The three thousand row stare
The data is clean. The columns are named. You scroll to the bottom, you scroll back to the top, and you feel exactly as informed as you did ten minutes ago. Somewhere in these rows is the thing your manager wants to know, and you cannot see it, because human eyes are not built to read three thousand rows.
This is the moment people type the worst question in analytics into a chat box: what is interesting here? It feels like the right question. It is the sort of thing a clever consultant would ask. What comes back is four paragraphs of confident mush: sales show variation across regions, some products perform better than others, further analysis is recommended. You have learned nothing, and it took ninety seconds.
In plain English
- Segment:
- A slice of your rows that share something. One region, one product, one month.
- Driver:
- The thing doing most of the moving. If revenue fell, the driver is whatever fell hardest inside it.
- Outlier:
- A row far outside the normal range. Either your most interesting customer or your worst typo.
- Pivot table:
- A built-in tool that groups your rows and totals them. Sales by region by month, in about four clicks.
- Baseline:
- The thing you are comparing against. Last month, last year, the average, or target.
Why vague questions produce vague answers
An AI answering "what is interesting here" has no idea what you consider interesting, what normal looks like in your business, or what decision this is feeding. So it hedges. It describes your data back to you in a slightly more expensive vocabulary.
A good data question has three parts: a metric, a comparison and a reason for caring. Miss the comparison and you get a number with nothing to lean against. Miss the reason and you get an answer nobody can act on.
โ Weak prompt
Prompt
Here is my sales data. What is interesting?
Output
Sales vary by region and month. The North region shows notable performance. Product mix differs across segments. Further analysis may reveal additional insights.
No metric, no comparison, no decision. You get a summary of the obvious, dressed up in consultant language.
โ Good prompt
Prompt
Here is monthly revenue by region for the last 12 months. Total revenue is down 8 percent in Q2 against Q1. I need to tell my manager which regions caused that drop. Rank the regions by how much of the 8 percent each one accounts for, in pounds and as a share of the fall. Ignore anything under 1 percent.
Output
A ranked list showing that two regions account for most of the drop, with the pound amount and share for each, and the rest grouped as immaterial.
Metric, comparison, decision and a threshold for what counts. The answer arrives ready to say out loud.
Checkpoint
Every good data question names a metric, a comparison and the decision it feeds. Without a comparison, a number has nothing to be big or small against.
Question shape one: what is driving this
When something has moved and you need to explain why, ask for a ranked contribution rather than a description. You want to walk into the meeting able to say two regions caused four fifths of it, not sales were mixed.
Here is [what the data is] with columns [list them].
[Metric] changed from [X] to [Y] between [period A] and [period B].
Break that change down by [region / product / channel].
Rank each one by how much of the total change it accounts for,
showing both the amount and the percentage share.
Stop when you have explained 80 percent of the movement.
Say clearly which single item mattered most.
The last two lines do the real work. Without them you get a full list of every category including the fourteen that moved by nothing.
Question shape two: what is unusual
Outliers are where the money and the mistakes both live. A customer ordering forty times the average is either your best account or a broken integration, and you want to know which before Friday.
Here is [what the data is] with columns [list them].
Find rows that are unusual on [Amount / Quantity / Days to pay].
For each one tell me: the value, how far it sits from the typical value,
and your best guess at whether it is real or a data error.
List the top 10, highest first.
Do not remove anything. I want to look at them myself.
Never let a clean-up step delete outliers automatically. The row you drop as an anomaly is quite often the one your manager is about to ask about by name.
Question shape three: what changed
The most common real-world request is a comparison against a baseline. Last month against this month, this year against last, actual against target. Name the baseline explicitly, because if you do not, the AI will pick one for you and not mention it.
Compare [this period] against [that period] for [metric].
Give me a table with: category, the old value, the new value,
the change in absolute terms, and the change as a percentage.
Sort by absolute change, biggest first.
Then in two sentences, tell me the story a busy person needs to hear.
Flag any category where the percentage looks dramatic
only because the starting number was tiny.
That final instruction saves you from the classic embarrassment: proudly reporting that the South West grew 400 percent, when it grew from one order to five.
Give it structure, not a data dump
Two practical notes. First, most chat tools cope far better with a summarised table than with 40,000 pasted rows, and a pivot table takes about four clicks to build. Summarise first, then ask your question about the summary.
Second, keep personal details out of the box. Customer names, email addresses, account numbers and anything covered by an agreement should be swapped for placeholders or removed before you paste. Your question is nearly always about the pattern, not about Mrs Henderson in particular.
End any analysis prompt with this line: tell me what you would check next, and what in this data could make your answer wrong. It costs nothing and it is often the most useful paragraph you get back.
๐ Quiz
Question 1 of 4Why does asking what is interesting here produce such weak answers?