Skip to main content
Call Analysis and Data Extraction

Post-Call Analysis: Summary, Sentiment and Call Success

Get a summary, the caller's sentiment and a success verdict for every VAW Native call, and choose the AI model that writes them.

Written By voiceaiwrapper

Last updated 26 days ago

Post-call analysis reads each finished call and writes a short report: what happened, how the caller felt, and whether the call did its job. You can then skim calls without listening to recordings, and filter Call Logs to find the calls that need attention, such as unhappy callers or failed bookings.

Analysis is on for every new agent. This article explains what each part does, how to adjust it, and where the results appear.

Applies to: agents on a VAW Native Pod (an engine or provider connection). Vapi, Retell AI and ElevenLabs agents use their provider's own call analysis.

Before you start

  • An agent on a VAW Native Pod. See Create an Agent from Scratch.
  • AI credits in your account. Each analysis run uses credits. See "What analysis costs" below.

Open the Analysis tab

  1. In the sidebar, open Agents and click your agent.
  2. Open the Analysis tab.
  3. Make your changes, then click Save changes. Calls that end after you save use the new settings.

The left side holds the settings. The right side, What you'll see, shows an example report built from sample data, so you can see the effect of each setting before any real call happens.

Agent page β†’ Analysis tab with analysis on. Left side: the Analyze calls switch (on), the AI model for analysis and extractions dropdown open so the credits/run figure next to each model is visible, and the three rows Summary, User sentiment and Call success, all switched on. Right side: the What you'll see panel showing the sample report. The Call Data Extractions card is visible at the bottom.

The settings at a glance

SettingWhat it doesNew agents start with
Analyze callsTurns the whole report on or off. "Analysis runs after the call ends."On
AI model for analysis and extractionsThe model that reads the transcript and writes the report.GPT-5 mini
SummaryTwo or three sentences about the call.On
User sentimentHow the caller seemed to feel.On, reported as Positive / Neutral / Negative
Call successWhether the call achieved what the agent is there for.On, judged as Pass or fail

Analyze calls

This is the master switch. When it's off, the tab shows: "Analysis is off. Calls get no report, and attached extractions do not run." Your other settings are kept and come back as soon as you turn it on again.

AI model for analysis and extractions

Pick the model that writes the report. The list holds the same models your agent can talk with, plus GPT-5.6 Sol, which is offered only for analysis. Next to each model you'll see an estimate of its cost in credits/run. The estimate is per report, not per minute of the call.

The same model also fills in any Call Data Extractions attached to this agent. Larger models cost more per run. New agents start on GPT-5 mini.

Summary, User sentiment and Call success

Under Select what you want in your call analysis, each part has its own switch. Click a row to open its settings.

  • Summary: "Two or three sentences about the call, written for a person skimming the log." Under Add your own instructions, you can tell it what to always cover, for example "Always mention any objections the caller raised."
  • User sentiment: "How the caller seemed to feel, judged only from what they said." Under How should it be reported?, choose one of the formats below. You can add instructions here too, for example "Frustration with hold time counts as negative."
  • Call success: whether the call did what the agent is for. You choose how it's judged (pass or fail, a score, a checklist and more). See Define What Counts as a Successful Call.

Instructions are optional and can be up to 1,000 characters. They add to the standard behaviour; they never replace it.

Sentiment formatWhat you get
Positive / Neutral / Negative (default)The standard three-way read.
Good / BadA simple two-way read. Good shows as Positive, Bad as Negative.
Score from 1 to 10A number for sorting and charting. 7 and up shows as Positive, 3 and below as Negative, anything between as Neutral.

When a caller says too little to judge, sentiment comes back as Unknown rather than a guess. Call success works the same way.

Below these rows, Custom data extracted from your call lets you add your own fields to the report, such as a booking date or an order number. See Capture Structured Data with Call Data Extractions.

When analysis runs, and when it's skipped

The report is written about a minute after the call ends. While a call is live, the Analysis column in Call Logs shows After the call; then Pending or Running; then Completed.

Before spending any credits, VoiceAIWrapper checks that there's a conversation worth reading. A report needs at least one turn from each side: one thing the caller said and one thing the agent said. If the check fails, the call shows Skipped with a reason:

Reason shownWhat happened
Call never connectedNobody answered, so there was nothing to read.
No conversation recordedThe call connected, but nobody said anything.
Conversation too shortOnly one side spoke, for example the agent's greeting and then a hang-up.
Analysis disabledAnalyze calls was off for this agent.
Agent deletedThe agent was deleted, and its analysis settings went with it.
Transcript deleted by retentionThe transcript was removed under your Pod's retention settings.

Skipped calls cost nothing for analysis. Call Data Extractions only run on calls whose analysis completed.

If a report shows Failed, the AI model couldn't produce one. The transcript and recording are unaffected, and the sheet shows the error.

Where the results appear

In the call sheet. Open any call in Call Logs and select the Analysis tab. You'll see the Summary, the User sentiment and Call success badges with a one-line reason for each, and any Custom data from extractions. The footer says which model wrote the report, and How this is judged takes you straight to the agent's Analysis tab. For more on the call sheet, see Call Details: Transcript, Recording, Analysis, Variables and What Ran.

Call Logs β†’ a finished call opened in the call sheet, Analysis tab selected. Shows the Summary text, a User sentiment badge (for example Positive) with its one-line reason, a Call success badge (for example Success) with its reason, and the footer line 'Written by gpt-5-mini Β· N tokens.' (the model's ID, as the sheet shows it) with the How this is judged link. Blur the caller's name and phone number in the sheet header.

In Call Logs. The list has Analysis, Call success and User sentiment columns, and filters with the same names. Call success shows Success, Failure or Unknown whatever scale you chose, so you can filter on it across agents. See Find and Filter Calls in Call Logs.

In chats. Chats with your agent get the same report, using the same settings. A chat is analysed once the person has sent at least one message. You'll find it on the chat's Analysis tab in Chat Logs.

In campaigns and webhooks. A campaign can count a call as successful when the analysis says it succeeded, and the agent's webhook receives the full report once analysis finishes. See Define a Campaign Success Metric.

What analysis costs

Each report is one run of the analysis model, charged in AI credits. The credits/run figure in the model menu is an estimate; the charge is based on what the run actually used. The cost is added to the call's own cost, so it appears in that call's Credits figure in Call Logs. Skipped calls aren't charged for analysis. Each attached extraction is a separate run and is charged separately.

To spend less, keep the default model, or turn off Analyze calls on agents whose calls you never review, such as internal test agents.

Frequently asked questions

No. Calls that end after you click Save changes use the new settings. Earlier calls keep the report they already have.

The check counts turns in the transcript, not sound on the line. If nothing the caller said made it into the transcript, the call counts as one-sided. Check the transcript in the call sheet to see what was captured.

Yes. What a client can see and change on their workspace is set in the client's permissions. See Set Client Permissions on VAW Native Workspaces.