Level 2 · Copilot in Excel

2.2Insights & automatic analysis

Goal: Ask Copilot to analyse data: trends, outliers, correlations.

Estimated time: 12 min · self-paced reading · If you only have 5 min: read “The essentials in 30s” + the diagram.

The essentials in 30 seconds

Three doors into analysis: automatic insights to rough out, natural question to target, anomaly detection to make reliable — in that order.
Correlation is not causation: Copilot flags, you interpret — two parallel curves prove nothing.
Ask back for the “how”: column and range used, before reusing a figure in a decision.

Running mini-case — Hicham explores a year of sales

Hicham's “Ventes2025” Table covers 12 months and 8 regions in Algiers and beyond. Instead of building each analysis by hand, he first looks at the automatic insights, then asks: “Which month fell the most?”, “Which region is the most stable?”, “Spot the outliers”. Copilot offers a PivotTable and a chart to insert. Before reusing them in a meeting, Hicham asks back “which column are you relying on?” and traces a suspicious peak back to its source row. In 2.3, he will turn an observation into a formula; in 2.4, he will turn it into a truthful chart.

Continuation of the 2.1 running thread: Hicham's “Ventes2025” Table speaks — and continues in 2.3, 2.4 and 2.5.

1. The analyst who never tires

On a Table, Copilot offers automatic insights: trends, highlights, suggested PivotTables and charts, in one click. That is data exploration without building each analysis yourself. But the real power comes from the natural-language question: “Which month fell the most?”, “Is there a correlation between discount and volume?”. Hicham thus questions his data like a fellow analyst — without writing a formula.

2. The anomaly detector

A very powerful use: anomaly detection. “Spot the outliers”, “Are there any duplicates or inconsistencies?”. Copilot performs a first quality pass that a human eye would take hours to do on a large Table. Beware, though: correlation is not causation. Copilot may flag that two columns move together; it is up to Hicham to interpret whether one explains the other or whether it is a coincidence.

3. Insert and verify

The proposed analyses insert into the sheet (PivotTable, chart) in one click — they become real Excel objects, not mere pictures. Hicham then works from there to refine by hand. Rigour reflex: ask back for the “how”. “Which column are you relying on?”, “Which range did you take?”. Understanding the calculation avoids turning a wrong analysis into a decision — especially on figures that commit.

Common misinterpretation: “Copilot found a correlation, so I found the cause”. No. Flagging that two columns move together explains nothing: common cause, coincidence or real link, it is up to you to decide by going back to the source rows. Hicham's rule: an insight is a lead to confirm (column, range, source row), never a verdict to present as is.

From the automatic click to the precise question

Text version of the diagram (SEO / screen readers / translation): 1) Hicham ROUGHS OUT with Copilot's automatic insights on “Ventes2025” (trends, suggested PivotTables and charts in one click). 2) He TARGETS with natural-language questions (“which month fell?”, “which region is stable?”, “discount/volume correlation?”). 3) He MAKES RELIABLE through anomaly detection (outliers, duplicates) then VERIFIES: “on which column?”, tracing back to the source row, inserting the PivotTable as a real Excel object. Lesson: Copilot flags, the data decides, you interpret.

Did you know?

Hicham nearly presented a “historic record” that was actually an invoice entered twice — Copilot had flagged it as an outlier.

True: a detected anomaly is a lead to confirm at the source before any slide.

A peak is first a question, not a chart title.

🔎 Concrete example — 🔎 Concrete example — Questioning a year of sales without a single formula

Hicham wants to understand his 12 months and 8 regions, not to calculate:

“Which month saw the sharpest fall, and by how much?”
“Which region is the most stable over the year?”
“Spot the outliers and explain why.”

Reading: quantified answers with, often, a chart or PivotTable ready to insert.

Lesson: ask “on which column?” before reusing a figure in a decision.

Section vocabulary

Automatic insights

Trends, highlights, PivotTables and charts suggested by Copilot in one click on a Table: exploration without building each analysis.

Simple example: Hicham scans “Ventes2025” in one click before questioning.

What it is not: definitive truths to present as is.

Reflex: rough out = automatic insights first.

Natural-language question

Questioning your data in English (“which month fell the most?”) without writing a formula: the real power of Copilot analysis.

Simple example: Hicham asks for the most stable region in one sentence.

What it is not: a formula hand-written in each cell.

Reflex: target = ask the business question.

Anomaly detection

“Spot outliers / duplicates / inconsistencies”: an automatic first quality pass over large Tables.

Simple example: Hicham traces a suspicious peak back to its source row.

What it is not: a definitive verdict on the cause of the anomaly.

Reflex: anomaly = lead to confirm.

Correlation ≠ causation

Two columns moving together does not prove that one causes the other: Copilot flags, the human interprets.

Simple example: Hicham checks whether discount and volume are really linked.

What it is not: proof of cause and effect.

Reflex: correlation = question, not conclusion.

In practice — In practice — Run a question-guided data exploration

On Hicham's “Ventes2025” Table (or yours), first look at the proposed automatic insights.
Ask three business questions in natural language (trend, comparison, stability).
Launch an anomaly detection: “outliers, duplicates, inconsistencies”.
For each quantified answer you keep, ask “on which column / which range?” before using it.

An overview of your data in minutes, with the reflex of checking the calculation before deciding.

Ready-to-say script — present an insight without overselling (30 s)

“On Hicham's sales, Copilot flags a peak in May and an erratic North region — these are leads, not conclusions. I checked: Amount column, January–December range, source row of the peak found. The PivotTable is inserted in the sheet, we can rework it. Correlation is not causation: I tell you what the data shows, not why — that, we decide together.”

Say while showing the insight, the cited column and the source row.

Copy-paste template — 4-prompt exploration frame

PROMPT 1 (overview): “Analyse Ventes2025 and give the 3 main findings.”
PROMPT 2 (targeting): “Which month fell the most? Which region is the most stable?”
PROMPT 3 (reliability): “Spot outliers, duplicates and inconsistencies.”
PROMPT 4 (verification): “Which column and which range are you relying on for each figure?”

Month-end close

Hicham speeds up the close: 3 automatic findings, then “which salesperson is progressing most?”.

Quality check

Before any reporting in Algiers, detection of duplicates and outlier amounts, source row verified.

Meeting preparation

“Is there any seasonality by month?”: Hicham arrives with the overview and two settled questions.

Figure that commits

A total reused in a decision: column and range asked back, cross-checked with a formula.

Caution: A Copilot insight is never presented as is when it commits (budget, headcount, contract): check the column, the range and the source row, and cross-check any decisive figure. An anomaly is a lead, not an accusation.

✓ Best practices

Start with automatic insights to rough out, then refine with targeted questions.
Use Copilot as an anomaly detector before any reporting: it spots duplicates and outliers.
Insert the proposed PivotTables/charts: they are real Excel objects, to rework by hand afterwards.

✗ Common pitfalls

Confusing correlation and causation: two parallel curves do not prove a cause-and-effect link.
Reusing a Copilot figure in a decision without knowing on which column it was calculated.
Taking detected anomalies at face value: they are leads to confirm, not verdicts.

Key takeaways

Copilot analyses in natural language: trends, comparisons, correlations, without writing a formula.
Flagship use: anomaly detection (outliers, duplicates) — an automatic first quality pass.
Correlation is not causation: Copilot flags, you interpret.
Ask back for the “how” (column, range) before turning an analysis into a decision.

48h challenge — apply tomorrow

On a real Table (Hicham's or yours): launch the automatic insights, ask 3 business questions, detect 1 anomaly traced back to its source row, and insert the most useful PivotTable.

Section exit checklist