AI Triage: Anomaly Detection

How VersionForge uses AI to detect anomalies in sync changes and prioritize the Safety Gate review queue.

What AI Triage Does

When a sync pipeline produces changes for the Safety Gate review queue, the volume can range from a handful of rows to thousands. Reviewing every row manually is not realistic at scale. AI Triage analyzes the incoming changes and flags the ones that look unusual -- the rows that a human reviewer should look at first.

AI Triage does not approve or reject anything. It annotates changes with anomaly scores and priority labels so that reviewers focus their time where it matters most.

What Counts as an Anomaly

VersionForge's anomaly detection evaluates each change against multiple signals:

Unusual magnitude -- A salary figure that jumped 40% in one period, a GL balance that doubled, or a cost center budget that went negative. The model compares the change against the historical range for that field and flags outliers.
Pattern deviation -- A department that has posted consistent monthly expenses suddenly shows zero activity, or a revenue line that has always been positive flips to a credit. The model looks for breaks in established patterns.
Sensitive field changes -- Certain fields carry higher inherent risk: compensation rates, bank account numbers, intercompany elimination accounts, executive cost centers. Changes to these fields receive elevated scrutiny regardless of magnitude.
First-time values -- A new account code, cost center, or dimension member that has never appeared in prior syncs. New values are not necessarily wrong, but they warrant a second look to confirm they map correctly.

Priority Scoring

Each flagged change receives a priority label:

| Priority | Criteria | Reviewer Action | |----------|----------|-----------------| | HIGH | Large magnitude outlier, sensitive field change, or multiple anomaly signals on the same row | Review immediately -- do not batch-approve | | MEDIUM | Moderate deviation from historical pattern or a first-time dimension value | Review before approving the batch | | LOW | Minor pattern deviation that is likely legitimate but worth noting | Review if time permits; safe to batch-approve in most cases |

Rows with no anomaly signals are left unmarked and can be approved in bulk.

AI Triage priority labels are suggestions, not gates. You can approve a HIGH-priority row if you determine the change is legitimate. The label is there to direct your attention, not block your workflow.

How AI Triage Complements Manual Risk Scoring

VersionForge also supports manual risk scoring rules -- for example, "flag any change where the amount exceeds $100,000" or "flag any row touching the intercompany elimination account." AI Triage works alongside these rules:

Manual rules catch known, predictable risks that you can define in advance.
AI Triage catches unknown or emergent risks that do not fit a predefined rule -- the anomalies you did not think to write a rule for.

Both types of flags appear as annotations in the review queue. Manual rule flags show the rule name; AI Triage flags show the anomaly type and confidence score.

Where Triage Results Appear

In the Safety Gate review queue, flagged rows display a colored badge next to the row:

Red badge -- HIGH priority
Amber badge -- MEDIUM priority
Blue badge -- LOW priority

Click the badge to expand the triage detail panel, which shows:

The specific anomaly type(s) detected
The historical comparison data (e.g., "This field averaged $12,400 over the last 6 periods; the current value is $31,000")
A confidence score from 0 to 100 indicating how certain the model is that this is a genuine anomaly

AI Triage requires at least three prior sync runs with approved data to build a meaningful baseline. On new connections, triage annotations will not appear until sufficient history exists.

Tuning Triage Sensitivity