Copy page

Guides

Post-Call Analysis

After every call ends, Neuratel runs an AI analysis pipeline that generates structured insights from the transcript. Configure what gets analyzed via the agent's analytics.post_call settings.

What Gets Analyzed

Four analysis types run in parallel:

Type	Output
Summary	AI-generated text summary of the call
Success Evaluation	Rubric-based score measuring whether the call succeeded
Data Collection	Structured JSON extraction of fields you define
Sentiment Timeline	Per-turn sentiment score (-1.0 to 1.0) with trend analysis

Configuration

Configure via the agent's analytics section:

Code
{
  "analytics": {
    "enabled": true,
    "track_sentiment": true,
    "post_call": {
      "summary": {
        "enabled": true,
        "prompt": "Summarize what the caller needed and how the agent resolved it. Include any follow-up actions mentioned."
      },
      "success_evaluation": {
        "enabled": true,
        "criteria": [
          {
            "id": "issue_resolved",
            "name": "Issue Resolved",
            "description": "Did the agent successfully resolve the caller's primary issue?",
            "weight": 2.0
          },
          {
            "id": "tone_professional",
            "name": "Professional Tone",
            "description": "Did the agent maintain a professional and empathetic tone?",
            "weight": 1.0
          }
        ],
        "threshold": 0.7
      },
      "data_collection": {
        "enabled": true,
        "fields": [
          {
            "id": "caller_intent",
            "name": "Caller Intent",
            "description": "The primary reason the caller contacted us",
            "type": "string",
            "required": true
          },
          {
            "id": "issue_resolved",
            "name": "Issue Resolved",
            "description": "Whether the caller's issue was resolved",
            "type": "boolean",
            "required": true
          },
          {
            "id": "follow_up_needed",
            "name": "Follow-Up Needed",
            "description": "Whether a human follow-up is required",
            "type": "boolean",
            "required": false
          }
        ]
      }
    }
  }
}

Success Evaluation Rubrics

Choose a rubric type for success evaluation:

Rubric value	Use case
PassFail	Binary: did the call succeed or fail?
NumericScale	Score from 1–10
DescriptiveScale	Excellent / Good / Fair / Poor
PercentageScale	0–100 % score
LikertScale	Strongly agree → strongly disagree
Checklist	Pass/fail per criterion
Matrix	Multiple dimensions scored independently
AutomaticRubric	AI determines the best rubric automatically

The values are PascalCase string literals (matches the SuccessEvaluationRubric enum on the backend). Configure under analytics.post_call.analysisPlan:

Code
{
  "analytics": {
    "post_call": {
      "analysisPlan": {
        "enabled": true,
        "successEvaluationRubric": "NumericScale",
        "successEvaluationPrompt": "Score this call from 1-10 based on whether the agent resolved the customer's issue completely and efficiently.",
        "summaryPrompt": "Summarize the call in 2-3 sentences.",
        "structuredDataPrompt": "Extract the following fields from the call.",
        "structuredDataSchema": {
          "type": "object",
          "properties": {
            "appointment_booked": {"type": "boolean"},
            "follow_up_date": {"type": "string"}
          }
        },
        "extract_topics": false,
        "extract_action_items": false,
        "compute_user_sentiment": false,
        "minMessagesThreshold": 2
      }
    }
  }
}

enabled: true is the master switch — analysis is opt-in. With it false, none of the prompts run even if populated.

extract_topics, extract_action_items, and compute_user_sentiment default to false for cost discipline (each enabled flag adds one LLM call per session). Turn on the ones you actually consume.

minMessagesThreshold: 2 skips analysis entirely on dead-air / voicemail-only calls (matches Vapi parity). Increase if you want to skip very short courtesy calls too.

Accessing Results

Analysis runs asynchronously after the call ends. Poll the voice session and check analysis_status:

Code
curl "https://api.neuratel.ai/v1/voice-sessions/{session_id}?include=full" \
  -H "Authorization: Bearer nk_live_your_api_key"

Code
from neuratelai import NeuratelAI
client = NeuratelAI()

session = client.voice_sessions.get(session_id)
status = session["analysis_status"]   # "pending" | "completed" | "failed"

if status == "completed":
    print(session["call_summary"])
    print(session["call_successful"], session["call_successful_rationale"])
    print(session["user_sentiment"], session["user_sentiment_score"])
    print(session["topics_discussed"])
    print(session["action_items"])
    print(session["extracted_data"])

Code
import { NeuratelAI, type Types } from "@neuratelai/sdk";
const client = new NeuratelAI();

const session = (await client.voiceSessions.get(sessionId)) as Types.VoiceSession;
if (session.analysis_status === "completed") {
  console.log(session.call_summary);
  console.log(session.call_successful, session.call_successful_rationale);
  console.log(session.user_sentiment, session.user_sentiment_score);
}

Flat fields exposed on VoiceSessionResponse:

Field	Type	Description
analysis_status	string	"pending" while async analysis is running, "completed" when results land, "failed" on a hard error. The four eval fields below are populated only when completed.
call_summary	string	LLM-generated text summary (controlled by summaryPrompt).
user_sentiment	string	Categorical sentiment: "Positive" / "Neutral" / "Negative" / "Unknown" (set when compute_user_sentiment: true).
user_sentiment_score	float	Numeric sentiment in [-1, 1].
call_successful	string	Success eval verdict: "success" / "failure" / "unknown". Populated when the rubric runs.
call_successful_rationale	string	Human-readable explanation for the call_successful value.
topics_discussed	string[]	Set when extract_topics: true.
action_items	string[]	Set when extract_action_items: true.
extracted_data	object	Matches the JSON schema you supplied in structuredDataSchema.

Don't act on these fields if analysis_status != "completed" — they'll be null until analysis lands.

Receiving via Webhook

Configure a webhook to receive analysis results as they complete:

Code
{
  "analytics": {
    "webhook": {
      "enabled": true,
      "url": "https://your-app.com/webhooks/neuratel",
      "events": ["call.ended", "call.summary.ready"],
      "include_transcript": true,
      "include_analytics": true,
      "include_recording_url": true
    }
  }
}

The call.summary.ready event fires once analysis is complete and includes the full payload (transcript, summary, evaluation, recording URL). The call.ended event fires immediately when the call disconnects, before analysis runs.

Sentiment Timeline

When track_sentiment: true, each turn in the conversation receives a sentiment score from -1.0 (very negative) to 1.0 (very positive). The timeline also includes an overall trend: improving, declining, stable, or volatile.

Access the timeline via GET /v1/voice-sessions/{id}?include=all — look for performance_metrics.sentiment_timeline in the response.

Related

• Webhook Events — full payload reference for call.summary.ready
• Agents — analytics configuration in the full agent schema
• Analytics — dashboard and reporting endpoints