Descriptive Analysis in Excel: What the Table Means

Most guides to descriptive analysis in Excel stop at the button. Open the Analysis ToolPak, run Descriptive Statistics on a column, and Excel hands back a table: mean, median, mode, standard deviation, skewness, kurtosis, range, count. How to Add Data Analysis in Excel covers turning the ToolPak on if it is not already there. What that tutorial does not usually cover is what those seven numbers actually tell you about your data, and that is where most people get stuck — not at the button, at the table it produces.

What each number in the table means

• —Mean — the average. Sensitive to a single extreme value, which is the whole reason the next number exists.
• —Median — the middle value once everything is sorted. Barely moves when one row is an outlier, which is why a mean and median that disagree by a lot is itself the finding.
• —Mode — the most frequent value. On continuous data (prices, durations, measurements) this often comes back as an error, which just means no value repeated exactly — not a sign anything is wrong with the data.
• —Standard deviation — how spread out the values are around the mean. A small number means most rows sit close to it; a large one means they don't.
• —Skewness — which side the data leans. A positive number means a long tail of high values pulling the mean above the median; negative means the opposite.
• —Kurtosis — how much of the spread comes from rare, extreme values rather than the ordinary middle of the distribution. High kurtosis is a warning that a handful of rows carry outsized weight.
• —Range, minimum and maximum — the floor and ceiling, and the gap between them. Useful for spotting an impossible value (a negative age, a nine-digit price) before it poisons every other number on the list.

A worked example

Say you ran a short onboarding survey, 20 new hires rating "how useful was week one" from 1 to 10. The Descriptive Statistics output shows a mean of 6.8 and a median of 8. Those two numbers disagreeing is the actual story: a handful of very low scores are pulling the mean down while most people answered near the top. Reporting only the mean — "onboarding scored 6.8 out of 10" — understates how most new hires actually felt, and reporting only the median hides that a real minority had a bad week one worth investigating.

Weak prompt: "What does this data show?" — with a screenshot of the table attached.

Better prompt: "Descriptive Statistics output for 20 onboarding survey scores: mean 6.8, median 8, mode 9, standard deviation 2.4, skewness -0.9. Explain in two sentences what the gap between the mean and median suggests about the shape of these scores, for someone who has not seen the raw data. Do not recalculate anything — treat these numbers as fixed." Handing over the exact figures, rather than a picture of the table, is the same structural fix OpenAI's own prompting guidance recommends for any task: give the model the facts it needs rather than asking it to extract them itself.

Where AI genuinely helps, and where it overreaches

Turning seven numbers into a sentence a non-statistician can act on is a task AI does well, provided the arithmetic stays in Excel's output and the model is only asked to describe it. OpenAI's own documentation on data analysis with ChatGPT and the ChatGPT spreadsheets tool are built around exactly this — reading a table you already have, not inventing one.

Where it overreaches is explaining why the distribution looks the way it does. Negative skew in the onboarding example is consistent with a few people having a genuinely bad week, but it is equally consistent with a scale problem, a confusing survey question, or three people who misread the 1–10 direction. A model given only the summary table has no way to tell those apart, and a fluent, confident guess at the cause is not the same as a checked one — a survey of hallucination in large language models documents this exact pattern: the model's confidence carries no information about whether the underlying claim is actually true.

If you want the model to look at the raw rows rather than the summary table, uploading the actual file is the more reliable path — OpenAI's file uploads FAQ sets the real limits on what it can open directly, which matters before you assume it has seen every row rather than a description of them.

One tool answers a different question

Descriptive Statistics describes one column on its own. It does not tell you whether a difference between groups — say, onboarding scores for people managers versus individual contributors — is a real effect or noise; that is what ANOVA in Excel tests for, and it is a genuinely different question from anything in the table above. It is also a different feature from the small lightning-bolt icon Excel shows when you select a range — Quick Analysis tools in Excel covers that one, which applies formatting and quick totals and does not run a statistical test at all. People searching for descriptive analysis sometimes land on the wrong one of the two.

Checks before you repeat a summary to anyone

1. 01Look at the mean and median side by side before reading anything else. If they are close, the data is roughly symmetric and either number is a fair headline; if they are far apart, say both, not just the one that sounds better.
2. 02Treat an AI explanation of why a distribution looks a certain way as a hypothesis, not a finding — it was not shown the raw rows, only the summary.
3. 03Re-check the minimum and maximum against what is actually plausible for the thing being measured, the same way checking an AI answer when you are not the expert recommends for any claim you cannot independently verify from memory.
4. 04Keep the source data in an Excel Table rather than a plain range, so the ToolPak's input stays correct as new rows get added instead of silently leaving them out next time you re-run it.

What to do Monday

1. 01Run Descriptive Statistics on one column you already understand well, so you can sanity-check the output format before trusting it on something new.
2. 02Compare the mean and median first, every time, before reading skewness or kurtosis — it is the fastest way to notice a distribution that is not what a single average would suggest.
3. 03Once the numbers are right, AI data visualization tools covers picking a chart that actually shows the shape of a distribution rather than the default Excel suggests.
4. 04If the real question is whether two groups genuinely differ, not just what one column looks like on its own, move to ANOVA rather than trying to read that answer out of Descriptive Statistics.

The judgement that separates a useful summary from a misleading one is the same skill the labour market is already pricing in: PwC's 2026 Global AI Jobs Barometer found a 62% wage premium for workers with measurable AI skills, up from 57% the year before. Coursium's AI Excel course teaches that judgement directly — reading what a result actually supports, not just what a summary claims. Stay ahead of AI by learning the tools on your phone.