Spot trends and anomalies with the Analyst agent

Step 1: Give the Analyst agent the full datasetExcel

Point it at the complete Table, not a pre-filtered view or a summary you built yourself — part of its actual value is deciding for itself which breakdowns are worth checking, and a pre-filtered dataset has already made that decision for it, possibly filtering out the exact anomaly you're trying to find.

Step 2: Ask for chain-of-thought, not just conclusionsCopilot

The Analyst agent can show its reasoning steps if you ask for them, which is what makes its output checkable instead of a black box.

Prompt idea:

Analyze this dataset for usage trends and anomalies over the last 12 months. Show your reasoning step by step — what you checked, what looked unusual, and why — not just a final summary.

Step 3: Follow the reasoning, not just the conclusionCopilot

Read through the steps it shows before accepting the anomalies it flags — if a step doesn't hold up, the conclusion built on it doesn't either.

Step 4: Pull one anomaly and verify it manuallyExcel

Pick the anomaly you're most likely to act on and check it against the raw data yourself before treating it as fact.

Check the work

• Read the agent's stated reasoning steps, not just its summary — if a step is wrong, don't trust what follows from it.
• Manually verify at least one flagged anomaly against the source data.
• Check whether the "trend" it found holds up over a longer time window, not just the period it was shown.

Source: Northwestern University IT, "Using Researcher and Analyst Agents in M365 Copilot" (2026)