> ## 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.

# Investigate a failed customer conversation

> Follow one low Evaluation score into the customer exchange and execution logs. The worked example traces an incorrect 60-day return promise against a 30-day policy and…

Follow one low Evaluation score into the customer exchange and execution logs. The worked example traces an incorrect 60-day return promise against a 30-day policy and an ineligible order lookup.

## Before you begin

- A scored [Evaluation](/docs/documentation/measure/evaluations#viewing-evaluation-results) and access to its saved [Transcript](/docs/documentation/measure/transcripts).
- The expected policy or behavior. The example uses the [Return policy accuracy criterion](/docs/guides/studio/monitoring/check-support-policy).

## Trace the result to its evidence

### 1. Find a low-scoring conversation

Open **Return policy accuracy** from **Analytics**. Confirm the **Environment** and dates in **Results**. Click the table’s **Result** heading to put low scores first, then choose a **1** result.

Find a low policy-accuracy score — Sort policy scores to find a conversation that needs investigation.

### 2. Read the Evaluation log

Click the result row to open its **Evaluation log**. Read the judgment and reasoning alongside the conversation. For this case, the agent promised a 60-day window although the policy permits 30 days.

Use the **Prompt** tab when you need to check the exact criterion. Return to **Results** and reopen the same row before continuing.

Inspect the selected Evaluation log — The selected Evaluation log ties the low score to the incorrect policy promise.

### 3. Inspect the same Transcript and lookup

Click **View full transcript**. Confirm that it contains the same customer request and response as the Evaluation log. Click **Show logs** (the bug icon), then expand **Look up order** to inspect its returned data.

In this example, order **VF-4821** arrived **35 days** earlier. The lookup returned a **30-day** return window and **return_eligible: false**, but the response said the order was eligible within 60 days. The failure is in the agent’s response; the lookup supplied the correct restriction.

Trace the same policy failure in its Transcript — The same conversation and lookup establish the policy contradiction.

### 4. Record a reproducible finding

Record the conversation identifier, time, Environment, expected behavior and observed response. Include the relevant log result.

For the example: “The agent promised eligibility after 35 days using a 60-day window. The policy and order lookup specify 30 days. The expected response is to explain the restriction and offer a specialist review.”

## Verify the investigation

- The finding cites the selected conversation and its execution evidence.
- The policy criterion and lookup support the finding; it is not inferred from the aggregate score.
- The evidence identifies which response or operation needs correction.

## Act on the finding

- For response guidance, update the owning [Instructions](/docs/documentation/build/instructions) or [Playbook](/docs/documentation/build/playbooks).
- For incorrect tool data, investigate the tool’s input and returned result.
- For a correct response that was penalized, refine and retest the [Evaluation criteria](/docs/documentation/measure/evaluations#creating-an-evaluation).
- Preserve a confirmed agent failure as a [Test](/docs/documentation/measure/tests) before making the correction.

## If the conversation is missing

Recheck the Environment and dates, then review [Transcript saving settings](/docs/documentation/measure/transcripts#configuring-transcript-settings). Enabling saving now cannot recreate a conversation that was not saved.
