Reviewing Interactions (The Audit Tab)

The Audit tab is where you review individual evaluated interactions. This is the starting point for daily QA work: finding calls to review, examining scorecards, reading transcripts, and submitting feedback.

4.1 Navigate the Audit Tab

1. Log in to Ema at app.ema.co
2. Navigate to your Agent QA AI Employee
3. Click the Audit tab

The Audit area has three tabs (pills) across the top: Audit, Accepted Feedback, and Accuracy. The sections below cover the Audit table and interaction detail drawer. See the Accuracy tab for monthly report cards and reviewer workflows.

The list syncs automatically, and shows when it last did so (e.g. "Synced a few seconds ago" or "Not synced yet"). Click Refresh at any time to pull the latest data without waiting.

Each row in the interaction list displays:

QA Status Values

Failed Evaluation Error Types

When an evaluation fails due to a problem with the submitted input, the audit row surfaces a specific error type. These are terminal failures -- the system will not retry them, because the outcome cannot change without the caller supplying different input.

4.2 Filters and Sorting

Available Filters

Sorting Options

Pagination

Results display 25 interactions per page. Navigation uses token-based pagination.

4.3 Rerunning an Interaction

Permission required: Builder, Manager, or Admin.

A rerun re-processes an interaction end to end, re-transcribing/re-redacting and re-scoring, and refreshes the same row in place with new verdicts, score, insights, and CSAT. Use it after editing a rule, fixing a processing issue, or when an interaction is stuck or failed.

1. In the Audit tab, select one interaction, or multi-select any number of them.
2. Click Rerun.
3. Confirm the warning -- rerunning deletes any existing feedback on the selected interaction(s). For a multi-select, the warning states how many of the selected interactions have feedback that will be lost.

The row shows "QA in progress" while it reprocesses, then settles with refreshed results. Its Ingested At timestamp updates to the rerun time (floating it to the top of the list), while Call Time never changes.

Rerun Ineligibility Reasons

When a row cannot be re-run, the audit row surfaces a reason. The server reports the first applicable reason in this order: row state (in progress or already queued) takes priority over attempt limit, which takes priority over retention.

4.4 Bulk Actions

Permission required: Builder, Manager, or Admin.

You can select multiple interactions in the Audit tab and act on them together. A Deselect all control clears the entire selection.

Bulk Archive

Select up to 50 interactions at once and click Archive to archive them in bulk. A warning is shown if any of the selected interactions have pending Continuous Learning (CL) data that will be lost when archived.

Bulk Delete

Select multiple interactions and click Delete to delete their evaluations in bulk. This calls the BulkDeleteQATranscriptEvaluation RPC and permanently removes the selected evaluation records.

4.5 View Interaction Details

Click View Details on any interaction to open the detailed view.

When the interaction has open flagged rule results, the drawer shows a review queue position (for example, "1 of 5") so you can step through flagged interactions without returning to the table. Use the forward and back arrows to move to the next or previous flagged interaction in the queue.

Interaction Summary

A collapsible panel at the top of the detail view gives an AI-generated plain-language summary of the call -- what the customer needed, what the agent did, and how it was resolved. It uses the same PII redaction as the transcript, so sensitive values appear as redaction placeholders (e.g. a masked name or account number) rather than the original text.

The drawer header also shows a language badge indicating the transcription (input) language used for the call. The badge displays the language the caller chose for speech-to-text (e.g. "Spanish" for a call submitted with language code es). When no language was specified and auto-detect ran, the badge shows "Auto" followed by the language that speech-to-text found (e.g. "Auto · Portuguese"). If auto-detect ran but no language was identified, or if no language information is available for the interaction, the badge is not shown.

Below the summary, the detail view has the following tabs:

Scorecard Tab

The core evaluation view. A summary bar at the top shows the total parameter count with Pass and Fail totals.

The scorecard tab offers two views you can switch between. Your preference is saved locally and persists across sessions.

Flagged Rule Results

Rule evaluation cards with an amber border have been flagged for manual review because the model's confidence was low or the transcript did not contain enough data to support a verdict. Each flagged card shows the flag reason:

Expand the flag line on a flagged card to see What's uncertain -- a one-sentence note from the model explaining what it was unsure about. For a categorical rule, the note names the categories that appeared to fit.

Each flagged card shows confidence and data sufficiency. Confidence is displayed as a word: High (0.85 or above), Medium (0.70 to below 0.85), or Low (below 0.70).

Reviewer Decisions

When a rule result is flagged, the card shows actions for a reviewer to decide it. The available actions depend on the rule type:

Each decision resolves that flag. When all flagged results on an interaction are resolved, the Reviewed mark appears on the row in the table, and any provisional score indicator is removed.

You can change a decision at any time by reopening the card and selecting a different action.

Provisional Score

When a pass/fail rule result is flagged and no decision has been made yet, the QA score is provisional -- the model's answer counts toward the score, but a reviewer decision can still change it. A provisional score is shown with a dashed border on the score badge, both in the table and in the drawer. Once all pass/fail flags on the interaction are resolved, the dashed border is removed and the score is final.

Classic View

New View (ScorecardV2)

The new view opens by default. Failed rules are expanded automatically, since those are what a reviewer typically looks for first. A New badge in the toolbar opens a "What's new" popover describing the changes. Click Switch to classic view to go back.

Each rule row in the new view also shows the model's confidence level (High, Medium, or Low) and data sufficiency for that result.

Transcript Tab

Full conversation text with speaker labels and timestamps. Every speaker the transcript carries is shown under its own label -- not just Agent and Customer. A supervisor, an IVR/bot segment, or an external third party (e.g. a hotel receptionist or interpreter brought into the call) appears under its own descriptive label. QA scoring still evaluates only the interaction's designated agent; every other party, including third parties, is treated as context.

Strengths Tab

Positive agent behaviors identified during the interaction.

Opportunities Tab

Areas for improvement with coaching recommendations.

CSAT Mapping Tab

Shows Ema's AI-derived customer satisfaction analysis for the interaction, plus any customer-submitted survey feedback ingested via the CSAT Ingestion API.

Additional Data

4.6 Feedback Interaction Deeplink

When feedback is accepted on a scorecard parameter, the feedback item now links directly back to the source interaction in the Audit tab. Clicking the link opens the interaction's detail view, so you can navigate from an accepted suggestion straight to the call it came from without searching manually.