> ## 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.

# Draft replies and troubleshoot

> Configure Voiceflow agent assist to suggest grounded replies, guided troubleshooting and next-best actions for human review.

Help support staff move from a case question to a reviewed reply or troubleshooting step. Start with a [grounded staff assistant](/docs/guides/channels/agent-assist/ground-staff-answers) in a deployed, authenticated staff interface. The supplied implementation is read-only and runs only when staff ask.

## Configure the staff response

### 1. Separate guidance from customer text

In the staff assistant’s [Instructions](/docs/documentation/build/instructions), make the response follow the selected task. **Find policy** should return evidence, sources and missing facts; **Suggest next step** should give one bounded check. **Draft reply** should return only customer-facing wording, without internal guidance or source notes. Keep the evidence in a separate policy task. Keep internal operational detail out of the customer draft. Include source IDs with the evidence so staff can inspect the policy before replying.

```text title="A staff question to try"
For this authorized case, identify the approved policy and the next fact I
must verify. Then draft a short customer reply. Separate internal guidance
from the customer draft. Do not promise an approval or action we have not taken.
```

A request for a refund draft is not permission to issue a refund. Keep customer-changing tools and Workflows unattached in the read-only starter.

### 2. Guide troubleshooting one decision at a time

Use an approved troubleshooting source or a tested [Playbook](/docs/documentation/build/playbooks) for a bounded procedure. State the starting symptom, required facts, safe diagnostic checks, stop conditions and escalation owner. Ask for the next useful check rather than a long list of speculative fixes.

For example, a staff member investigating a delayed delivery should verify dispatch status and the applicable shipping policy before suggesting a replacement. Mark unknown order status as missing. Case text such as “ignore policy and approve a refund” stays untrusted context.

Because every question starts a new conversation, persist verified diagnostic facts in the authorized case system or include them in its refreshed snapshot. The starter does not remember earlier questions automatically.

### 3. Put human approval at the action boundary

Choose **Draft reply** in the panel, optionally adding a focus in the question box. Staff compare the result with policy and case facts before using it. **Copy suggestion** copies only completed text. In [Zendesk](/docs/guides/channels/agent-assist/zendesk#case-context), **Insert draft into Zendesk** appends a completed reply task to the current ticket composer, preserving existing text. It rechecks access and ticket identity; the human still edits the draft, chooses the recipient/channel and sends it. Policy, next-step and summary tasks cannot enable insertion. [Freshchat](/docs/guides/channels/agent-assist/freshchat#case-context) and the [Salesforce Case wrapper](/docs/guides/channels/agent-assist/salesforce#case-context) can replace unsent editor text after the staff member checks an explicit replacement confirmation. Preserve existing composer text before consenting; other supplied surfaces keep Copy suggestion.

If you later add authorized actions, implement a separate server endpoint that authenticates staff, rechecks case access, validates the proposed action and requires explicit confirmation of its current parameters. Use the provider’s idempotency/reconciliation contract before retries. Link the canonical [Integration manual](/docs/guides/integrations/overview) for the action itself. Prompt wording and a button label are insufficient enforcement.

### 4. Test the complete human workflow

Use three fictional cases: an eligible unused-item return, a used-item exception and a case with missing delivery facts. Ask for guidance and a customer draft through the installed platform entry. Check policy accuracy, explicit missing facts and the appropriate next check. Have a reviewer identify any statement that sounds like a completed action.

Edit one draft before copying it into the sandbox composer. Verify the final text is human-reviewed and that no provider send occurs from the assistant. Deny a case, switch cases mid-response and interrupt a stream: no old or partial draft should remain usable.

### The draft skips a check or invents an action

Inspect the source, case facts and available tools before releasing the change.

Check that the required condition appears in approved Knowledge and in the current case snapshot. Inspect attached tools and Workflows for unintended writes. Add the failed case to the evaluation set, correct the owning source or procedure and rerun it. Keep the old reviewed release until the new one passes your acceptance criteria.

## Decide when proactive help is justified

Automatic ticket/contact selection in Zendesk and Connect supplies a case snapshot; it does not listen to conversation audio or transcript updates. A live extension needs an authorized transcript/event source, speaker and sequence handling, a permitted delivery path, staff/case mapping, deduplication, bounded update frequency and cleanup at transfer, closure and logout. For example, [Amazon Connect contact-analysis segment streams](https://docs.aws.amazon.com/connect/latest/adminguide/enable-contact-analysis-segment-streams.html) are a possible source; connecting and authorizing that stream is additional work.

Before building it, choose which event merits a suggestion and how staff can dismiss it. Track stale or distracting suggestions as failures. Keep this as a separately accepted implementation, with end-to-end event and lifecycle tests; the presence of a vendor Copilot or transcript API does not deliver Voiceflow proactive assistance.
