Conversation Review & Extraction Agent

Set up the chain-only reviewer to classify completed conversations and extract explicitly supported facts into private structured outputs.

Conversation Review & Extraction

Use this internal agent to turn a completed customer conversation into a compact, structured record for your team. It runs as a chain-only step after a conversational agent completes. It does not reply to the customer, ask follow-up questions, or take over the conversation.

Configure the reviewer

  1. In the chatbot dashboard, open Agents and enable Conversation Review & Extraction.
  2. Choose the classifications that are useful to your team: Sentiment, Intent, Resolution, Urgency, and Follow-up needed.
  3. Add only the custom fields you plan to use. Give each field a clear name and an instruction that defines the value to extract. Supported types are Text, Number, Yes / No, Date, and Choice. For a Choice field, define the allowed choices.
  4. Keep Only use explicitly stated facts enabled when you need conservative, auditable extraction. With it enabled, unsupported custom-field values are left empty instead of guessed.
  5. Set the maximum input and output token limits for the review. The input limit bounds the transcript supplied to the reviewer; the output limit bounds its structured result. These are model-processing limits, not a promise that every transcript will fit in full.
  6. Add the reviewer as an Agent step in the completion chain of the conversational agent you want reviewed.

Design the upstream conversation for good data

The reviewer can only analyze the saved conversation and any form data passed by the chain. It cannot recover information the visitor did not provide. If you need a visitor's region, product interest, or preferred contact method, configure the customer-facing agent or a smart form to ask for it. Describe the field narrowly—for example, “Return a suburb only when the visitor explicitly names one”—and avoid collecting sensitive details you do not need.

The reviewer adds evidence quotes and confidence values to its result. Treat these as review aids, not as a substitute for checking the original conversation before making a consequential decision.

Find and retry results

Open Submissions → Agent Outputs for the chatbot. Each result includes its status, summary, classifications, custom fields, evidence, source-chat link, and model-processing counts. You can expand the structured result, search or filter it, and export results to CSV.

If a review fails, open its failed output and choose Retry review. This reruns only the reviewer against the saved thread using the current reviewer settings; it does not replay the customer conversation or earlier chain steps. If the source conversation is missing or was not persisted, the review cannot be reconstructed from the output row alone.

For the full chain behavior and the other Submissions sections, see Workflows & Submissions.