Post-Call Sentiment + Lead Scoring: CallSphere vs Vapi Analytics Gap

TL;DR

Every call CallSphere handles is automatically post-processed by GPT-4o-mini into structured analytics: sentiment (-1.0 to 1.0), lead score (0-100), intent, topic extraction, satisfaction (1-5), escalation flag, and an AI-written summary. The data lands in call_log_analytics and powers the staff dashboard. Vapi.ai gives you the raw recording, the transcript, and a webhook. The analytics pipeline — what to score, how, where to store, how to display — is yours to build. This post walks the pipeline architecture and what it would take to replicate.

Why Auto-Analytics Beats Raw Recordings

A 200-unit property management firm or a 6-clinician medical practice handles roughly 80-120 calls a day. Nobody listens back to all of them. If your voice analytics is a folder of recordings, your analytics is whoever happens to listen to the angry call that escalated.

Auto-analytics flips that. Every call gets the same six dimensions, scored consistently, stored structurally, queryable. You can ask:

• "Show me every call this week with sentiment below -0.4."
• "What's the average lead score by source?"
• "Which provider's calls trend lowest on satisfaction?"
• "Which callers had escalation flags in the last 30 days?"

That is operational data, not anecdote.

Vapi's Analytics Story

Vapi's analytics is at the platform-operations level: latency, error rate, call duration. Per-call business analytics — sentiment, lead score, intent — is not a built-in concept. To replicate:

1. Pull the transcript via webhook.
2. Send it to an LLM with a structured-output prompt.
3. Parse the response into a row.
4. Store it (your database).
5. Build a dashboard.
6. Tune the prompt for consistency across thousands of calls.
7. Add domain-specific scoring (e.g., insurance-verified vs not, pre-approved vs not).
8. Maintain prompt versioning as the business evolves.

That's a 4-6 week build for the basic pipeline, plus ongoing prompt tuning, plus dashboard work.

CallSphere's Analytics Pipeline

CallSphere ships the pipeline. Every call writes to call_logs (raw transcript + recording reference). A post-call worker fires GPT-4o-mini analysis and writes to call_log_analytics with the following columns:

• sentiment_score — -1.0 (very negative) to 1.0 (very positive).
• lead_score — 0 (no buying intent) to 100 (red-hot).
• intent — primary call intent (scheduling, rescheduling, billing, complaint, info, emergency, etc.).
• topics — extracted topic array (insurance, specific provider, specific service, specific property).
• satisfaction_score — 1 to 5.
• escalation_flag — boolean; true if the call needs human follow-up.
• summary — 2-3 sentence AI-written summary of the call.

The dashboard surfaces those rows with filters, charts, and an alerts panel for escalations.

Comparison Table

Capability	Vapi (DIY)	CallSphere
Per-call sentiment score	Build	Built-in
Per-call lead score	Build	Built-in
Intent classification	Build	Built-in
Topic extraction	Build	Built-in
Satisfaction score	Build	Built-in
Escalation flag	Build	Built-in
AI call summary	Build	Built-in
Analytics database schema	Build	Built-in
Dashboard	Build	Built-in
Time to first dashboard	4-6 weeks	Live

Analytics Pipeline Diagram

flowchart LR
    A[Live call ends] --> B[(call_logs: transcript + recording ref)]
    B --> C[Post-call worker]
    C --> D[GPT-4o-mini structured prompt]
    D --> E{Output JSON}
    E --> F1[sentiment_score]
    E --> F2[lead_score]
    E --> F3[intent]
    E --> F4[topics]
    E --> F5[satisfaction_score]
    E --> F6[escalation_flag]
    E --> F7[summary]
    F1 --> G[(call_log_analytics)]
    F2 --> G
    F3 --> G
    F4 --> G
    F5 --> G
    F6 --> G
    F7 --> G
    G --> H[Staff dashboard]
    G --> I{escalation_flag}
    I -->|true| J[Alert: SMS + email to manager]
    I -->|false| K[No alert]
    H --> L[Filter, chart, drill-in]
    H --> M[Daily summary digest]

Worked Example: A Medical Practice Catches a Frustrated Patient Early

Tuesday 2:47pm. A patient calls about a billing question. They were on hold last week, got transferred twice, never got an answer. The voice agent looks up the account, sees the unresolved ticket, escalates to a billing specialist. The patient is polite but tired.

After the call:

• sentiment_score: -0.5
• lead_score: n/a (existing patient)
• intent: billing_followup
• topics: ["unresolved_ticket", "previous_transfer", "balance_dispute"]
• satisfaction_score: 2
• escalation_flag: true
• summary: "Returning patient calling about an unresolved billing ticket from last week. Frustrated by previous transfers. Voice agent escalated to billing specialist; ticket needs same-day callback."

The escalation alert pings the office manager at 2:48pm. By 3:15pm a human has called the patient back. The 3-month relationship is preserved. Without auto-analytics, that call is one of 200 in a folder nobody reviews.

Migration / Decision Section

If you are running Vapi and your operational reporting is "let me re-listen to a call" — the analytics gap is an everyday cost. Two paths:

• Build the pipeline. Realistic for engineering-heavy companies. 4-6 weeks plus prompt tuning. Worth it if your differentiation is custom analytics dimensions (e.g., compliance scoring for healthcare, deal-stage scoring for B2B sales).
• Use CallSphere. The pipeline is included; the dashboard is included; the alerting is included.

Most operators we onboard pick CallSphere because the analytics pipeline is the moment Vapi goes from "cheap voice infrastructure" to "we built half a product."

FAQ

How accurate is the sentiment score?

GPT-4o-mini sentiment scoring has been validated against human-labeled call samples; agreement is high on extreme scores (very positive / very negative) and reasonable on neutral ranges. The score is a signal, not a verdict; it should drive operational triage, not punitive action.

Is the lead score the same in healthcare as real estate?

The model is the same; the prompt is vertical-tuned. In healthcare, lead score primarily flags new-patient acquisition and conversion intent. In real estate, it flags buyer or renter intent strength. The dimensions are documented per vertical.

Can I add custom scoring dimensions?

Yes. Enterprise plans support custom analytic fields (e.g., "compliance_topics_mentioned" for regulated industries, "preferred_communication_channel" for CRM enrichment).

How fresh is the analytics row?

Default is post-call (typically within 30 seconds of call end). Real-time scoring during the call is available on enterprise plans for use cases that need mid-call routing decisions.

Does it work in non-English calls?

Sentiment and intent extraction work across the major languages GPT-4o-mini supports. Per-language prompt tuning is available on enterprise plans for non-English-dominant deployments.

Can I export the analytics?

Yes. Standard exports include CSV, JSON, and a streaming webhook. CRM integrations push the analytics row directly to leading CRMs (Salesforce, HubSpot) on enterprise plans.

See the analytics dashboard live at /demo. Pricing at /pricing.