- Use AI to speed up qualitative and quantitative analysis while keeping results verifiable.
- Measure the reliability of AI-assisted coding through double coding of a sample.
- Protect the data of respondents and complainants at every step.
This module complements the 12 weeks of the programme. It adds no exam: certificates and the diploma are earned as before. Tools are described as known in November 2026; they change quickly.
1Where AI helps the MEAL team
In the MEAL cycle, generative AI is useful for repetitive text and structuring tasks:
- Qualitative analysis: proposing a first codebook from 50 open-ended answers, classifying verbatims, summarising focus group discussions.
- Data collection tools: reviewing a questionnaire (double-barrelled, leading or overly long questions), suggesting answer options, helping write the logic of a KoboToolbox or ODK form.
- Quantitative analysis: writing Excel formulas, R or Python scripts, explaining a statistical test.
- Reporting: a first draft of a report, a summary sheet, key messages for communities.
It is not reliable for producing figures: asking "what is the satisfaction rate?" of an assistant that has read a file can give a wrong result. Have it calculated by a formula or script that you can read and rerun.
2AI-assisted qualitative coding: the method
- Anonymise the answers before sending anything: remove names, numbers, precise locations and identifying details.
- Build the codebook with AI on a sample, then revise it as a team: clear definitions, examples, an "other" code.
- Code the full set with AI, giving it the codebook.
- Verify: an analyst codes a random sample by hand (for example 10%, at least 50 answers) and the agreement rate is calculated. Below about 80%, revise the definitions and recode.
- Document in the report: tool used, size of the verified sample, agreement rate, limitations.
Be careful with languages: a model often understands answers in local languages or phonetic transcriptions less well; check those answers more closely.
3Quantitative data: cleaning and calculating without inventing
Good practice:
- Ask the AI for the code or formula, not the result. Example: "Write an Excel formula that counts households with a food consumption score below 21".
- Test the code on a small sample whose answer you know.
- Keep a record: commented script, data version, cleaning steps (duplicates, outliers, inconsistencies).
- Never send the raw dataset with identifiers to a consumer tool; work on an anonymised dataset or an approved professional account.
For questionnaire translation, AI can produce a first draft, but back-translation and testing with a few respondents remain essential: one badly chosen word can change the meaning of an indicator.
4Complaints, feedback and sensitive data
Feedback and complaints mechanisms sometimes receive highly sensitive messages: sexual exploitation, abuse and harassment (PSEA), fraud, threats. Rules:
- PSEA allegations, serious fraud or threats are handled by trained people, following the confidential procedure, and never pasted into an AI tool.
- For routine feedback (distribution times, quality of an item), assisted classification can help, on a professional account, after anonymisation, and with human review of the "sensitive" categories.
- An automated system must never make a complaint "disappear": every message receives an acknowledgement and follow-up.
Remember: people who complain often take a risk. Trust in the mechanism depends on confidentiality.
5Reports, dashboards and evaluation ethics
AI can propose the structure of a report, rephrase for a non-technical audience, or turn findings into key messages. But:
- Every figure in the report must come from your analysis, with its source.
- Participant quotes must be real, never generated or "improved".
- An evaluation remains independent: do not let the tool soften negative findings.
- State the use of AI and its limitations in the methodology, as you would for statistical software.
Finally, share results with communities in their language: AI can help simplify a message, but a team member who speaks the language validates the final version.
Exercise: checking AI-assisted coding
Situation. A post-distribution survey collected 640 open-ended answers. AI coded them using an 8-theme codebook. You recode a random sample of 64 answers by hand: you agree with the AI on 50 of them. Of the 14 disagreements, 9 concern answers in a local language.
Questions. 1) What is the sample size as a percentage? 2) What is the agreement rate? 3) What do you do?
Answer key. 1) 64 / 640 = 10%. 2) 50 / 64 ≈ 78%, below the threshold of about 80%. 3) Revise the definitions of the codes concerned, have the local-language answers coded by an analyst who speaks the language (9 of the 14 disagreements, about 64%), recode and then check a new sample; document the approach and the final rate in the report.
Key points
- AI helps code, structure and draft; figures come from formulas or scripts that can be read again.
- Anonymise before sending anything; never a raw dataset with identifiers in a consumer tool.
- Check coding by double coding a sample and aim for at least about 80% agreement.
- Sensitive complaints (PSEA, fraud, threats) are handled by trained people, never by AI.
- Real quotes, findings not softened, use of AI stated in the methodology.
Self-assessment quiz
6 self-assessment questions, corrected immediately. They do not count towards the exams.
The Mini MBA AI and Humanitarian Action (Claude and Gemini) covers these topics in depth over 12 weeks.