AI Managed and Authored Anomalies¶

Every anomaly comes from checks, and checks come from two places. AI Managed checks are created by Qualytics from what it observes in the data. Authored checks are written by your team for what the business expects. An anomaly can come from either kind, or from both at once when several checks fail together.

Knowing which kind produced an anomaly helps you judge it. An AI Managed check knows what the data usually looks like, so its anomaly says this is unusual. An Authored check knows a business rule, so its anomaly says this breaks a rule we set. This page explains how the two differ for triage, how to tell them apart, and how your triage decisions feed back into AI Managed checks.

Two Kinds of Checks¶

	AI Managed check	Authored check
Created by	Qualytics. Profile creates most of them from patterns it observes in the data. Volumetric and Freshness checks come from Observability measurements.	A person on your team, in the app or through the API.
What it catches	Data that breaks a pattern it used to follow, such as a field that was always filled in and now has empty values.	Data that breaks a business rule, such as a regulatory limit or an agreed range that the data alone cannot show.
Owner	Qualytics, until someone takes ownership.	The person who owns the rule.
How it changes	Later Profile runs refresh or remove it as the data changes, and Observability keeps Volumetric and Freshness limits up to date.	Only when someone edits it.
What triage changes	Archiving its anomalies as Resolved or Invalid changes how much Qualytics trusts the check.	Nothing. The check stays as it is.

Only Active checks produce anomalies. When the datastore's Oversight setting keeps new AI Managed checks as drafts for review, they produce no anomalies until someone activates them. See AI Managed Checks Introduction and Authored Checks.

The AI Badge¶

When every check that failed in an anomaly is AI Managed, the anomaly shows an AI badge, a purple pill with a four-point star, in the anomaly lists and on the anomaly's page. Hovering it shows AI Detected Anomaly and Identified by AI managed checks. An anomaly with at least one Authored check among its failed checks has no badge.

The badge reflects the checks as they were when the anomaly was created. If an AI Managed check is later edited into an Authored check, the anomalies it already produced keep their badge.

The badge says where the checks came from. It does not say the anomaly is more or less likely to be real, and it does not replace human review.

Finding them in a list

The anomaly lists have no filter by check kind. Scan a list for the AI badge, narrow it with the Rule filter to the rule types you want to review, or open a check's Anomalies tab to see only the anomalies that check produced. See Filter and Sort Anomalies.

Shaping Anomalies from AI Managed Checks¶

You do not have to take over an AI Managed check to control the anomalies it produces. These edits on the check are organizational: they keep it AI Managed, and they apply to the anomalies it creates from then on:

Edit on the check	Effect on its new anomalies
Anomaly Assignee	New anomalies start with that user as an assignee.
Tags	New anomalies start with the check's tags. The tags also set the check's Importance, which raises or lowers the Impact of its anomalies, including the open ones.
Custom Anomaly Description	Record anomaly messages show the value of a column on the same row, instead of the standard wording.
Status	A check switched to Draft stops producing anomalies until it is activated again.

Anomalies that already exist keep their tags and assignees. Edit them on the anomaly itself. See How Anomalies Work.

Editing how the check evaluates data, such as its threshold, coverage, filter, or fields, is different: it turns the check into an Authored check that Profile no longer changes. See Edits that turn an AI Managed check into an Authored check.

How Triage Tunes AI Managed Checks¶

The resolution an anomaly is archived with also tells Qualytics how well each AI Managed check behind it is doing. Qualytics keeps a measure of trust for each AI Managed check, and the resolution moves it:

Resolution	What it tells Qualytics	Effect on the check
Resolved	The check caught a real issue.	Qualytics trusts it more.
Invalid	The check raised a false alarm.	Qualytics trusts it less.
Duplicate	Nothing about the check.	None.
Discarded	Nothing about the check.	None.

Acknowledging an anomaly does not move the trust either. A Full Scan with Auto Resolve Anomalies on counts like a person archiving as Resolved: each anomaly it resolves raises the trust in its checks. So does an anomaly archived by a Flow or by a linked ticket.

Trust matters in two places:

Profile keeps trusted checks. When a later Profile no longer finds the pattern behind a check it created, it removes the check, unless the check is trusted. A check that keeps catching real issues stays, even through a quiet period.
Volumetric limits follow trust. An AI Managed Volumetric check draws tighter limits the more it is trusted, and wider limits the less it is trusted, so a check with many false alarms becomes quieter.

Authored checks are never tuned this way. To change an Authored check, edit it.

Choose the resolution on purpose

Archiving a real issue as Invalid, or a false alarm as Resolved, teaches the check the wrong lesson. When neither fits, for example an anomaly that is no longer relevant, use Discarded.

When the Check Changes¶

An anomaly keeps each failed check as it was when the anomaly was created, so later changes to the check do not rewrite the anomaly:

The check is edited into an Authored check. Its existing anomalies keep their AI badge and their failed check messages. Only new anomalies come from the Authored version.
Profile removes the check. The anomalies it already produced are not removed with it, and their Failed Checks section still shows the check as it was.
Profile creates the original check again. After you edit an AI Managed check into an Authored one, a later Profile can create a new AI Managed check with the original definition beside it. Archive that new check as Invalid to keep Profile from creating it again.

Triage Tips¶

For an AI Managed anomaly, ask whether the change was expected. A new campaign, a migration, or a seasonal peak can explain data that breaks an old pattern. If it was expected, archive the anomaly as Invalid with a comment.
For an Authored anomaly, ask whether the rule still holds. The rule was written on purpose, so a failure usually means a real problem, unless the business has changed the rule.
When the pattern changed for good, update the check. Archiving the same false alarm every week hides the problem. Adjust the check's limit or coverage instead, which makes it Authored.
When both kinds fail together, start with the Authored check. Its message states the business rule that was broken, and the AI Managed failures on the same row usually describe the same problem from another angle.