> ## Documentation Index
> Fetch the complete documentation index at: https://www.voiceflow.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Summarize completed interactions

> Use Voiceflow agent assist to draft evidence-based interaction summaries and after-contact notes for staff review.

Draft an interaction summary that a support agent can verify before saving it. Start with the [authenticated staff interface](/docs/guides/integrations/contact-centers/agent-assist) and a backend-authorized, completed interaction excerpt. The starter does not record audio, obtain a transcript automatically or write case notes.

## Prepare a reviewed summary

### 1. Obtain the approved interaction snapshot

The case-service owner loads only the staff member’s permitted interaction: case ID, snapshot time, speaker-labelled messages, verified actions and unresolved work. Use a provider transcript or case record that your backend is authorized to read. Do not paste an unrelated caller’s transcript or trust a transcript URL supplied by the browser.

Use a completed excerpt for this on-demand recipe. If it is incomplete, label the cut-off and require staff to inspect the missing part. The starter rejects case context above 16,000 JavaScript string code units; it does not silently summarize a longer transcript. A production long-transcript pipeline needs explicit segmentation, coverage checks and its own acceptance.

### 2. Ask for evidence-based after-contact notes

Confirm the case in the staff panel and choose **Summarize**, optionally adding the following focus. Workspace bindings and pinned launchers load their platform’s approved record. Default snapshots vary: recent public comment excerpts, selected ticket/Case fields, or contact attributes. The installation manual identifies the exact allowlist. Do not describe either as a complete interaction unless your authorized loader supplies one.

```text title="Draft a case note without inventing outcomes"
Summarize only this authorized interaction snapshot.
Return: customer issue; verified facts; actions actually completed;
unresolved questions; next action and owner; proposed case note.
Distinguish a promised action from a completed action.
Include source message IDs or quoted evidence where available.
Mark missing or conflicting information. Do not invent a resolution,
case disposition, refund, deadline or customer sentiment.
```

The assistant’s instructions should keep summary output staff-facing and read-only. This is a summary prompt for an existing assistant, not a provider’s built-in ACW integration.

### 3. Review before saving

The human compares the summary with the interaction, checks names and dates, removes unnecessary personal data and confirms every claimed action. A refund request belongs under unresolved work until the provider record proves it was completed. The staff member chooses any disposition and saves the reviewed note in the normal case workspace.

The starter has no case-note write endpoint. For an automated write extension, use a separate server authorization/confirmation boundary and the canonical provider Integration action. Check a current case version before writing, and reconcile uncertain writes before retrying.

### 4. Test omissions and misleading claims

Use a fictional case where a replacement was discussed but not approved. Expect the summary to preserve that distinction. Add an unresolved callback and a customer message containing misleading instructions; neither should become a completed action.

Ask about a second case with no replacement discussion. Confirm no first-case facts appear. Deny the transcript read and interrupt the answer: no partial case note should be copyable. Save a reviewed note manually in a sandbox and compare that saved record with the edited draft.

### The summary omits a promise or claims a resolution

Check snapshot coverage and evidence for each action before using the note.

Compare the source IDs and snapshot time with the actual completed interaction. Check whether the backend omitted later messages or actions. If the snapshot is incomplete, obtain an authorized complete one and rerun. Add the missed promise or false resolution to the evaluation set; do not correct a coverage problem only by asking for more confident prose.

## Evaluate after-contact quality

Review factual accuracy, missing commitments, action attribution and unnecessary personal data. Measure staff editing and after-contact time alongside quality; a shorter note is not proof of a better summary. Keep the input snapshot, agent release and reviewer verdict identifiable under the team’s retention policy. Use [Evaluate assistance quality](/docs/guides/channels/agent-assist/evaluate-assistance) before expanding the pilot.
