# Prompt 4 — Detect Anomalies in a Time Series

> Adapted from Chapter 10 (Asha worked example, support-ticket volume line). The pattern for the Anomaly Detection AI visual.
>
> **Verified against**: Power BI Copilot, June 2026 surface (Anomaly Detection on a line chart)

## Crystallize

Enable Anomaly Detection on the relevant line chart with the sensitivity slider at default. The Crystallize work is in **the choice of which series to monitor** and **whether default sensitivity matches the metric's underlying volatility**.

The discipline: Anomaly Detection's default sensitivity is calibrated for a generic time series. Real business metrics have their own volatility signatures. Default sensitivity will flag too many points for high-variance metrics and too few for low-variance ones.

## Scope

No conversation in this surface. Treat the sensitivity-slider position as the contract; document it alongside the visual ("anomaly threshold: default; revisit per Ch 10 if false-positive rate > 10%").

## Assemble

The visual flags one or more points as anomalies, with a confidence-interval band drawn around the expected range.

## Refine (per-visual move)

- **Ch 10 business-context check** (the per-visual move for Anomaly Detection): For each flagged anomaly, ask whether the spike or dip has a known operational explanation in the business context. In the Asha case, the flagged morning was the publicized rollout of a customer-facing identity integration — the spike was expected, not anomalous. The callout was a false positive.

A false positive is not a failure of the visual; it is a failure of context that the visual surfaces. Documenting recurring false positives is how you tune the sensitivity over time.

## Ship decision

Do not surface a false-positive callout. Note in the working file (or on the [errata page](/errata/) if the visual is published) that this anomaly type recurs and the sensitivity slider should be adjusted for this metric on the next iteration.

---

**When to use this prompt**: You have a time series with potential anomalies you want surfaced without prior bias about where to look. The business-context check is what distinguishes the discipline from treating every visual-flagged spike as a real signal.

**Dataset**: [asha-escalation-drivers](/datasets/asha-escalation-drivers/) — download CSV + JSON to reproduce the Anomaly Detection pass end-to-end. Shared with the other three Chapter 10 AI-visuals prompts.