Post-call analysis

When a call ends, Tried reads the transcript with an LLM and writes structured results to the call's call_analysis. Four preset fields are always produced: a summary, the caller's sentiment, whether the call succeeded, and whether it reached voicemail. On top of those you define your own fields (text, selector, boolean, number) to pull out the facts your business needs, such as whether an appointment was booked or why the person called.

Use post-call analysis to:

• Read a one-line summary instead of a full transcript in Call history.
• Filter and chart calls by outcome in Analytics.
• Send extracted data to your systems through the call_analyzed webhook.
• Fill contact fields from what was said. See Contacts.
• Choose which calls AI QA scores. See AI QA overview.

Where to configure it#

Open an agent and expand Post-Call Extraction in the settings column. The section shows:

• The three presets, Call Summary, Call Successful and User Sentiment, each marked Preset.
• Call Summary prompt, Call Successful prompt and User Sentiment prompt, each showing Default or Custom. Click one to write your own instructions.
• Your custom fields with their type, an If badge when the field has a condition, and edit and delete buttons.
• Add to create a field, Backfill to re-run analysis on past calls, and the analysis model selector.

Analysis settings are part of the agent, so they are versioned with it. Each call uses the settings of the agent version that handled it.

What analysis produces#

{
  "call_summary": "The caller booked a check-up for Tuesday at 4pm and asked about parking.",
  "in_voicemail": false,
  "user_sentiment": "Positive",
  "call_successful": true,
  "custom_analysis_data": {
    "appointment_booked": true,
    "call_reason": "New booking",
    "caller_name": "Sam Taylor"
  }
}

Preset prompts#

Each preset prompt is optional, up to 2000 characters, and is added to the analysis instructions:

The Call Successful prompt is phrased as the condition for success: the model is told that call_successful is true only when it holds. Leave a prompt empty to use the default behaviour.

Custom fields#

Custom fields live in the agent's post_call_analysis_data array. Each has a type, a name, a description, optional examples or choices, a Required flag and an optional Only extract when condition. See Define extraction fields.

The analysis model#

post_call_analysis_model picks the LLM that runs analysis. The default is gpt-5.6-terra. Any model in the agent model list can be used: OpenAI, Anthropic Claude or Google Gemini models. The selector sits next to Backfill in the section.

When analysis runs#

1. The call ends and call_ended is sent to your webhook.
2. Analysis runs on the full transcript together with the disconnection reason.
3. call_analysis is saved. user_sentiment and call_successful are also stored for filtering and charts. For a batch call, the batch's successful count is updated.
4. call_analyzed is sent, always after call_ended.
5. In the background: AI QA scores the call for every cohort it matches, the post-call workflow runs, and contact field mappings, contact memory and CRM sync run.

Analysis runs whenever a call ends through a runtime or telephony event, including calls that never connected (busy, no answer, a web caller who never joined). Calls that are stored as failed at creation without an end event (dial_failed with call_status: error, and batch attempts refused with no_valid_payment) are not analysed.

Calls with no conversation#

If the transcript is empty, no LLM request is made. The summary reads "The call ended before any conversation took place.", sentiment is Unknown, and call_successful is false.

When the LLM fails#

A failed request is retried once. If it fails again, the result says so instead of guessing: the summary reads "Post-call analysis could not be generated because the AI request failed.", sentiment is Unknown, call_successful is false and custom_analysis_data is empty. in_voicemail still follows the disconnection reason. You can re-run analysis later. See Re-run and backfill analysis.

Chats#

Chat agents use the same engine when a chat ends, with post_chat_analysis_data and post_chat_analysis_model. The result is chat_analysis with chat_summary, chat_successful, user_sentiment and custom_analysis_data. See Chat history.

Pages in this section#