Post-Call Analysis

Post-call analysis runs an LLM against the call transcript right after each call ends and writes structured JSON back to the call record. Useful for lead qualification, sentiment tracking, and automated CRM updates.

What it produces

For each call where analysis is enabled, the agent's chosen analyzer model extracts:

Field	When emitted
Your declared schema fields (see below)	Always
`summary`	When `enableSummary: true` (default on) — 2-3 sentence outcome summary
`sentiment`	When `enableSentiment: true` (default on) — `positive` / `neutral` / `negative`
`success`	When `enableSuccessEvaluation: true` — boolean against your `successCriteria`

Result is stored on calls.analysis as JSON, accessible via the dashboard call-details page or GET /api/calls/:id.

Configure

In the agent config, under Advanced → Post-Call Analysis, or via API:

{
  "postCallAnalysis": {
    "enabled": true,
    "schema": [
      {
        "name": "lead_qualified",
        "type": "boolean",
        "description": "Did the caller express interest in our product?",
        "required": true
      },
      {
        "name": "budget_inr",
        "type": "number",
        "description": "Stated annual budget in Indian Rupees, or null if not stated"
      },
      {
        "name": "next_action",
        "type": "enum",
        "enumValues": ["book_demo", "send_pricing", "callback_later", "not_interested"],
        "description": "What's the next step?",
        "required": true
      },
      {
        "name": "company_name",
        "type": "string",
        "description": "Caller's company name"
      }
    ],
    "summaryPrompt": "Summarize the call in 2-3 sentences focused on what was promised and what happens next.",
    "enableSummary": true,
    "enableSentiment": true,
    "enableSuccessEvaluation": true,
    "successCriteria": "The caller agreed to a demo or provided their email for follow-up",
    "analyzerModel": "gpt-4o-mini"
  }
}

Field types

`type`	Notes
`string`	Free text
`number`	Numeric value (integer or decimal)
`boolean`	true / false
`enum`	Must also provide `enumValues: string[]` — analyzer picks one

Set required: true to fail the analysis when a field can't be extracted (the call gets analysisStatus: "failed" in that case).

Analyzer model

By default the analyzer uses the same provider and model as the agent's main LLM. Override via analyzerModel:

gpt-4o-mini — cheapest, fast (~₹0.05/call for typical transcripts)
gpt-4o — better field-level accuracy
claude-haiku-4-5-20251001 — fast, multilingual
claude-opus-4-7 — most accurate, expensive

For lead-gen workloads at scale, gpt-4o-mini hits the right cost/accuracy balance.

Sample result

For a 90-second call where the caller agreed to a Friday demo:

{
  "lead_qualified": true,
  "budget_inr": 250000,
  "next_action": "book_demo",
  "company_name": "Acme India",
  "summary": "Caller expressed strong interest in the Pro tier for their 12-person sales team. Confirmed a budget of ₹2.5L/yr. Agreed to a demo at 3pm IST Friday with stakeholder Priya. Will receive a calendar invite within the hour.",
  "sentiment": "positive",
  "success": true
}

Webhook delivery

When the analyzer completes, osmTalk fires an analysis_completed event to your default webhook URL (set under Settings → Webhooks):

{
  "event": "call.analysis_completed",
  "timestamp": "2026-05-06T07:34:21.245Z",
  "call": {
    "id": "call_xxx",
    "agentId": "agent_xxx",
    "agentVersion": 7,
    "organizationId": "org_xxx",
    "channel": "phone",
    "durationSeconds": 92,
    "callerPhone": "+919876543210",
    "recordingUrl": "https://storage.osmtalk.com/...",
    "recordingMultiChannelUrl": null,
    "dynamicVariables": { "first_name": "Arjun" },
    "metadata": { "campaign_id": "..." },
    "campaignId": null,
    "campaignLeadId": null
  },
  "analysis": { /* same shape as above */ }
}

If a webhook secret is configured, requests are signed with X-OsmTalk-Signature: sha256=<hex> over the raw body — verify with HMAC-SHA256.

Pricing impact

Every analyzed call charges the analyzer model's per-token rate against your credit balance. Disable by setting postCallAnalysis.enabled: false if you don't need it.

When to use vs. inline tools

Use post-call analysis	Use an HTTP tool inside the call
Extracting data once at the end	Capturing fields as they're stated
Sentiment / summary needed	Real-time CRM updates
Asynchronous processing OK	Bot needs the result to decide what to say next
Many fields, complex schema	One or two values

On this page