# Review AI analysis proposals (/core/build/analyze-your-database-with-ai/how-to/review-ai-analysis-proposals)













<CalloutContainer type="info">
  <CalloutDescription>
    **Target Audience**: Non technical users
  </CalloutDescription>
</CalloutContainer>

### Goal [#goal]

Review the metadata proposals produced by an AI analysis run, decide what to keep or discard, and confirm each section so accepted changes update your metadata in Toucan AI.

***

### Prerequisites [#prerequisites]

* At least one [AI analysis run](/core/build/analyze-your-database-with-ai/how-to/analyze-your-database-with-ai) has completed on the database.
* You have access to the Databases section as a builder or admin.

***

### Steps [#steps]

#### 1. Open the AI analyses list [#1-open-the-ai-analyses-list]

* Navigate to **Databases** and open the database you analyzed.
* Open the **More options** menu (⋮) in the database header.
* Select **AI analyses**.

You land on a list of every analysis run for that database, newest first. Runs are never replaced by newer ones: you can always go back to an earlier review.

<img alt="AI analyses list keeping every past run, newest first" src="__img0" />

Each row shows:

* The **scope** of the run (which schemas and tables were selected)
* When it started and who started it
* A status badge when action is needed (**To review**)

<img alt="Analysis run carrying the To review badge" src="__img1" />

#### 2. Open a completed run [#2-open-a-completed-run]

* Click a run marked **To review** (or any completed run that still has pending proposals).
* If the run is still **running**, you see live progress and can **Stop the analysis**. Proposals are only reviewable after the run completes successfully.
* If the run **failed** or was **cancelled**, open it to see how far the pipeline got; there is nothing to confirm until you launch a new analysis.

#### 3. Navigate review sections [#3-navigate-review-sections]

A completed run with pending proposals opens a review layout:

* A **sidebar** lists review sections that still need attention (in review order):
  * **Type casts** — type casts inferred for columns. For example, text to datetime
  * **Display names** — human-readable labels for the UI and chat
  * **Descriptions** — short explanations of what columns represent
  * **Relationships** — suggested joins between tables (reviewed separately; see step 6)
* The main pane shows a **diff table** for the selected review section: current value vs proposed value, plus **Justification** (rationale) and **Confidence** when available — shown only to help you during review.

<CalloutContainer type="info">
  <CalloutDescription>
    Justification and confidence are not saved with your metadata and are not available to the AI assistant.
  </CalloutDescription>
</CalloutContainer>

Only review sections that still have something to review appear in the sidebar. When every proposal for a section is confirmed and applied, that section disappears from the list.

#### 4. Decide on proposals [#4-decide-on-proposals]

For each row you can:

* **Accept** the proposal (keep it for application)
* **Reject** it (discard it; it will not be applied to your metadata)
* Leave it **undecided** for later

You can mix accepts and rejects in the same review section before confirming.

Use the sticky action bar at the top of the pane:

* **Accept all** / **Reject all** — apply to the rows currently visible (respects the table filter)
* **Reset** — clear local decisions on the visible rows
* Badges show how many proposals are **to apply** and **to discard** across the whole review section (including decisions made under a previous filter)

Optionally filter the table by one or more tables so you can review a large run in smaller batches.

<CalloutContainer type="info">
  <CalloutDescription>
    Leaving a proposal undecided does not block confirmation of the others. Undecided rows stay pending for a later pass on the same review section.
  </CalloutDescription>
</CalloutContainer>

#### 5. Confirm and apply a section [#5-confirm-and-apply-a-section]

* When you are ready, click the confirm button for that review section (for example **Confirm type casts**).
* Confirm the dialog (&#x2A;*Confirm this section?**). Toucan AI locks that section and starts applying accepted proposals; rejected ones are removed. Apply cannot be cancelled.
* While applying, the UI shows that changes are being applied. Rows disappear as writes land on your metadata.
* If apply fails part-way, what already landed stays applied; remaining confirmed proposals are retried on the next apply of that section.

<img alt="Confirm this section dialog over the type casts diff table" src="__img2" />

Repeat for each remaining column review section until the run shows nothing left to review for casts, display names, and descriptions.

#### 6. Relationship suggestions [#6-relationship-suggestions]

During analysis, Toucan AI also looks for **relationships** between tables. From the run review sidebar, open **Relationships** to land on **Suggested joins**.

* Decide row by row (**Keep** / **Reject**), or use **Keep all** / **Reject all** / **Reset**.
* Each suggestion shows relevance (confidence) and a justification.
* Click **Confirm** to apply kept joins. Suggested joins are not applied through the column proposal confirm buttons.

<img alt="Suggested joins tab with relevance and justification per join" src="__img3" />

#### 7. When everything is reviewed [#7-when-everything-is-reviewed]

After all column proposals for a run are confirmed, the AI analyses list shows **All AI proposals have been reviewed** for that run (instead of **To review**).

<img alt="Analysis run showing All AI proposals have been reviewed" src="__img4" />

***

### Edge cases [#edge-cases]

| Situation                           | What to do                                                                                                                  |
| ----------------------------------- | --------------------------------------------------------------------------------------------------------------------------- |
| Run completed with **no proposals** | Nothing to review: the analysis found nothing to change, or everything was already confirmed.                               |
| Run **failed** or **cancelled**     | Launch a new analysis on the same or a smaller set of tables. Previous completed runs and their proposals remain available. |
| Partial re-analysis                 | Starting a new run on a subset of tables does **not** delete older runs. Review each run independently.                     |
| Schema changed after the run        | Prefer a new analysis (and/or **Refresh Schema**) so proposals match the current structure.                                 |

***

### Constraints [#constraints]

* Proposals are **not** applied until you confirm a review section.
* Confirmation is **per review section**: you can apply casts today and leave descriptions for tomorrow.
* Analysis consumes [AI credits](/core/credits-and-usage); reviewing and confirming does not add a separate enrichment charge.

***

### Conclusion [#conclusion]

Confirmed proposals update your semantic metadata. Tables included in the analysis are marked **Available for chat and charts** by default.

**Suggested Next Step**: [Complete metadata information](/core/build/semantic-layer/how-to/complete-metadata-information) or [Table Visibility Governance](/core/govern/table-visibility-governance)
