ChatGPT for evaluators 20 techniques

EvalCommunity Academy · AI Tools for M&E Practice

20 Things You Can Do With ChatGPT That Most Evaluators Never Try

Stop using AI as a search box. Twenty practical techniques that turn ChatGPT into a critical friend, document reviewer, data checker and research assistant for monitoring, evaluation, accountability and learning.

Level: Beginner to intermediate · Reading time: about 15 minutes · Updated: November 2026

Before you start

Features such as file analysis, Projects, voice, memory and deep research vary by ChatGPT plan and account, and they change often. Most techniques here also work in other assistants such as Claude. Above all: never upload identifiable respondent data, safeguarding information or client material your contract does not allow you to share with third-party AI services. Section 7 covers this in detail.

Learning objectives

By the end of this tutorial you will be able to:

• Use AI to sharpen evaluation questions and stress-test emerging findings.
• Compare, review and check documents such as ToRs, logframes and reports against frameworks.
• Run first-pass data quality checks and spot misleading charts.
• Set up long-running workspaces and trusted-source research for evaluation assignments.
• Decide which parts of your work to delegate and which to keep, and protect data while doing so.

ToRs often list fifteen questions while the client’s real concern sits underneath. Instead of drafting questions yourself, describe the situation and let ChatGPT interview you until the core issue is clear.

The client wants a mid-term evaluation of a youth employment project, but I can’t tell whether their real concern is low enrolment, poor job placement or partner performance. Ask me one question at a time until we can define the main evaluation question.

AI is least useful when it agrees with you. Before a finding goes into the report, ask for the strongest reasonable case against it. This is a quick rehearsal of the challenge you will face at the validation workshop.

My emerging finding: the training component improved women’s income. Identify the assumptions behind this, give me the strongest alternative explanations, and tell me what evidence would distinguish between them.

This turns a prediction into a test, which is especially valuable for contribution claims. You get a list of checks to run before you commit to a conclusion, and the AI does not become the decision-maker.

I believe the project contributed to reduced school dropout. What would I need to find in the data or interviews for that conclusion to be wrong? List the checks I should run before finalising.

Most evaluators repeat a few costly patterns: underestimating fieldwork time, accepting vague ToRs, agreeing to a sample the budget can’t support. Describe a few past cases and ask for the pattern and a reusable checklist.

Here are three assignments where my timeline slipped and why. Identify the recurring pattern and turn it into a short go/no-go checklist I can use before accepting a new evaluation.

Summaries side by side hide contradictions. Upload two documents (two evaluations of similar programmes, a baseline and midline report, a project document and a partner report) and ask for disagreements, not summaries.

Don’t summarise these separately. Identify where they actually disagree, where they rely on different evidence, and where they seem to agree but define the same term differently (for example “beneficiary” or “employment”).

A revised ToR, logframe or contract arrives and you don’t want to reread everything. Upload both versions and say what counts as a meaningful change.

Compare version 1 and version 2 of this ToR. Ignore formatting. List every substantive change to scope, evaluation questions, deliverables, deadlines, budget or required expertise, and flag anything that affects my workload.

This turns ChatGPT from a summariser into a reviewer. Pair a framework with a draft: a donor’s report quality checklist and your draft report, the OECD-DAC criteria and an evaluation matrix, a proposal template and your technical proposal.

Review my draft evaluation report against the attached quality checklist. For each criterion, rate it met, partly met or not met, quote the relevant passage, and suggest what is missing.

Tip: run this before submitting to a client, as a self-assessment. It catches structural gaps; it cannot judge the quality of your evidence.

Stuck on a KoboToolbox form error, a Power BI setting or an Excel formula? Upload a screenshot instead of describing it. Asking what it cannot tell from the image keeps the answer honest.

This XLSForm validation error appears when I upload my survey. What does it seem to be telling me, and what other information would you need before suggesting a fix?

Before analysing an Indicator Tracking Table or survey export, ask what questions the data is capable of answering. This stops you designing an analysis the data cannot support.

Before analysing this anonymised monitoring dataset, tell me which evaluation questions it can and cannot answer, and why. Then identify the three patterns or anomalies most worth investigating.

A fast first-pass data quality check: outliers, blanks, duplicates, impossible combinations and sudden jumps. Asking it not to assume errors keeps judgement with you.

Flag values that look unusual in this partner monthly report: blanks, duplicates, figures over 150% of target, month-on-month changes over 30%, or impossible combinations. Don’t assume they are mistakes; explain why each one stands out.

Useful for donor dashboards, partner presentations and your own report graphics. Ask about axes, scale, time period, missing context and visual framing.

Explain what this results chart appears to show, then tell me which design choices or missing information (baseline, denominator, time period, disaggregation) could lead a reader to the wrong conclusion.

Stop re-explaining the context in every new chat. A Project keeps related conversations, reference files and instructions together. Create one per evaluation with the ToR, inception report and evaluation matrix, plus standing instructions.

Project instructions: This is the final evaluation of [project]. Use the ToR and evaluation matrix as the reference for all questions. Write in [your preferred language and spelling], use the OECD-DAC criteria, and always cite which file a statement comes from.

Fill a Project with sector evidence you trust: past evaluations, systematic reviews, guidance notes and your own notes. Then question it during desk reviews and proposal writing.

Using only the files in this Project, what does the evidence say about cash transfers and school attendance for adolescent girls? Separate strong findings from weak ones and name the source of each.

Where voice supports files and Projects, you can review a long inception report on a commute or a walk between field sites. Interrupt, question and ask it to go back, as you would with a colleague.

Talk me through the three parts of this evaluation methodology that a sceptical client is most likely to challenge.

That sampling note, KII guide or theory of change you drafted months ago is still in your history. Search your chats and Projects before recreating it from scratch.

Find the key informant interview guide for local government officials I worked on earlier this year, and adapt it for a WASH evaluation.

With memory or personalisation turned on, ChatGPT can look across your planning conversations for recurring blind spots. Treat the answer as a prompt for reflection, not a verdict.

Looking at the evaluation workplans we’ve discussed, which constraints do I usually forget to mention until late: translation, ethics approval, security, travel time, or something else?

Deep research can investigate a question across many sources. Where your plan allows, focus it on selected sites, such as evaluation offices, evidence repositories and official statistics, instead of the whole web.

Research what rigorous evaluations say about graduation programmes in fragile settings. Prioritise UN evaluation offices, the World Bank IEG, 3ie and peer-reviewed studies. Exclude blogs and news. Cite every source.

Tip: open and check every cited source. AI research can still misstate findings or cite the wrong document.

Desk reviews can expand without end. Ask how much research a question deserves given the budget, timeline and stakes, and set a stopping rule.

I have three days for the desk review in a 25-day evaluation. Given the evaluation questions, which sources are essential, which are nice to have, and what stopping rule should I use?

Presenting negative findings to a programme team, negotiating scope with a client, or interviewing a guarded official. Ask ChatGPT to play the other person, first cooperative, then sceptical, then unpredictable.

Play a defensive programme director hearing that the project’s main outcome was not achieved. I’ll present the finding; push back the way a real director might. After three rounds, give me feedback on my delivery.

Perhaps the most important question of all. Ask AI to separate tedious work from the parts where your judgement, experience and ethics matter, and to help only with the first.

I’m writing the findings chapter of a final evaluation. Separate the work into parts where you can remove tedious effort and parts where my evaluative judgement must lead. Help me with the first category without replacing the second.

AI can organise evidence, check references and spot unclear passages. Deciding what the evidence means, how strong it is, and what to recommend remains the evaluator’s job.

Never upload

• Identifiable data on participants, respondents or interviewees, including transcripts with names.
• Safeguarding, protection or PSEA case information.
• Data from contexts where exposure could put people at risk.
• Client documents your contract, data-sharing agreement or ethics approval does not allow you to share.

Responsible-use checklist

• I have checked my organisation’s AI policy and my client contracts.
• I have reviewed my data and training settings in my ChatGPT account.
• Datasets are anonymised or aggregated before upload.
• I verify every figure, quote and citation against the original source.
• I disclose AI assistance where the client or evaluation policy requires it.
• Findings, conclusions and recommendations are my own professional judgement.

Build your AI skills for evaluation

Explore more EvalCommunity Academy tutorials and tools on responsible AI in monitoring, evaluation, accountability and learning.

Explore EvalCommunity Academy