CODED × KJO · Participant workbook
Day 2. From Data to Insight
What you leave with today
1A check you run before you ask anything: which columns, and how many rows have them.
2One number with something beside it, and the row count behind both.
3Your Comparison card: what it means, and what would have to be true for it to be wrong.
Words used today
Grounding: pointing a tool at a file, instead of pasting the contents into the chat.
Track A: Copilot inside Excel.
Track B: Copilot Chat with the file attached, plus manual Excel steps.
Outlier: a value far outside the normal range of a column.
Signal: an outlier that means something about the business.
Information: a number with something beside it.
Knowledge: what that comparison means, and what would change if it were true.
Clean: what is in the file is right. Complete: everything you need is there. They are not the same thing.
Page 1
Block 1: Two hundred and fifty-seven
Yesterday's three artifacts:
Industrial Branch handled 257 service requests this year.
Is that good or bad?
What would you need to know before you could answer?
A number on its own is not information. It is just a number.
The ladder
| Rung | The question it answers | Where |
|---|---|---|
| Data | What have we actually got? | Yesterday |
| Information | What does it add up to? | This morning |
| Knowledge | What does it mean, compared with what? | This afternoon |
| Wisdom | What should we show, and to whom? | Tomorrow |
Exercise 1.1.Yesterday's question, on the full file.
Which site had the worst unplanned downtime last quarter? Use
work-orders-full.xlsx.What happened when you tried?
Did the tool tell you how many rows it used?
Page 2
Block 2: Grounding
The problem
Every file at your desk is this size or bigger. Yesterday's method was practice.
The issue
A model holds a limited amount of text in view at once. That limit is the context window. Paste more than fits and the earlier part falls out of view. The tool does not stop. It answers using what it can still see.
The concept
Grounding: you point the tool at the file instead of pasting the file into the tool.
Three things change. It sees every row. It knows which column holds a date and which holds text. It can count.
Yesterday you learned that a model predicts rather than computes. A grounded tool runs real operations on a real table. That is why the dashboard you saw yesterday morning had numbers you could trust.
The fix: two routes
| Track A: Copilot in Excel | Track B: Copilot Chat with file | |
|---|---|---|
| Where you work | Inside the workbook | In the browser |
| Writes back into your file | Yes | No, you copy across |
| Best for | Adding columns, tables, dashboards | Explaining a decision, writing your three sentences |
| Needs | A Copilot license for Excel | The file attached to the chat |
Which track do you have today?
Worked grounding prompt
You are a maintenance data analyst. I have attached a work order file. Each row is one work order. Before answering, state: - the number of rows you can see - the date range and which column holds it Then answer this question: [your question here] With your answer, state the number of rows you used and the number you excluded. For every excluded row, say why.
Grounding does not replace the verification ask. It makes the verification ask answerable.
In depth: what grounding does not fix
It still does not know what your data means. It will confidently analyze a column it has misread.
Track A needs the data as a table and works best on one sheet.
Track B cannot write into your workbook, so you copy results across by hand. That is where transcription errors enter.
A grounded tool is right far more often. Your trust in it will rise. Page 3 is about what that costs you.
Page 3
Block 5: Can this file answer that question?
Before you ask the tool anything, you ask yourself two things.
1. Which columns does this question need? Name every one of them.
2. How many rows have all of them filled in? Not the file's row count. The count for these columns.
My confidence in this tool right now, out of 10:
Exercise 5.1.The columns in
work-orders-full.xlsx. Fill this in from the check card.| Column | Rows filled | Column | Rows filled |
|---|---|---|---|
WorkOrderID | CloseDate | ||
Site | DowntimeHours | ||
Asset | Cost_KWD | ||
Priority | Contractor | ||
RaisedDate | Total rows |
Which column is not like the others?
Exercise 5.2.Your own questions. Run the check before you open anything.
| The question you want answered | Columns it needs | Rows with all of them | GO / STOP |
|---|---|---|---|
Now answer the ones that cleared. Row count on every answer.
| The question | Your answer | Rows the tool used |
|---|---|---|
My confidence after those, out of 10:
Page 3b
Block 6: The fifth question
The manager's fifth question: average time to close a work order, by site.
Your answer:
Rows the tool used:Rows in the file:The difference:
Did you run the check on this one before you asked? yes / no
Which columns does it need?
How many rows have all of them?
Where had you already seen this problem?
Which site ranked fastest, and where does it actually rank?
The concept
Confidence is not evidence. Ask for the row count on every question, not on the hard ones, because you cannot tell in advance which one is the hard one.
Nothing went wrong with the tool. Every number in the first table was arithmetically correct. It answered the question it was asked, using the rows it could use, and said nothing about the rest.
This was not a hallucination. It was a silence.
Page 4
Block 6: The fix, and why it works
Exercise 6.1.Rewrite the fifth question so a wrong answer becomes visible.
The version that works
What is the average time to close a work order, by site? Time to close means the days between RaisedDate and CloseDate. Before you answer: - report how many work orders have no CloseDate - report that count by site, as a percentage of each site's rows Then give me two tables: Table A: the average using only work orders that have a CloseDate. Table B: the average counting work orders with no CloseDate as still open, measured from RaisedDate to the latest date in the file. Tell me which sites change rank between the two tables.
Same tool. Same file. Same question. Three extra lines.
You did this in your workbook in ninety seconds in Block 5. These lines do it inside the prompt, every time, whether or not you remember.
Two sentences to keep
Clean is not the same as complete.
Clean means what is in it is right. It does not mean everything you need is there.
Run the check on every question, not on the hard ones.
You cannot tell in advance which one is the hard one.
Page 5
Blocks 7 and 9: Compared with what?
A number becomes information when you put something beside it. There are four things you can put there.
| # | Compare against | The question it answers | Watch out for |
|---|---|---|---|
| 1 | The same thing last period | Is this getting better or worse? | Both windows need the same chance to age |
| 2 | The target | Are we where we said we would be? | If nobody set one, say so. Do not invent one |
| 3 | Another group | Who is doing this better? | The groups have to be alike enough |
| 4 | A benchmark outside | Is this normal for work like ours? | Where did it come from, and who paid for it |
A fifth option is always available: compare against nothing, and say so.
Exercise 9.1.Take one number from the Work Order Set. Produce four versions of it, one per comparison. All four will be true.
My number:Rows it used:
| # | Compared against | The finding that comes out |
|---|---|---|
| 1 | Last period | |
| 2 | The target | |
| 3 | Another group | |
| 4 | A benchmark |
Which of those four would change a decision?
Choosing the comparison is choosing what matters. The tool will not do that for you.
Page 6
Block 11: Outlier or signal
Your answer from Day 1, page 8. Did you delete the 40x cost row?
What it actually was:
WO-2087. Ahmadi Refinery. 72 hours. 390,000 KWD. A three-day compressor failure, and the most expensive single event in fourteen months of data.It is also one of the twenty duplicate IDs. Removing duplicates by ID could have deleted it without anybody choosing to.
Three questions for any outlier
1Is it a data error, or a real event?
2If it is real, is it a one-off or the first of a pattern?
3Who in the business would know?
An outlier is a value far outside the normal range. A signal is one that means something.
The data cannot tell you which one you are holding. The business can.
The cleaner chart is not the more correct chart. It is the more comfortable one.
Page 7
Block 12: What it means
A true comparison is not yet a finding. Meaning is a different question: what changes if this is true?
The three tests
| Test | The question |
|---|---|
| So what | What does this cost or save, and roughly how much? |
| Who | Who does something different because of it? |
| What would break it | What would have to be true for this to be wrong? |
Exercise 12.1.Finish your Comparison card, then copy it here.
The number:
Compared against:
The comparison, in one line:
Rows it used:
What it means. Not what it is. What changes if it is true.
What would have to be true for this to be wrong:
The test before you hand it over: read it out loud and ask who does something different tomorrow because of it? If the answer is nobody, you have a number and a neighbor, not a finding.
Page 8
Day 2 key takeaways
Write these in your own words.
1
2
3
What today built
The four things worth carrying to your desk
Ground the file. Do not paste it. Then ask for the row count every time.
Before you ask: which columns does this need, and how many rows have them all? Clean is not complete.
Keep the strange row until somebody explains it. The data cannot tell you what it means.
A number becomes information when you put something beside it. Choosing what to put there is choosing what matters.
Before tomorrow
Tomorrow: you get a dashboard where every number is correct and nobody in the room can use it. Then you build one that a stranger can read without you standing next to them.