looking calendars and dirty-looking accounts receivable. The hygienist shows up to the room and the chart says "cleaning"; she does not know if the patient is due for a basic prophy or a periodontal maintenance, the front desk has to call back, and the practice loses 15 minutes per visit.

Three coding standards matter most for outpatient voice AI:

• ICD-10-CM — diagnoses. ~70,000 codes. Used on claims, in problem lists, and in clinical notes.
• CPT — procedures (medical). ~10,000 codes. Used for billing E/M visits, procedures, and labs.
• CDT — procedures (dental). ~750 codes. Maintained by the ADA. Used for every billed dental service in the U.S.

A voice agent that doesn't know about these is a chatbot that pretends to be a clinic.

Vapi's Clinical Data Layer: There Isn't One

Vapi's data model is "whatever you put in your webhook response." The platform has no notion of patients, no notion of diagnoses, no notion of services. That isn't a defect — it is the explicit positioning of the product. Vapi is voice infrastructure; clinical data is your problem.

What that means in practice:

• You design the patient table.
• You design the medical_records table and decide whether to use ICD-10 (you should).
• You design the services table and decide whether to use CPT and CDT (you should).
• You design the prescriptions table and decide on RxNorm/NDC.
• You write each tool that lets the agent read or write to those tables.
• You handle ICD-10 mid-year code changes (every October).
• You handle the CDT annual update (every January).
• You handle CPT updates (multiple times per year).

For a generalist team, that is a 6-12 month buildout before the first claim is filed cleanly.

CallSphere Healthcare's Clinical Data Layer

CallSphere ships the schema. Tables include:

• medical_records — patient_id, provider_id, encounter_date, icd10_codes (array), chief_complaint, assessment, plan, free-text note.
• prescriptions — patient_id, provider_id, drug_name, dose, frequency, refills, status (active/discontinued).
• services — name, cpt_code, cdt_code, default_duration_minutes, default_price.
• appointments — patient_id, provider_id, service_id (FK), start, end, status.
• patient_insurance — patient_id, payer_name, member_id, group_id, eligibility_cached_at.

That schema is referenced by the 14 voice tools. get_services returns services with their CPT/CDT codes. schedule_appointment writes the service_id into the appointment row. The provider's printed schedule the next morning is coded correctly because the voice agent never wrote a free-text "cleaning" — it wrote a structured pointer to D1110.

Comparison Table

Clinical capability	Vapi.ai	CallSphere Healthcare
Patient demographics table	DIY	Built-in
ICD-10-CM medical records	DIY	Built-in
Prescription table	DIY	Built-in
CPT-coded services	DIY	Built-in
CDT-coded dental services	DIY	Built-in
Insurance / eligibility	DIY	Built-in
Provider schedules	DIY	Built-in
Encounter notes (chief complaint, assessment, plan)	DIY	Built-in
Annual code update process	DIY	Built-in
Voice tools coded against schema	DIY	14 tools
Practice-level isolation (multi-tenant)	DIY	Built-in

Medical Record Entity Diagram

erDiagram
    PRACTICES ||--o{ PROVIDERS : employs
    PRACTICES ||--o{ PATIENTS : registers
    PRACTICES ||--o{ SERVICES : offers
    PROVIDERS ||--o{ PROVIDER_SCHEDULES : has
    PROVIDERS ||--o{ APPOINTMENTS : sees
    PATIENTS ||--o{ PATIENT_INSURANCE : holds
    PATIENTS ||--o{ APPOINTMENTS : books
    PATIENTS ||--o{ MEDICAL_RECORDS : owns
    PATIENTS ||--o{ PRESCRIPTIONS : prescribed
    SERVICES ||--o{ APPOINTMENTS : type_of
    APPOINTMENTS ||--o{ MEDICAL_RECORDS : results_in
    MEDICAL_RECORDS ||--o{ PRESCRIPTIONS : may_include
    PRACTICES {
        uuid id PK
        text name
        text type "medical or dental"
    }
    PATIENTS {
        uuid id PK
        text first_name
        text last_name
        date date_of_birth
        text phone
    }
    MEDICAL_RECORDS {
        uuid id PK
        uuid patient_id FK
        uuid provider_id FK
        date encounter_date
        text[] icd10_codes
        text chief_complaint
        text assessment
        text plan
    }
    PRESCRIPTIONS {
        uuid id PK
        uuid patient_id FK
        text drug_name
        text dose
        text frequency
        int refills
        text status
    }
    SERVICES {
        uuid id PK
        text name
        text cpt_code
        text cdt_code
        int duration_minutes
        numeric price
    }

Worked Example: Dental Hygiene Recall vs Periodontal Maintenance

A returning patient calls. They are between D1110 (Adult Prophylaxis) and D4910 (Periodontal Maintenance) depending on their last cleaning history and insurance benefits.

A voice agent that doesn't know CDT codes will book a "cleaning." The next morning, the hygienist starts a D1110, then mid-procedure realizes the patient was last seen for active perio therapy and is in maintenance. Two outcomes: (1) the practice undercharges; (2) the insurance claim is denied for code mismatch.

A CallSphere voice agent calls get_services, sees both codes with their criteria, calls get_patient_appointments to look at the last visit type, and either schedules D4910 or escalates to staff if the data is ambiguous. The hygienist's room is set up correctly. The claim files cleanly. The patient is happy.

That is the practical value of having a clinically-coded schema sitting under the voice agent.

Migration / Decision Section

If you are evaluating Vapi for a clinic and the question on the table is "can we replicate this?" — yes, you can. The honest cost is roughly:

• 6-12 months of engineering for the clinical data layer.
• Annual recurring effort to track ICD-10, CPT, and CDT updates.
• A clinical informaticist on staff or on retainer to validate code mappings.
• A QA loop that catches mis-coded appointments before they hit billing.

Most clinic operators choose to skip that path. A voice AI that produces clean billing data on day one is worth more than the cost difference, because every dirty claim costs $25 to rework and every coding error compounds across thousands of monthly visits.

FAQ

Why does the voice agent need to know CPT/CDT codes?

Because the calendar entry it creates is the upstream of every billing and claim downstream. A coded calendar produces clean revenue. An uncoded calendar produces a billing manager rebuilding the day every morning.

How are ICD-10 updates handled?

CallSphere ships an annual update (October 1, when CMS publishes new ICD-10-CM) and rolls deprecated codes through a documented migration path. Existing records keep their historical codes; new records use the active set.

Can I use SNOMED CT or LOINC?

SNOMED CT is supported as an optional vocabulary on enterprise plans. LOINC is used where lab data is in scope. Most outpatient practices live primarily in ICD-10 + CPT/CDT.

What about RxNorm and NDC for prescriptions?

The prescriptions table captures drug_name, dose, frequency, and refills. RxNorm/NDC tagging is supported on enterprise plans for practices that need formulary alignment or e-prescribing integration.

Does the voice agent ever generate clinical notes?

The agent does not author clinical assessments — that is a provider responsibility. It does write structured chief complaint and reason-for-visit fields, captured during the voice intake, that the provider reviews and signs off on inside the dashboard.

How does CallSphere prevent the agent from offering services not in the practice's catalog?

Services are scoped to the practice. get_services only returns rows where practice_id matches the inbound call's tenant. The agent cannot offer a service the practice does not actually deliver.

See the entity model wired to a real voice agent at /demo. Healthcare details at /industries/healthcare.