AI Assisted Data Analysis

Welcome

Why this matters

AI tools are fast, fluent and confident; even when they are wrong. They will happily run a t-test on the wrong variable, call a correlation a cause, or produce a beautiful chart from a mistaken assumption, and say all of it in the same authoritative tone as when they’re right. The skill this session builds isn’t prompting. It’s knowing enough about your own data and question to catch it when that happens.

Learning objective

Use AI to support your analysis while remaining in control of the analytical process.

By the end of this session, you should be able to use AI to:

• Explore a dataset and understand its structure, variables and contents.
• Investigate an interesting relationship and generate potential insights.
• Decide what you want AI to do rather than asking AI to decide the whole analysis.
• Evaluate and validate AI generated analysis.
• Explain and justify your conclusions in your own words.

The key principle

AI can help you analyse data. It should not replace your analytical judgement.

A useful mental model is:

You → decide → AI assists → you check → you interpret → you conclude

What’s ahead

1

The principle

2

A simple framework

3

Your turn: the dataset challenge

4

Reflect & take home

This is deliberately hands-on. You do not need to become an expert prompt engineer — the aim is to practise making analytical decisions and then using AI to help execute, explain or check them.

Effective AI-assisted data analysis

Start with the question, not the AI

A weak starting point is:

“Analyse this dataset and tell me what you find.”

This gives AI too much control over the analytical process.

A stronger starting point is:

“I want to understand whether study time is associated with assessment score. First help me inspect the variables and identify appropriate ways to investigate this relationship. Do not choose an analysis for me without explaining the reasoning.”

The difference is important:

• Weak: AI decides the question, method and interpretation.
• Strong: You define the purpose; AI supports the process.

A simple framework:

1. PLAN

Before opening Claude or Copilot, decide:

• What am I trying to understand?
• Which variables might help answer the question?
• What kind of relationship or pattern am I interested in?
• What would count as useful evidence?

2. ASK

Give AI enough context to help.

A useful prompt can include:

Context + Goal + Data + Task + Constraints

Example:

I am exploring a small student dataset.

My question is whether study hours are associated with
assessment score.

The variables include study_hours and assessment_score.

Please:
1. describe the relevant variables,
2. suggest suitable ways to explore the relationship,
3. explain why each approach would be appropriate,
4. do not make the final analytical decision for me.

Flag anything I should check before interpreting the results.

3. CHECK

AI output is a suggestion, not evidence.

Check:

• Did AI understand the variables correctly?
• Are the data types what it thinks they are?
• Did it use an appropriate method?
• Are there missing values or unusual observations?
• Are the calculations correct?
• Does the graph actually support the claim?
• Is the conclusion stronger than the data justify?

4. EXPLAIN

If you cannot explain the result yourself, you do not yet own the analysis.

Try completing this sentence without looking at the AI response:

“The data suggest that ______ because ______. However, ______.”

That final qualification is often where good analytical judgement appears.

TO DO AND NOT TO DO

If the AI does this…	You do this
Jumps straight to a statistical test without asking what you want to know	Stop it and restate your question first
Describes a variable confidently but incorrectly	Check the variable against the raw data yourself
Reports a correlation as if it were a cause	Ask it to name at least one alternative explanation
Gives a p-value with no explanation of what test was run or why	Ask “what test did you use, and why is it appropriate here?”
Produces a polished graph you haven’t actually looked at	Open the plot and read the axes before trusting the story
Makes a claim you can’t reproduce or verify	Ask for the code or working, then run it yourself

Your dataset challenge

The scenario

Imagine you are exploring data from students completing an introductory course.

Your broad question is:

What factors appear to be related to assessment performance?

You are given the dataset below.

Your task

Work independently for the next 8–10 minutes.

Use Claude or Copilot to support your analysis.

Do not ask AI to “analyse everything”.

Instead, make your own choice about one relationship that you find interesting.

What a strong answer looks like: you can point to a specific relationship, show a chart or number that supports it, and say (in your own words) one thing that would make you doubt the conclusion.

Step 1 — Explore

Ask AI to help you understand:

• how many observations there are;
• what each variable represents;
• variable types;
• possible missing or unusual values;
• which variables might be useful for your question.

Your decision: What do you want to investigate?

Step 2 — Investigate

Choose one relationship:

• study hours ↔︎ assessment score
• attendance ↔︎ assessment score
• prior score ↔︎ assessment score

Now run it through ASK from the framework above — swap in your chosen variables and question. If you want a ready-made version, adapt this:

I have chosen to investigate the relationship between
attendance and assessment score.

Help me explore this relationship. Please:
- suggest an appropriate visualisation,
- calculate a useful summary statistic,
- explain what the result means,
- state important limitations,
- distinguish what the data show from what is only
  a possible explanation.

Show your calculations/code so I can check them.

Whatever it gives you, hold onto it loosely — Step 3 is where you test it.

Step 3 — Challenge the AI

Now deliberately ask a checking question:

Before I accept this conclusion, act as a critical reviewer.

What could be wrong with this analysis?
What assumptions or alternative explanations should I consider?
What should I independently verify?

Then actually verify at least one thing.

For example:

• inspect the raw values;
• recreate the calculation;
• inspect the scatterplot;
• check whether an apparent pattern is driven by one or two observations;
• compare the AI’s interpretation with the descriptive statistics.

Dataset

⬇️ Download the dataset (CSV)

Student dataset

student_id	study_hours	attendance	prior_score	assessment_score
1	3	62	52	55
2	5	74	61	62
3	2	55	48	49
4	8	91	70	78
5	6	83	66	69
6	4	70	58	61
7	10	96	74	84
8	7	88	68	74
9	1	48	43	45
10	5	76	63	65
11	9	94	71	81
12	6	81	65	71
13	3	64	54	57
14	11	97	78	88
15	4	69	60	63
16	8	89	72	79
17	2	57	50	51
18	7	85	67	73
19	12	98	82	91
20	5	72	59	64
21	6	79	64	68
22	9	93	75	83
23	3	61	53	56
24	10	95	77	86

Optional: inspect the data yourself

str(students)

## 'data.frame':    24 obs. of  5 variables:
##  $ student_id      : int  1 2 3 4 5 6 7 8 9 10 ...
##  $ study_hours     : num  3 5 2 8 6 4 10 7 1 5 ...
##  $ attendance      : num  62 74 55 91 83 70 96 88 48 76 ...
##  $ prior_score     : num  52 61 48 70 66 58 74 68 43 63 ...
##  $ assessment_score: num  55 62 49 78 69 61 84 74 45 65 ...

summary(students)

##    student_id     study_hours       attendance     prior_score   
##  Min.   : 1.00   Min.   : 1.000   Min.   :48.00   Min.   :43.00  
##  1st Qu.: 6.75   1st Qu.: 3.750   1st Qu.:67.75   1st Qu.:57.00  
##  Median :12.50   Median : 6.000   Median :80.00   Median :64.50  
##  Mean   :12.50   Mean   : 6.083   Mean   :78.21   Mean   :63.75  
##  3rd Qu.:18.25   3rd Qu.: 8.250   3rd Qu.:91.50   3rd Qu.:71.25  
##  Max.   :24.00   Max.   :12.000   Max.   :98.00   Max.   :82.00  
##  assessment_score
##  Min.   :45.00   
##  1st Qu.:60.00   
##  Median :68.50   
##  Mean   :68.88   
##  3rd Qu.:79.50   
##  Max.   :91.00

Quick visual exploration

plot(
  students$attendance,
  students$assessment_score,
  xlab = "Attendance (%)",
  ylab = "Assessment score",
  main = "Attendance and assessment score",
  pch = 19,
  col = "#7c6ff2"
)

You can also investigate another relationship by changing the variables:

plot(students$study_hours, students$assessment_score)
plot(students$prior_score, students$assessment_score)

The critical evaluation checkpoint

Stop and decide

Before reading the model answer, answer these questions yourself:

1. What did you investigate, and why?

One sentence on the relationship, one on what made it worth looking at.

2. What does the evidence show — and not show?

Be specific. “There is a relationship” isn’t specific; a direction, a rough size, and its limits are. Remember: an association is not a cause.

3. What did you check yourself?

Name one thing you independently verified rather than took on trust.

4. Would you have trusted the AI’s first answer without checking it?

No — and that’s the point. The workflow exists to keep the analytical reasoning with you, not to make AI faster to agree with.

Debrief

What good AI-assisted analysis looks like

A strong workflow is not:

Dataset → AI → answer

It is:

Question → your decision → AI assistance → evidence → validation → your interpretation

Three questions to remember

1. Did I decide what I wanted to know?

If not, AI may be driving the analysis.

2. Can I verify what AI did?

If not, ask for the code, calculations, assumptions or explanation.

3. Can I explain the conclusion without AI?

If not, you have not yet taken ownership of the result.

Take-home message

Use AI as an analytical assistant, not as the analyst.

Let it help you:

• explore,
• generate options,
• write code,
• explain methods,
• identify things to check,
• challenge your interpretation.

But you remain responsible for deciding what the analysis means.

A reusable prompt template

Copy and adapt this for future analyses:

CONTEXT
I am analysing [dataset/context].

GOAL
I want to understand [specific question].

DATA
The relevant variables are [variables].
Here is the data / an attached file: [data].

TASK
Please help me:
1. [explore / visualise / calculate / compare]
2. explain the reasoning behind your suggestions
3. show the code/calculations so I can verify them

CONSTRAINTS
Do not decide the overall analytical approach for me.
Do not make causal claims unless the design supports them.
Flag assumptions, limitations, missing information and anything
I should independently verify.

OUTPUT
Give me concise, interpretable results and distinguish:
- what the data show,
- what is uncertain,
- what would require further checking.