Topic Critique Agent

The Topic Critique Agent is a specialist inside Insight Agent that focuses on the quality of your topic assignment. It inspects a text-to-analyze (TTA) column’s topic collection, diagnoses assignment problems, and — with your confirmation — edits the topic collection to fix them.

​

When to use it

Reach for the Topic Critique Agent when the AI-assigned topics don’t match your expectations. Typical questions it handles:

• “Why isn’t topic X assigned to these rows?”
• “How can I improve the AI score on this column?”
• “Why do so many rows have no topic?”
• “The topics overlap — can you fix them?”
• “Help me improve the topic descriptions.”
• “The sentiment on topic X is often wrong — how do I fix it?”

It can also diagnose and improve per-topic sentiment (positive / neutral / negative) alongside topic labels and descriptions — checking whether sentiment is enabled on the right topics, measuring its accuracy against reviewed rows, and toggling it on or off as part of its proposed edits. If sentiment is disabled at the column level, the agent skips sentiment discussion entirely.

​

How to invoke it

There’s no separate tab — you use it through Insight Agent. Open Insight Agent from the project overview, the Insight Agent tab, or a report, and ask a topic-assignment question. Insight Agent hands the conversation off to the Topic Critique Agent automatically when the question is about assignment quality or the topic collection. If your project has more than one TTA column, mention which column you want reviewed in your prompt — for example, “Critique the topic assignment on the ‘Overall Feedback’ column.”

​

Page context

Insight Agent now knows which page you started the conversation from. If you open it — or click a magic-wand shortcut — from the Topic Assignment page of a specific TTA column, the Topic Critique Agent automatically targets that column, so you don’t need to name it in your prompt.

• Say “Improve the topic descriptions” while viewing the “Overall Feedback” column and the agent will critique Overall Feedback.
• Switch to a different column and ask the same question, and the agent will critique that column instead.
• Page context is a hint per message, never a filter — it isn’t persisted with the conversation, and you can always override it by naming a column explicitly (“Critique the ‘NPS Reason’ column instead”).

​

What the agent does

The Topic Critique Agent works in phases and always explains its reasoning before proposing a change:

1. Analyzes the topic collection — detects overlapping or ambiguous descriptions, labels or categories that contradict the description, descriptions that are too narrow or too broad, and topics whose sentiment configuration looks off (for example sentiment enabled on a purely factual topic, or disabled on one where tone clearly matters).
2. Checks manual reviews — looks at how consistently reviewed rows have been labeled and how much review coverage exists, including how accurate current sentiment assignments are versus reviewer choices.
3. Grounds findings in real rows — pulls concrete example rows and topic statistics so its diagnosis is tied to your actual data.
4. Proposes changes — presents a prioritized list of recommendations (usually description refinements first, then structural changes like merges, splits, adding new topics, or enabling/disabling per-topic sentiment, then manual-review suggestions).
5. Optionally tests before applying — before committing, the agent can run the proposed edits on a small subset of rows so you can see the effect. Subset tests support description and label tweaks, toggling per-topic sentiment, adding brand-new topics, and excluding existing ones — so you can preview structural changes, not just wording changes, before applying them.
6. Edits the topic collection — only after you confirm. All edits from a single conversation are grouped into one entry in the version history, so they can be reverted together.

​

Understanding the AI score

A common reason to call the Topic Critique Agent is to raise a column’s AI score. The score reflects how confident the AI is in its own topic assignments across the whole column — it isn’t a comparison against manually reviewed rows. That means:

• Reviewing more rows on its own does not raise the AI score. Manual reviews are useful for sanity-checking assignments and spotting ambiguous topics, but they don’t change what the model predicts on unreviewed rows.
• Improving topic descriptions almost always does. A low score usually indicates overlapping, ambiguous, or poorly worded descriptions — exactly what the Topic Critique Agent is built to find and fix.

If you’re chasing a higher AI score, ask the agent to critique your topic descriptions first.

​

Reviewing changes the agent made

Every edit the Topic Critique Agent applies is recorded in the topic assignment version history, alongside your manual edits and AI reassignments. Agent edits are marked with a distinct sparkles badge on each affected topic, so you can quickly tell which changes came from the agent. To review or revert them:

1. Open the column’s Topic Assignment page.
2. Click History in the top-right corner.
3. Pick the timestamp just before the agent’s changes.
4. Topics that were edited by the agent are highlighted with the sparkles badge; topics changed by manual or other AI edits use the standard “changed” badge.
5. Use Restore to roll the topic collection back to that point if you want to undo the agent’s changes.

See Topic Assignment History for the full history workflow.

​

Tips for better critiques

• Be specific about the column and the problem. “Why is ‘Delivery Speed’ rarely assigned even though customers mention it a lot?” is much easier for the agent to act on than “the topics look wrong.”
• Let it plan before it edits. The agent is designed to explain its diagnosis first — read the reasoning before confirming an edit.
• Use subset tests for risky changes. For big description rewrites, category restructures, adding brand-new topics, or removing existing ones, ask the agent to validate on a subset first — it will show how the change moves assignments (and sentiment, when applicable) before you commit.
• Follow up with a full run. After confirming edits, kick off a full AI reassignment so every row benefits from the new topic collection.