AI for Data & Analysis
Learning Outcomes
- Prepare a dataset so AI can read it accurately
- Interpret patterns, totals, and outliers in raw numbers through conversation
- Generate working spreadsheet formulas from a plain-English description
- Summarise a table of figures into an executive-ready narrative
- Validate AI-generated analysis before you rely on it
Lesson Plan
| Segment | Duration | Topic |
|---|---|---|
| Intro | 3 min | Why data work no longer needs a spreadsheet wizard |
| Prepare | 6 min | Getting data into an AI-readable shape |
| Interpret | 8 min | Asking AI what the numbers say |
| Formulas | 8 min | Generating formulas from plain English |
| Trends & outliers | 8 min | Spotting movement, flags, and anomalies |
| Summarise | 6 min | Turning a table into an executive narrative |
| Validate | 4 min | Checking AI's maths before you trust it |
| Wrap-up | 2 min | Key takeaways, preview next lesson |
Before You Begin
Pre-work:
- Complete Lesson 3: Your First AI Conversation so you're comfortable structuring a request
- Review the confidentiality habits in Lesson 4: AI for Document Work — they apply here too
- Have one small, real spreadsheet handy (a sales export, a budget, a survey result)
Shopping List:
- Any AI assistant open in your web browser (the free tier is fine)
- One dataset of 10–200 rows you understand well enough to sanity-check
- A spreadsheet app (Excel, Google Sheets, or similar) to paste formulas into
AI reads text, not spreadsheet cells. To hand over a small table, copy the rows and paste them into the chat — most assistants read tab- or comma-separated data instantly, and many let you upload the file directly.
Before you paste, tidy the data: one clear header row, one record per row, no blank rows or merged cells, no colour-coding that carries meaning (AI can't see colour), and a quick note on any cryptic column name. Then frame the paste with context:
> Here is monthly sales data for our four regions. Each row is one
> month. "Units" is items sold, "Revenue" is in GBP. Read it and
> tell me back in one sentence what this table contains.
>
> Month Region Units Revenue
> Jan North 420 18900
> Jan South 380 17100
> ...
That last instruction — tell me back what you see — is cheap insurance against a misread column.
Revenue is in GBP or Date is DD/MM/YYYY stops the AI misreading 03/04 as the wrong month or treating text as a number.Once the AI confirms it has read your table correctly, start broad before you drill in:
> Based on the table above:
> 1. Which region had the highest total revenue this year?
> 2. Which month was strongest overall, across all regions?
> 3. What's the average revenue per region per month?
> Show the key numbers you used so I can follow your reasoning.
Asking the AI to show the numbers it used is the most useful habit in this lesson — it turns a black-box answer into something you can check.
You can layer follow-ups without re-pasting — the AI remembers the table within the conversation. Mix question styles: orienting ("summarise this in plain English"), pointed ("which row is highest?"), comparative ("how does Q1 compare to Q2?"), and diagnostic ("why might South be lagging?").
This is where hours come back to your week. Instead of remembering spreadsheet syntax, describe what you want in words and let the AI write the formula — then paste it in and adjust the references.
> In Google Sheets, I have months in column A, region in column B,
> and revenue in column D. Write a formula that totals revenue for
> the "North" region only. Explain what each part does.
A good assistant returns the formula plus a short explanation — a SUMIF that adds column D where column B equals "North". Read the explanation; it's how you learn to trust or correct the result. The same pattern handles counting rows, looking up a value, or month-on-month percentage change — just describe the goal.
If a formula errors when you paste it, hand the error back:
> That formula gives a #REF! error in Excel. My data runs from row 2
> to row 47. Fix it and tell me what was wrong.
in Excel or in Google Sheets — they differ in small ways — to get a formula that works first time. Then check it against three rows by hand before applying it to thousands: 'almost right' across 5,000 rows is worse than no formula at all.AI is good at noticing what stands out — the spike, the dip, the row that doesn't fit. But you must define what "unusual" means, or it'll guess.
> Look at the monthly revenue for each region. Point out:
> - Any month more than 20% above or below that region's average
> - The overall trend for each region: rising, flat, or falling
> - Anything that looks like a possible data-entry error
> List each finding with the region, the month, and the figure.
The prompt gives a concrete rule (20% from average) instead of the vague word "outlier." For trends, ask the AI to characterise direction and pace — "describe the trend in two sentences: is growth speeding up, slowing, or steady?" — not predict the future.
describe what already happened is reliable; asking it to forecast next quarter is far riskier — it has no knowledge of your pipeline or market. And a flagged 'possible error' means 'worth a look,' not 'this is wrong': a record-breaking month and a typo look identical to AI. You decide which it is.A table of numbers rarely lands in a meeting; a short, plain-English story does. Ask the AI to translate it for a specific reader.
> Summarise this sales data for our leadership team in four bullet
> points an executive can act on. Lead with the headline number,
> name the strongest and weakest regions, and end with one
> recommendation. No jargon. Round figures to the nearest thousand.
The same data can be retold for different readers — "headline number first, no jargon" for leadership, "exact figures, variance vs budget" for finance, or "neutral, one paragraph" for a board pack.
rewrite in a more direct, confident tone, then edit the wording yourself.People skip this step and later regret it. AI does arithmetic, but it can also state a wrong total with full confidence. Catch it with a three-check routine:
- Spot-check one number in the spreadsheet. If it checks out, trust the rest.
- Sanity-check the scale. "£4.2 million" when you sell a few hundred units a month is a red flag.
- Re-derive it — make the AI show its working a second way:
> You said total revenue was 312,400. Recalculate it by adding the
> four regional totals separately, then show those four numbers and
> the new sum. Do they match your earlier figure?
If the two numbers disagree, you've found a problem before it cost anything. Common risks and their quick checks:
| Risk | How to catch it |
|---|---|
| Silent arithmetic error | Spot-check one figure by hand |
| Misread column | Ask "which column did you treat as revenue?" |
| Made-up data point | Confirm every figure traces back to your table |
| Overconfident "why" | Treat all causes as hypotheses to verify |
Questions & Answers
Key Takeaways
- Prep beats power. A clean table with clear headers and stated units beats the cleverest prompt on messy data.
- Make the AI show its working. Revealing the numbers behind every answer turns a guess you can't check into one you can.
- Describe what you want, not how. Plain-English requests produce working formulas — AI handles syntax, you handle intent.
- Define "unusual" yourself. Concrete rules ("more than 20% from average") give reproducible findings; vague words give vague answers.
- Translate numbers into a story. Data retold for the right audience with a clear "so what" is what moves a meeting.
- Validate before you trust. Spot-check, sanity-check, re-derive. In your report, the number is yours.
Next Steps: Lesson 6: AI for Project Management