How Can NGOs Use AI for M&E Data Analysis?

TL;DR:

• AI is useful for pulling evidence out of documents, a first sorting of open-ended answers, and drafting narrative from results you have already checked.
• It is unreliable for counting. Evaluators at Itad found counts produced by Copilot inconsistent, and relied on spreadsheet formulas for totals.
• Build your codebook and analysis plan first. Break the work into small steps, and have a person check a sample of every AI output.
• Why a number moved, and what the programme should do next, stays with the M&E officer.
• Clean the data before it goes in. Removing names is not enough when location, age and a rare event can still point to one person.

Who this guide is for

This guide is for M&E officers, MEAL staff and programme managers in NGOs, including teams where one person covers several projects. You need basic Excel. You do not need to know how to code.

It covers the analysis step: from cleaned data to findings you can defend. For turning those findings into a finished report, read our guide to writing donor reports faster with AI.

What can AI do well in M&E analysis?

AI does best on bounded jobs with clear instructions. In practice that means finding and extracting information from documents, suggesting possible themes, sorting text against definitions you supply, and drafting text from a results table.

The most useful public evidence comes from Itad, an evaluation firm that published a learning brief in March 2026. It reports three real evaluations where its teams tried AI tools.

Itad's summary of its own experience is short and worth remembering: AI is good at extraction, and people are good at making sense of it.

The wider picture is thinner than the marketing suggests. A December 2025 report from the Innovation and Research Caucus, which supports UKRI, found that evaluator practice is "in some areas moving faster than the published evidence, particularly in qualitative synthesis". It also found strong consensus for keeping a human in the loop.

Can you trust AI to code qualitative data?

Treat it as a first pass that a person reviews. In a 2024 preprint from Penn State researchers, two human coders reached a kappa score of about 0.73 on a deductive coding task. Agreement between those coders and GPT-4 was lower, at 0.44 to 0.50. The newer models tested, GPT-4o and Claude 3.5, scored 0.38 to 0.42 against the humans.

Kappa measures how much coders agree beyond chance, and 1.0 would be perfect agreement. That test used 200 data entries and models from 2024, so do not treat those figures as a verdict on today's tools. It also found that three separate GPT-4 runs agreed strongly with each other, at about 0.87. A tool can be consistent and still differ from your experienced coders. Run your own check on your own data.

One gap you should know about: we did not find a published test that we could open on Sheng or mother-tongue interview transcripts. Treat AI quality on those as unknown until you have tested it with a fluent speaker.

Where AI fits across the M&E cycle

The rule that prevents most errors: counts stay in the spreadsheet

Do every count, percentage and disaggregation in Excel or a statistics package, then give AI only the verified results.

Itad reports that aggregated counts from Copilot were inconsistent, so its team relied on spreadsheet formulas and human quality checks for totals and derived figures. A formula gives the same answer each time and anyone can inspect it. That is what you need when a donor or auditor asks where a number came from.

AI is still helpful here, as a tutor. Ask it to write or explain the formula, then let Excel do the calculation. If you want to build this skill, our Excel and AI training in Nairobi covers it.

A safe method in seven steps

Itad's lessons, and basic data protection, point to the same working order.

1. Write your analysis plan first. Set out the evaluation questions, indicators, comparison groups and codes before opening an AI tool. Itad found that prompts built around a framework, with clear definitions and example answers, reduced drift and off-topic content.

2. Clean the data. Remove names, phone numbers, ID numbers and household identifiers. Then deal with the less obvious ones. The UN's Centre for Humanitarian Data warns that combining key variables such as location or ethnicity can still allow re-identification, even after direct identifiers are gone. In Kenya's Data Protection Act, anonymisation means removing identifiers so the person is no longer identifiable. If someone could still work it out, the data is not anonymised.

3. Do the numbers yourself. Produce your indicator tables in Excel. Merge very small groups into wider categories. Our data privacy guide gives a practical small-numbers rule.

4. Split the work into steps. Itad's suggested order is to define the framework, extract evidence, classify where appropriate, synthesise by component, then integrate across sources. Giving a tool one small job at a time reduces drift and makes checking easier.

5. Limit the context. For nuanced material, analyse one code, case or component at a time. Itad's interview analysis only worked well after the team narrowed what the tool had to handle.

6. Check a sample. Read a sample of source material against the AI output. Score it for completeness, relevance, fidelity and invented content. Itad's approach included a check of about 15% of proposals in one evaluation. The right share depends on the stakes, so decide it in advance and write it down.

7. Keep a record. Save prompts, tool versions, verification notes and decisions. If a client commissioned the work, agree tool use with them and explain in your method where AI was used.

Worked example: from indicator table to careful narrative

Everything in this example is invented for training. Suppose a programme runs a life-skills course for girls at three sites.

In Excel, if the target is in column B and the actual is in column C, the achievement in row 2 is =C2/B2. The programme total comes from =SUM(C2:C4)/SUM(B2:B4).

Here is a trap to look for. The average of the three site percentages is 85.3%, which is a different figure from the true programme result of 88.8%. A narrative can quote the wrong one and still sound correct. This is one more reason to supply the checked total yourself.

Now give AI only the verified table and what you actually know.

You are helping an M&E officer write a results paragraph. Use only the table and field notes below. Do not calculate new numbers. Do not say why a result happened unless the field notes say so. Do not use words like "caused", "led to" or "due to" unless the notes support a cause. If something is missing, write [QUESTION FOR M&E OFFICER]. Maximum 120 words, plain English, for a programme review meeting. Table: [paste verified table]. Field notes: [paste cleaned notes]

A weak draft from any tool might say: "The programme's strong community engagement drove results at Site B, while Site C lagged due to low interest." Nothing in the table supports either claim.

A sound paragraph looks like this, once the officer has added what happened on the ground:

The course reached 710 girls against a target of 800 (88.8%). Site B exceeded its target at 106%, and Site A reached 88%. Site C reached 62%. The field team reports that sessions at Site C started late. We have not yet established whether this explains the shortfall, and we will review attendance records with the site team before the next quarter.

The M&E officer has to write the final two sentences, because only the officer knows what happened at Site C.

Five prompts you can copy

Use these only on cleaned material. Replace the square brackets.

1. Check your indicators

Below are the indicators from a programme logframe. For each one, tell me whether it is specific and measurable, what is unclear, and what data source it needs. Do not rewrite the programme's objectives. Give your answer as a table. Context: this is a [sector] programme in Kenya. Indicators: [paste]

2. Get a formula and an explanation

I have an Excel sheet with these columns: [list column names and what each holds]. Give me the formula to calculate [measure] for each row, and a separate formula for the total. Explain each one in a sentence. Tell me one way to test that the formula is correct. I use Microsoft Excel.

3. Draft a codebook from a small sample

Here are 30 cleaned open-ended answers to the question: "[question]". Suggest up to 8 codes. For each code, give a name, a one-sentence definition, one example of what belongs and one example of what does not. Add a code for "unclear". Do not count anything. Answers: [paste]

Treat the result as a draft. Add the codes that matter to your evaluation questions, and remove those that do not. Then have your coders test it.

4. Apply your codebook, with evidence

Apply the codebook below to each answer. Give a table with the answer number, up to two codes, the exact words from the answer that support each code, and your confidence as high, medium or low. If no code fits, write "unclear". Never invent a quotation. Do not give totals; I will count in Excel. Codebook: [paste]. Answers: [paste in batches]

Asking for the exact words makes checking far quicker. Put the output in Excel and count there.

5. Challenge your own draft

Read this draft findings section as a strict external evaluator. List every claim that goes further than the evidence shown, every causal claim without support, and every place where a group or site is treated as if it were the whole programme. Quote the sentence and explain the problem. Do not rewrite it. Draft: [paste]

This one is often the best use of AI in an M&E team, because it adds a second reader to work you already own.

Which tools can you use? (as of November 2026)

No single tool is right for everyone. Choose by data sensitivity first, then by what your organisation already has.

Dedicated qualitative software, such as NVivo or MAXQDA, may suit evaluations with heavy coding. Test any automated feature first. Itad chose not to use MaxQDA's automated coding because it over-coded transcripts.

For very sensitive material, some teams consider running a model on their own computer. Our guide to running open-source AI models locally explains what that takes.

What you must keep out

Do not paste individual beneficiary records, case notes, safeguarding reports, or raw interview transcripts that still have identifying details into a public AI tool.

M&E data often falls within what Kenya's Data Protection Act calls "sensitive personal data", which includes health status, ethnic social origin, family details and sexual orientation. Section 31 requires a data protection impact assessment where processing is likely to result in high risk.

Section 53 has provisions on research and statistics, but it does not give M&E teams a free pass. It depends on conditions, including that results are not made available in a form that identifies people. Take advice before relying on it.

If your programme works with children or survivors, read our guide to AI and safeguarding first. Some data should not go near these tools, even if cleaned.

What stays with the M&E officer

AI can describe the pattern in the data you give it. It has no knowledge of what happened on the ground unless you tell it.

These tasks stay with you:

• Deciding whether the data is good enough to use at all.
• Explaining why a number moved: a drought, a teachers' strike, a change in staff or a late disbursement.
• Judging whether a change can be linked to the programme.
• Noticing who is missing from the data.
• Choosing what the team should do differently, and telling leaders what they would prefer not to hear.

AI can reduce the routine load so that more of your time goes to these tasks.

What goes wrong, and how to avoid it

A final checklist before findings leave your desk

• [ ] Every number traces back to a spreadsheet cell or source table.
• [ ] Totals and percentages were calculated outside the AI tool.
• [ ] A person checked an agreed sample of AI coding or extraction.
• [ ] Every quotation has been found, word for word, in the source.
• [ ] Results were checked for key groups, sites and any disaggregation you promised.
• [ ] No causal claim goes further than your evaluation design allows.
• [ ] No small group or rare detail could identify someone.
• [ ] Translated material was checked by a fluent speaker.
• [ ] Prompts, tools, versions and checks are recorded.
• [ ] The donor or commissioner's rules on AI use were followed.

Questions people ask

Can AI analyse my survey data for me?

It can help, but keep calculation in Excel or a statistics package. Use AI for formulas, first sorting of open-ended answers and drafting from verified results. In one Itad evaluation, counts from Copilot were not reliable enough to use.

Is it safe to upload our dataset to a free AI tool?

Do not upload raw records with personal data to a public tool. Prefer aggregated, cleaned data. Check each tool's settings, because some free consumer services may review or use conversations.

How much of the AI's coding should we check?

Decide in advance, based on the stakes. Itad checked about 15% of proposals in one evaluation, and in another its team reviewed and corrected every transcript. Start high with a new tool or a difficult topic. Reduce the share only if your checks show that it is reliable.

Can AI handle focus group discussions in Kiswahili or Sheng?

Some tools transcribe and translate many languages, so test yours first. Have a fluent speaker compare a sample with the audio. Sheng and code-switching need extra care, and we could not find a published test we could rely on.

Do we have to tell donors or commissioners we used AI?

Check your agreement first. Itad recommends agreeing tool use with clients and explaining where AI sits in your method. Our donor report guide sets out what funders have published.

Will AI replace M&E officers?

We see no sign that it can do the part that matters most: judging evidence and explaining what happened. Routine parts of the job may shrink. Officers who can design checks and question AI outputs will be better placed.

What is the best first task for a small M&E team?

Try the challenge prompt on a findings section you have already written. It uses text you own, and you can judge the output immediately. Then move on to formulas and first drafts from verified tables.

Sources

All links last checked 10 November 2026.

• Itad, The evaluator's guide to generative AI, by Chris Perry, Michael Moses and Max Pinter (March 2026)
• Innovation and Research Caucus, Using AI in Evaluation: Evidence, Practice and Principles for Responsible Use (17 December 2025)
• Zhang et al., When Qualitative Research Meets Large Language Model: Exploring the Potential of QualiGPT as a Tool for Qualitative Coding, arXiv (2024)
• Centre for Humanitarian Data (OCHA), Guidance Note on Statistical Disclosure Control
• The Data Protection Act, No. 24 of 2019 (ODPC copy of the Gazette version)
• The Data Protection Act, No. 24 of 2019 (ICNL text copy, used to confirm section numbers)
• KoboToolbox, Qualitative analysis of audio responses (updated 6 October 2026)
• KoboToolbox, Transcription and translation (updated 6 October 2026)
• Microsoft Support, Analyze Data in Excel
• Microsoft Support, Get started with Copilot in Excel
• Microsoft Support, Get started with Microsoft Copilot
• Microsoft Learn, Enterprise data protection in Microsoft Copilot and Microsoft Copilot Chat
• OpenAI, ChatGPT pricing
• Google, Gemini subscriptions
• Google, Gemini Apps Privacy Hub
• Anthropic, Claude pricing

Where FuKazee fits in

FuKazee trains NGO programme and M&E teams to use AI on their real reporting work, with the judgement staying where it belongs. We assume nobody in the room knows AI yet, we avoid jargon, and we show the free version of each tool. We also teach staff to check every output. Nobody becomes an expert in a few days, and we say so.

Our exercises use invented examples, so no real beneficiary data enters a tool during training. Our programme for NGOs includes M&E summaries alongside proposals, donor reports and communications. See the full outline on our page for AI training for NGOs in Kenya, or read how we suggest training NGO staff safely.

To talk through training for your M&E or programme team, email lead@fukazee.com.