How AI assistants like Claude can strengthen monitoring, evaluation and learning
A practical guide to where AI helps across the MEL cycle, where it should not be used, and the safeguards that keep evidence credible.
Draft. This piece is still being reviewed and may be revised before it is finalised.
AI assistants such as Claude, developed by Anthropic, have moved quickly from novelty to everyday tool in research and policy work. For monitoring, evaluation and learning (MEL) teams, the opportunity is not to hand over judgement, but to spend less time on routine drafting, coding and summarising, and more time on the parts of the work that need people: understanding context, talking to communities and deciding what the evidence means.
Used carelessly, the same tools can leak sensitive data or introduce confident-sounding errors into a report. This guide maps where AI assistants add value across the MEL cycle, where they do not belong, and the safeguards we recommend before any team adopts them.
Where AI assistants help across the MEL cycle
1. Design
An assistant can review a draft results framework or logframe and flag indicators that are vague, hard to measure or missing a baseline. It can suggest survey questions, point out double-barrelled or leading wording, and produce first drafts of translations. Translations still need checking by fluent speakers, especially in languages such as Somali or Swahili where AI performance can be less reliable than in English.
2. Data collection
Teams that build digital questionnaires in KoboToolbox or ODK can use an assistant to write and debug XLSForm logic, such as skip patterns, constraints and calculations, and to explain errors in plain language. It can also help prepare enumerator training materials and field manuals, which a supervisor then reviews and adapts.
3. Cleaning and analysis
This is where many teams see the clearest gains. Claude can write and explain code in Stata, R or Python for data cleaning, high-frequency checks and statistical analysis, and can help trace why a script fails. The result is reproducible code that an analyst can read, test and keep, rather than a black box. The analyst remains responsible for checking that the code does what the analysis plan requires.
4. Qualitative analysis
With properly anonymised transcripts, an assistant can propose a first pass of coding against an agreed codebook, summarise themes and pull out illustrative quotations. Treat this as a starting point, not a finding: a human researcher should review the coding, and a second coder or spot checks remain good practice. AI can miss irony, local idiom and what respondents leave unsaid.
5. Reporting and learning
Claude can work with long documents, including lengthy PDF reports, and can summarise them or pull out recommendations across several sources. It can help turn an evaluation report into a two-page brief, a plain-language summary for communities, or slides for a learning event. When asked to work from source documents, it can cite the passages it draws on, which makes checking much faster.
What AI should not do
- Make evaluative judgements. Whether a programme was relevant, effective or sustainable is a judgement for evaluators, grounded in evidence and context.
- Replace field presence. No tool substitutes for observing a site, listening to participants or understanding local power dynamics.
- Produce or fill in data. AI must never be used to generate responses, impute findings or invent figures.
- Be the final check. AI output can sound authoritative and still be wrong. Every figure, citation and claim needs verifying against its source.
Safeguards for responsible use
Protect respondents first. Never paste personally identifiable information, such as names, phone numbers, locations of individuals or anything that could identify a respondent, into an AI tool. Anonymise data before analysis. In Kenya, the Data Protection Act, 2019 applies to personal data, and donor and partner agreements often add their own rules. Use organisational accounts, and check how your provider handles and retains the data you share before you start.
Keep a human accountable. Name the person responsible for each AI-assisted output, and make review a step in the workflow rather than an afterthought.
Verify everything. Check figures against datasets, quotations against transcripts and references against the original documents.
Be transparent. State in methodology sections where and how AI tools were used, so that clients and readers can judge the evidence for themselves.
Watch for bias. AI models reflect the data they were trained on, which under-represents many African contexts and languages. Test outputs with people who know the context.
Getting started
Start with low-risk tasks that involve no personal data, such as writing and explaining analysis code, drafting survey tools or summarising published reports. Agree a short AI use policy that covers data protection, review and disclosure. Train the team, pilot on one assignment, and record both the time saved and any problems found. Scale up only once the safeguards are working in practice.
Key takeaways
- AI assistants such as Claude can speed up drafting, coding, summarising and first-pass qualitative work across the MEL cycle.
- Judgement, field presence and final verification stay with people.
- Protecting respondent data comes first: anonymise, use organisational accounts and follow the Data Protection Act, 2019 and partner agreements.
- Disclose AI use in methodology, and verify every figure, quotation and citation.
We offer training on the responsible use of AI tools in MEL and research, alongside our data analytics work. Get in touch to discuss how your team could adopt AI safely.
About Quantec Insight. A Kenya-based research, monitoring and evaluation, data analytics and economic policy consultancy. Learn more about us.
Acknowledgement: this piece was prepared and reviewed by the Quantec Insight team.