AI Data Analysis for Evaluators – Tutorial
EvalCommunity Academy · AI Tools for M&E Practice
How to Use AI for M&E Data Analysis Without Losing Your Analytical Skills
Five methods for using ChatGPT, Claude and similar tools as a mentor, tutor and reviewer, so your analysis gets better and your own judgement gets sharper, not weaker.
Level: Beginner to intermediate · Reading time: about 10 minutes · Updated: November 2026
Why this matters for evaluators
An AI tool can turn a survey export into tables, charts and “key insights” in seconds. But you are the one who signs the report, defends the numbers at the validation workshop and answers the donor’s questions. If you did not do the thinking, you cannot defend the findings. This tutorial shows how to get AI’s speed without giving up the analytical skills your credibility depends on.
Learning objectives
By the end of this tutorial you will be able to:
- Recognise the “illusion of understanding” that AI-generated analysis can create.
- Use AI as a mentor, tutor and reviewer instead of as the analyst.
- Write prompts that give hints and explanations rather than finished answers.
- Verify any AI-produced figure before it reaches a report.
- Build your analytical skills through one growing practice project.
1. The illusion of understanding
Before AI, learning to analyse M&E data was slow. A pivot table that didn’t add up, a disaggregation that contradicted the overall figure, a Stata command that threw an error: each one sent you to the manual, a colleague or a forum. That friction was frustrating, but it is exactly how analytical judgement develops.
Today you can upload a dataset and receive an analysis in seconds. You see the answer, so it feels as if you understand it. That feeling is the trap: recognising a result is not the same as being able to produce, check and defend it.
AI as the analyst
- “Analyse this endline survey for me”
- “What are the key findings?”
- “Why did the indicator decline?”
- Your brain stays passive
AI as mentor and reviewer
- “Set me analysis tasks, no answers”
- “Explain the concept, not the formula”
- “Review my analysis and challenge it”
- You own the analysis
2. Five methods that build skills instead of replacing them
1
Use AI as a task-setter and mentor, not the analyst
Ask AI to design the exercise, then do the analysis yourself in Excel, R, Stata, SPSS or Python. You still inspect the data, choose the questions, write the formulas or code, and interpret the results. AI only steps in when you are genuinely stuck.
Create a practice project using a synthetic household survey dataset for a livelihoods programme (baseline and endline, 400 households, with sex of household head, district and income). Give me a programme question to investigate and a series of analysis tasks that get progressively harder. Don’t give me the answers unless I ask.
When you hit a roadblock, ask for guidance, not the solution:
- Why are my disaggregated totals higher than the overall total?
- What concept am I misunderstanding in this pivot table?
- Explain this Stata error without giving me the corrected command.
- What should I check next, without telling me the answer?
Tip: grow one project instead of jumping between datasets. Start with descriptive statistics, then add disaggregation, then baseline-endline comparison, then a dashboard, then a two-page brief for the programme team. Each step builds on the last, and your skills compound.
2
Ask for concepts and “why”, not finished answers
Treat AI as a tutor. Instead of “write the formula for change from baseline”, ask it to explain the concepts you need, and then write the formula yourself.
Explain, as if I’m new to this, how I should calculate change between baseline and endline for a percentage indicator. Cover percentage points vs percent change and what to watch for if the sample changed. Don’t give me the formula.
Other good “why” questions for M&E data:
- When should I use survey weights, and what goes wrong if I ignore them?
- Why might an average hide what is happening to the poorest households?
- When is a median more honest than a mean for income data?
- My result looks strange. What should I check first: the data, the formula or the indicator definition?
3
Use AI as a reviewer after you finish
Some of the most valuable AI use comes after your analysis, not before. Complete the work yourself, then ask AI to critique your method, assumptions and interpretation, the way a senior evaluator would in a peer review.
Here is my analysis plan, my Excel formulas and my three conclusions. Review them like a senior evaluator. Point out problems with method, denominators, disaggregation, assumptions and interpretation, and say which conclusions the data doesn’t support. Don’t rewrite the analysis for me.
Typical problems a review surfaces: duplicated records from a bad merge, the wrong denominator, averages of percentages, ignored missing values, charts with truncated axes, and conclusions that claim attribution the design cannot support.
Self-test: ask AI for a new practice dataset and problem, solve it without help, then compare your approach with alternatives. Later, ask for a slightly different version and solve it without looking at your earlier work.
4
Build fundamentals first, then let AI deepen them
AI supplements structured learning; it does not replace it. Learn the core concepts (descriptive statistics, sampling, indicator calculation, disaggregation, data cleaning, visualisation) from courses, textbooks and official guidance. Then use AI to explore and test what you have learned.
The reason is simple: AI is not always right, and you need enough knowledge to notice when it is wrong. Tomorrow’s evaluators will need verification skills as much as prompting skills.
Example: “Girls’ enrolment increased by 20%”
It sounds like a strong finding. Before it goes in your report, ask:
- A 20% relative increase or 20 percentage points?
- Compared with which baseline, and is the comparison period appropriate?
- Did the number of schools or the population change?
- Were records missing at baseline or endline?
- Did the indicator definition change, and what happened in comparison areas?
5
Work in small, reviewable iterations
Stop thinking in single prompts. Set up a loop where AI gives one hint at a time and only moves on after you show your attempt. Work alongside your real tools (Excel, R, Stata, SPSS, Jupyter or Google Colab) so you test every idea on the data yourself.
Act as a senior M&E analyst and mentor. I’m analysing this anonymised monitoring dataset. Guide me step by step, but don’t do the analysis for me. Give me the next hint only after I show you my attempt.
As you progress, deepen the questions: “What assumptions am I making?”, “What alternative explanations fit this result?”, “Argue against my conclusion.”
3. The struggle-first loop
All five methods share one rhythm. Use it every time you analyse data with AI nearby.
STEP 1
Attempt
Work on the problem yourself first, even if it’s slow.
STEP 2
Ask for a hint
When stuck, request a concept or clue, never the answer.
STEP 3
Get a review
Once finished, ask AI to critique your method and conclusions.
STEP 4
Revise and repeat
Fix what you agree with, defend what you don’t, and go again.
4. Verify before you use: a checklist for AI-produced figures
Whether you or AI produced a number, run these checks before it reaches a report, dashboard or presentation.
5. A four-week practice plan
Grow one project over four weeks, using a synthetic dataset that AI generates for you. About two hours a week is enough.
6. Protecting data while you learn
- Practise on synthetic data. Ask AI to generate realistic fake datasets; nobody’s privacy is at risk.
- Never upload identifiable data such as names, phone numbers, GPS points or IDs, or small-group breakdowns that could identify people.
- Share code, not data. For review, you can often paste your formulas, syntax or analysis plan without the dataset itself.
- Check your agreements. Your contract, data-sharing agreement and ethics approval decide what you may share with third-party AI services.
7. Knowledge check
Tap each question to reveal the answer.
1. What is the “illusion of understanding”?
Feeling that you understand an analysis because you have seen the result, when you could not produce, check or defend it yourself.
2. Your disaggregated totals don’t match the overall total. What should you ask AI?
Ask what you should check first and which concept you might be misunderstanding (for example duplicates, overlapping categories or missing values), not for the corrected numbers.
3. An indicator went from 40% to 50%. Is that a 10% or 25% increase?
Both, depending on how you express it: it is a 10 percentage-point increase and a 25% relative increase. Always state which one you mean.
4. When is the best moment to use AI as a reviewer?
After you have completed your own analysis, so it critiques your method, assumptions and interpretation instead of replacing them.
5. Why do you still need fundamentals if AI can run the analysis?
Because AI can be wrong, and only someone who understands the method can notice wrong denominators, misapplied weights or unsupported conclusions.
6. How can you get AI feedback on an analysis of sensitive data?
Share your code, formulas or analysis plan instead of the data, or practise on a synthetic dataset with the same structure.
8. Key takeaways
- Use AI as a mentor, tutor and reviewer, not as the analyst. You own the analysis.
- Ask for hints, concepts and critique rather than finished answers.
- Build fundamentals from structured learning; verification is now a core analytical skill.
- Work in a loop: attempt, hint, review, revise.
- Analysis is detective work. You learn far more by examining the evidence yourself than by reading someone else’s conclusion. Struggle first, then consult AI.
Build your AI skills for evaluation
Explore more EvalCommunity Academy tutorials and tools on responsible AI in monitoring, evaluation, accountability and learning.
Example prompts are illustrative and work with most AI assistants, including ChatGPT and Claude. Always follow your organisation’s data protection and AI policies. EvalCommunity is not affiliated with any AI provider.
