Self-watching metrics

A certified metric does not sit still waiting to be looked at. It runs on a schedule, keeps a history, and watches itself for anomalies.

Scheduled runs

Metrics execute on a cadence you set. Each run records the value, when it ran, and how long it took, so history builds up as sparklines and deltas. The comparison that matters — this run against the last one — is always one glance away.

Anomaly detection with breakdowns

Automated checks flag outliers. The part that makes them useful is the breakdown by dimension: when a metric drops, the product tells you which slice moved rather than only that something moved.

A drop decomposed as:

• EMEA −14.0% — flagged technical, traced to a stale pipeline sync and fixed
• NA +2.1% — flagged genuine, a real demand shift worth monitoring

is a five-minute triage. The same drop without the breakdown is a day of investigation.

Drift triage in Chat

The classification matters because the two cases need opposite responses. A technical fault needs an engineer; a genuine shift needs a decision. Scoped Chat can pick up the investigation and return an answer with lineage and evidence attached, rather than leaving someone to reconstruct it from scratch.

Anomaly profiles

Sensitivity is configurable. An anomaly profile holds the rules deciding what counts as an outlier for a given metric — window size, deviation thresholds, and how sensitive it should be.

Note

A new tenant has no anomaly profiles of its own. Until you create one, anomaly detection falls back to a built-in default profile, so metrics still get checked — just with generic thresholds. See anomaly profiles and what a new tenant starts with.

Certification is the precondition

Only metrics that are certified should be watched. An uncertified metric is still a draft calculation; alerting on it is alerting on something nobody has signed off. Certification is what turns a metric into a watched number.

What you get

Instead of discovering drift when someone asks, you get it flagged while it is small, with enough breakdown to know whether to escalate it or dismiss it. That is the difference between drift eating trust and drift being handled.