HomeAboutProjectsHow I workContact
04Insights

AnalyzeFrom numbers to an answer.

An insight isn’t a curious fact: it’s something that changes what someone will do. I look for what happened, why, what’s next and what to do.

DAXPython
01The path

Four questions,from what happened to what to do.

An analysis that stops at the first question is a report, not an answer.

Descriptive

What happened

OTIF fell 1.1 points in week 42.

Diagnostic

Why

The north warehouse explains 70% of the drop.

Predictive

What’s next

With no changes, week 44 drops to 91%.

Recommendation

What to do

Check the North stock-out this week.

00Why it matters

A number without its whychanges nothing.

The dashboard says what happened. Someone has to say why, and what to do.

When…
1The number moves
2The meeting starts
3Someone asks why
4It’s time to decide
5Next month comes
6The quarter closes
✗Only the number
−1.1 ppA red arrow
40 minutesGuessing the cause
3 opinionsOne per area
Wait and seeNothing changes
It falls againFor the same cause
What worked?Nobody knows
✓With the analysis
North: 70%Where it comes from
5 minutesOn the cause, not the number
1 answerWith its check
One actionWith an owner and a date
It’s checkedDid the action work?
It’s knownWhat worked and what didn’t
Example figures.
00What I do and don’t

What I don’t do,and what I do.

1The number
✗I don’t

With no comparison

✓I do

Against target, history and peers

2The cause
✗I don’t

Correlation as cause

✓I do

I split signal from noise

3The conclusion
✗I don’t

Everything I found

✓I do

One sentence, before the chart

4The forecast
✗I don’t

One exact number

✓I do

With a range

5Machine learning
✗I don’t

Because it sounds good

✓I do

I start simple

00Signal or noise

Before explaining a change,I check that it is one.

Every metric goes up and down by itself. I explain what leaves its normal variation.

✗Noise93.0%: inside the normal range.

Looking for a cause means inventing one. Next week it comes back by itself.

✓Signal92.4%: outside the normal range.

Something changed in week 42. This one gets analysed.

Where the band comes from: from the metric’s own history. Average ± 2.66 times the average change from one week to the next.

Example figures. The method is Wheeler’s, “Understanding Variation”.
00Where it comes from

First, where it comes from.Then, why.

I split the change before explaining it. The parts have to add up to the total.

OTIF: from the 4-week average to week 42
Average of the last 4 weeks93.5%
North warehousestock-out− 0.77 pp
Center warehousecarrier− 0.22 pp
South warehouse− 0.11 pp
Week 4292.4%

The three parts add up to 1.1 points. The north explains 70%.

What I write

What happenedOTIF fell 1.1 points in week 42.

WhyThe north warehouse explains 70%: three products out of stock.

What’s nextWith no changes, week 44 drops to 91%.

What to doReview the reorder point of those three products this week.

Owner: north warehouse lead · Friday
Example figures.
00Six traps

Six trapsand how I avoid them.

An average

One number for everyone

Split by segment

The total can fall while every segment rises: the mix changed.
Two lines that rise together

“One causes the other”

A group without the change

Moving together isn’t causing. I compare with those who didn’t get the change.
A small base

“It doubled”

With how many cases

Going from 2 to 4 is +100%. I say how many cases are behind it.
Who is in the data

Only those who stayed

Those who left, too

A customer who left doesn’t answer the survey. I look for who is missing.
The worst of the month

“The plan worked”

Against its own history

After an extreme month, the next one is usually closer to normal. With or without a plan.
December against November

One month against the last

The same month last year

Seasons repeat. I compare each period with its equivalent.
00With AI

AI runs the queries.I answer for the number.

The draft is fast. The checks are the work.

✗
Not like thisThe AI says sales fell.

Nobody knows where the number came from.

✓
Like thisSales fell 8%. I checked it against accounting.

The number has a source, and someone answers for it.

  1. 1HypothesesMe

    Three guesses, written before any query.

    I askWhat would have to be true?
    I leave with✓3 hypotheses
  2. 2QueriesAI

    It runs them with read-only access.

    I askShow me the query and the row count.
    I leave with✓A draft with its SQL
  3. 3ChecksMe

    Rows, one total, and the JOIN read line by line.

    I askDoes it match what I already know?
    I leave with✓A number I can defend
  4. 4So what?Me

    I turn the number into a decision.

    I askWhat changes on Monday?
    I leave with✓An action and an owner
  5. 5Error logMe

    I note what the AI got wrong.

    I askWould it fail the same way again?
    I leave with✓One line in the log
02Data or insight

Data informs.An insight makes you act.

The same number, before and after asking “so what?”.

DataInsight
Sales: $412 M.Sales are up, but through one channel: if it stalls, we miss the target.
Gross margin: 31.8%.Margin is up because of mix, not price: keeping it depends on the premium line.
Top 10 products.Seven products sell below true cost: four get adjusted and three go.
00What I get asked

What I get asked,and what kind of question it is.

The kind of question decides the method. Not the tool.

What I get asked→What it is→What I do
“Why did sales fall?”
It’sA diagnosis

✓I split the drop until one part explains most of it.

“Is this normal?”
It’sA variation question

✓I compare it with its own history, not with last week.

“How much will we sell next month?”
It’sA forecast

✓A range, and the simplest method first.

“Did the campaign work?”
It’sA cause question

✓I compare with a group that didn’t get it.

“Which customers are leaving?”
It’sA segmentation

✓Cohorts by start month, before any model.

“Send me all the data.”
It’sA question not asked yet

✓I ask which decision is waiting.

“Can the AI do it?”
It’sA draft

✓It runs the queries. I check them and answer for the number.

00How I know I’m done

There are three written insights.Each with its action and owner.

Analysing ends when someone can act without opening a single chart.

I’m done when…
  1. ✓
    A conclusion in one sentenceThe north warehouse explains 70% of the drop.
  2. ✓
    Compared against target and history92.4% against a 95% target; −1.1 pp in four weeks.
  3. ✓
    Checked against the sourceRows, one total, and the JOIN read line by line.
  4. ✓
    An action and an ownerVisit the north warehouse. Owner: warehouse lead.
The north warehouse explains 70%. Check it first.Action: visit north · Owner: warehouse lead

With the insights written, they get told. Next comes Communicate.