How to Identify Outliers or Unusual Patterns in Your Data with AI
Outliers are where the interesting stuff lives — the early signals, the anomalies, the things you’d want to know about. Copilot finds them fast.
What it is
Most data analysis focuses on averages, trends, and headlines. But the outliers — the data points that don’t fit the pattern — are often where the most interesting stories are. The unusually high score, the sudden drop, the customer who behaved completely differently from everyone else. Copilot can scan datasets and surface these anomalies in seconds.
The challenge is that not every outlier is meaningful. Some are data entry errors. Some are seasonal effects. Some are real signal that something is changing. Copilot can identify the outliers, but you need to bring the context: ‘Is this December spike normal? Yes — happens every year.’ or ‘This region’s drop is concerning. Let’s investigate.’ The AI flags; you interpret.
Use outlier analysis as the start of a conversation, not the end. ‘These five values are outliers’ is a list. ‘These five values are outliers, here’s why each might be worth investigating, and here are the questions to ask the source teams’ is a starting point for action. Copilot can do all of this if you ask explicitly.
When to use this
- When you suspect something unusual is happening in your data.
- When you want to spot-check a dataset for errors or anomalies.
- When you’re investigating a sudden change or unexpected result.
- When preparing a report and want to call out the most notable points.
How to do it
- Define what ‘normal’ looks like for the dataset (range, average, expected pattern).
- Ask Copilot to identify outliers and explain why they stand out.
- Ask for possible reasons for each outlier (errors, seasonality, real change).
- Request follow-up questions or validation checks you should run.
- Verify the outliers against context (what was happening at that time?).
- Decide whether to investigate further, ignore (known cause), or fix (data error).
Best practices
- Compare like with like. Outliers in one segment may not be outliers in another.
- Check for data quality issues first. Often outliers are typos, not real anomalies.
- Look at outliers in context. Some are valid (a one-time event); others are concerning (early signal).
- Document assumptions. What you treated as normal vs anomalous matters for the analysis.
Common mistakes
- Treating every outlier as significant. Many are noise. Apply judgement.
- Ignoring outliers that don’t fit your hypothesis. The unwelcome outlier might be the most important signal.
- Acting on outliers without verification. One bad data point shouldn’t drive a major decision.
The Copilot Readiness Guide gives you the 25-question scorecard, the 4-category risk audit, and the 30-day plan to fix permissions, content quality, and sensitive content before go-live.
Get the Copilot Readiness Guide — $39 →FAQ
How do I find outliers in a dataset with Copilot?
In Excel, open Copilot and ask ‘Find outliers in this dataset and explain why each one is unusual.’ Copilot uses statistical methods (z-score, IQR) to flag values, then explains the context. The Excel Agent can also produce a separate outlier-flag column to mark the rows.
What statistical method does Copilot use for outliers?
Copilot defaults to standard deviation-based detection (typically flagging values more than 2-3 standard deviations from the mean) or IQR-based (1.5× IQR beyond Q1/Q3). For non-normal distributions, ask Copilot to use a different method — percentile-based or domain-specific thresholds work better for skewed data.
Why is Copilot flagging values that don’t seem like outliers?
Two common reasons: the dataset has a skewed distribution where standard deviation-based detection misfires; or Copilot is interpreting ‘outlier’ more loosely than you intended (highlighting ‘interesting’ values, not strictly statistical outliers). Specify the threshold in your prompt: ‘Flag values more than 3 SD from the mean’.
Can Copilot detect anomalies over time in Excel?
Yes — for time-series data, ask Copilot to ‘find unusual periods or breaks in trend’. Copilot can identify sudden spikes, level shifts, or breaks in seasonality. Verify the findings visually with a chart; outlier detection in time series is notoriously fragile and Copilot can misfire on cyclical patterns.