Build Your Personal AI Agentic System for Evaluators and Development Professionals

A practical, step-by-step guide for evaluators, M&E specialists, MEAL professionals, consultants, researchers and international development practitioners.

Difficulty: Beginner to intermediate

Technical skills: No coding required for Version 1

What you will build: A personal AI-powered M&E workspace with your own knowledge, workflows, tools and quality controls.

What Are You Actually Building?

You are not trying to create an AI that does everything.

You are building a workspace that can help you with the parts of M&E work that you repeatedly do — reading, organizing, comparing, analyzing, drafting, checking and improving.

For example, your system could eventually help you:

• review Terms of Reference
• develop evaluation designs
• develop and test theories of change
• review logframes
• develop indicators
• prepare M&E plans
• analyze interview and focus-group data
• analyze survey data
• review evaluation reports
• identify evidence gaps
• prepare donor reports
• support proposal development
• conduct desk research
• develop recommendations
• reuse your own templates and methodologies

A useful way to think about the system is:

That is the important difference between simply prompting AI and building a repeatable professional workflow.

Before You Start

You do not need programming skills to build the first version.

For this tutorial, start with:

• ChatGPT
• your own M&E documents
• your templates
• your previous work
• a few trusted reference documents

Later, you can add web research, data analysis, connected applications and more advanced automation.

From an AI Assistant to an Agentic System

A traditional chatbot often follows a simple pattern:

Your personal M&E system can become more structured:

This is the beginning of an agentic workflow. It does not necessarily mean that you have built a fully autonomous AI agent. It means that AI is organized around a defined objective, workflow, context, tools and verification process.

You do not need multiple AI models to do this. In Version 1, these stages can simply be different prompts, chats or workflow instructions inside your Project. Later, the same architecture can be implemented with specialized AI agents, retrieval systems, databases, external tools and automation.

Three Levels of Agentic Development

For this tutorial, you will primarily build Stage 1 and begin introducing the design principles of Stage 2. Stage 3 is optional.

Part 1 — Create Your Personal M&E Workspace

Step 1 — Open ChatGPT

Open ChatGPT in your browser.

Look at the left-hand sidebar and find the option for creating a new project.

Click New project.

A Project gives you a place to keep related conversations, files and project-specific instructions together.

Step 2 — Name Your Project

Give the project a name that you will recognize immediately.

I suggest:

You could also call it:

• My Evaluation Copilot
• My M&E Consultant AI
• My Evaluation Assistant
• My Development Practice AI

For this tutorial, we will use My Personal M&E AI System.

Step 3 — Organize Your Files on Your Computer

Before uploading anything, spend a few minutes organizing the documents you want to use.

Create a folder on your computer called:

Inside it, create these folders:

These folders are on your computer. They are simply a good way to keep your source material organized before you upload files to your ChatGPT Project.

Part 2 — Build Your Professional Knowledge Base

Step 4 — Create Your Professional Profile

Create a document called:

This does not need to be a long CV. It is simply a short description of how you work.

Include things such as:

This gives your AI some context about the professional environment in which you work.

Step 5 — Collect Your M&E Methodologies

Open the folder:

Add the guidance you actually use, such as:

• M&E frameworks
• evaluation methodologies
• Theory of Change guidance
• logframe guidance
• indicator guidance
• evaluation criteria
• UN evaluation guidance
• donor guidance
• qualitative research guidance
• survey methodology
• sampling guidance
• data quality guidance

Start small. Five good reference documents are more useful than 100 documents you never use.

Step 6 — Add Your Templates

Open:

Add the templates you use regularly:

• Evaluation Plan Template
• M&E Plan Template
• Theory of Change Template
• Logframe Template
• Indicator Reference Sheet
• Interview Guide
• Focus Group Guide
• Survey Questionnaire
• Data Analysis Plan
• Evaluation Report Template
• Donor Report Template
• Recommendation Matrix
• Risk Register
• Data Quality Assessment

This is where the system starts becoming genuinely personal.

Instead of saying:

you can say:

The difference is significant: AI is now working with your professional resources, not just its general knowledge.

Part 3 — Upload Your Knowledge

Step 7 — Upload Your Important Files

Return to your ChatGPT Project.

Look for the option to add files or Add files.

Start by uploading a small group of high-value documents.

For example:

• 01_My_Professional_Profile.docx
• M&E_Methodology.pdf
• Theory_of_Change_Guide.pdf
• Evaluation_Plan_Template.docx
• Evaluation_Report_Template.docx
• Indicator_Guide.pdf

Step 8 — Create a “START HERE” Document

Create:

This document tells the AI what the Project is for and how you want it to work.

Include the following:

How Your Knowledge Becomes Searchable as the System Grows

Retrieval-Augmented Generation (RAG)

When you have only a few documents, you can work directly with uploaded files. As your knowledge base grows, however, a new question appears: How does the AI find the right information among hundreds or thousands of documents?

This is where Retrieval-Augmented Generation (RAG) becomes useful. RAG introduces a retrieval layer before generation: the system searches a knowledge source for relevant information and then provides that context to the language model.

For evaluators, this leads to an important principle: retrieval quality affects answer quality. If important evidence is not retrieved, the model may produce an incomplete answer even when the information exists somewhere in the knowledge base.

As your system grows, ask not only “What does my knowledge base say?” but also:

Part 4 — Give Your AI Its Working Instructions

Step 9 — Open Project Settings

Inside your Project, look for the three-dot menu (…) near the top of the Project.

Click it and look for Project settings.

This is where you can add instructions that apply to the Project.

Step 10 — Add Your M&E Instructions

You can start with the following instructions and modify them as you gain experience:

Part 5 — Build Your First Workflow

Step 11 — Start With Evidence Review

Now you have your knowledge and your instructions. The next step is to create a repeatable workflow.

For M&E professionals, a very useful first workflow is an Evidence Review Workflow.

The question behind it is:

Step 12 — Create the Evidence Review Chat

Inside your Project, start a new chat.

Call it:

Use this prompt:

Use Both Local and Global Reasoning

Not every evaluation question requires the same type of reasoning.

Some questions are primarily local: where a report discusses sampling, what indicator was used, what value was reported, or which recommendation refers to a particular result.

Other questions are global: whether the methodology is coherent, whether evidence supports the conclusions, whether a Theory of Change is logically plausible, or whether recommendations are consistent with findings.

A mature M&E AI system should use both approaches: retrieve specific evidence for local questions, then step back and assess relationships across the document for global questions.

Part 6 — Create Your M&E Control Panel

Step 13 — Create Your Core Workflows

As you build the system, create separate chats for the work you do repeatedly.

You don’t have to create all of these today. Add them when you have a real need.

Your Workflows Can Eventually Become Specialized Agents

At this stage, your workflows may simply be separate chats with different instructions. That is perfectly acceptable. As your system becomes more sophisticated, you can think of each workflow as a specialized agent role.

You do not need to build these as separate technical agents today. The important design principle is separation of responsibilities.

Part 7 — Build Your Personal Knowledge Assistant

Step 14 — Ask Questions About Your Own Documents

Create a chat called:

Then test it with questions such as:

This is one of the biggest advantages of building your own knowledge base: you can ask questions about your own professional resources rather than starting from zero every time.

Part 8 — Build Your Evaluation Design Agent

Step 15 — Create “02 — Evaluation Design”

Use an instruction such as:

Then upload a Terms of Reference and ask:

Your workflow becomes:

Part 9 — Build Your Theory of Change Agent

Step 16 — Create “03 — Theory of Change”

Ask the AI to examine:

But don’t stop there.

Ask it to identify:

• assumptions
• causal mechanisms
• external factors
• risks
• unintended outcomes
• indicators
• evidence gaps

Part 10 — Build Your Indicator Assistant

Step 17 — Create “04 — Logframe & Indicators”

Give the workflow a clear instruction:

Then ask:

A useful output could look like this:

This is much more useful than simply asking AI to “give me 10 indicators.”

Part 11 — Build Your Evaluation Report Reviewer

Step 18 — Create “09 — Evaluation Report Reviewer”

Upload an evaluation report and ask:

Ask it to examine:

1. Evaluation questions
2. Methodology
3. Sampling
4. Data sources
5. Analysis
6. Findings
7. Evidence
8. Conclusions
9. Recommendations
10. Limitations

Then add the most important question:

You can go one step further:

Step 18A — Build an Evidence Chain

For important findings, ask the AI to trace the reasoning chain rather than simply summarize the report.

Then ask:

This turns the AI from a simple report summarizer into a quality-control assistant and makes the output easier for a human evaluator to audit.

Part 12 — Build Your Qualitative Analysis Workflow

Step 19 — Create “06 — Qualitative Analysis”

Depending on your project and data-protection requirements, you may work with:

• interview transcripts
• focus-group transcripts
• field notes
• coding frameworks
• research questions

Run the analysis in stages:

1. Develop an initial codebook.
2. Code the evidence.
3. Identify themes.
4. Look for contradictory or negative cases.
5. Compare themes across respondent groups.
6. Produce an evidence-grounded thematic analysis.

Step 19A — Separate Coding From Interpretation

For larger qualitative datasets, avoid asking the AI to interpret everything in one step.

This separation is useful because coding and interpretation are not the same activity. Ask the system to preserve the link between an important theme and the underlying excerpts or evidence.

Part 13 — Build Your Research Assistant

Step 20 — Create “08 — Research Assistant”

A good research workflow looks like this:

Ask:

When the question involves current information, use web research rather than relying only on your uploaded documents.

Part 14 — Add Your Working Preferences

Step 21 — Create “My_Working_Preferences.docx”

Write down how you prefer to work.

Upload this document to your Project.

Over time, your system becomes more useful because it reflects your way of doing professional work.

Part 15 — Add Tools Carefully

Step 22 — Start With the Safest Tools

Tools allow an AI system to interact with information outside a simple conversation.

Depending on the tools available to you, these might include:

• files
• web research
• data analysis
• cloud storage
• calendar
• email
• other connected applications

Don’t connect everything immediately.

A sensible progression is:

Start with tools that read and analyze information before giving an AI system permission to change or send information.

Part 16 — Keep Yourself in the Loop

Step 23 — Define Approval Rules

Create:

For example:

The principle is straightforward: let AI handle repetitive work, but keep professional judgment with the professional.

Part 17 — Put the System Together

Step 24 — Your Personal M&E AI Architecture

Each part has a job:

• Knowledge: your methodologies, reports, templates and research.
• Instructions: how you want AI to work.
• Workflows: repeatable professional processes.
• Tools: web, files, data and connected applications.
• Verification: checks whether the work is good enough.
• Human review: your professional judgment.

Part 18 — Test It With a Real Evaluation

Step 25 — Use a Real Assignment

Don’t test your system with:

That doesn’t tell you much.

Give it a real professional task.

A useful system should be able to help you:

1. understand the TOR
2. identify the evaluation purpose
3. identify stakeholders
4. identify evaluation questions
5. identify methodology requirements
6. identify missing information
7. identify risks
8. identify assumptions
9. identify expected deliverables
10. prepare clarification questions

Advanced Architecture — Designing the Agentic Workflow

Step 25A — Understand the Core Components

Your personal AI system can be understood as several cooperating components.

Step 25B — Use a Modular Workflow

This is more robust than asking one prompt to perform every task simultaneously.

Step 25C — Add Checkpoints

Each important stage should produce an intermediate result that can be inspected:

• What evidence was retrieved?
• What assumptions were identified?
• What analysis was performed?
• What uncertainties remain?
• What did the verification stage find?
• What changed during refinement?

Part 19 — Build the System in Five Levels

There is no need to reach Level 5 immediately.

For many evaluators, Levels 1–3 will already save a substantial amount of time.

Bonus — 15 AI Workflows for Evaluators

Once your basic system is working, you can add specialized workflows for the tasks you perform most often.

Your 60-Minute Setup Challenge

If you want to try this today, don’t overthink it.

The Most Important Rule

Don’t try to build:

“An AI that can do everything.”

Build:

“A collection of small AI workflows that do specific things well.”

For an evaluator, that might eventually become:

Why One Giant Prompt Is Not the Answer

When people discover AI, they often try to solve increasingly complex problems by making the prompt increasingly long.

For complex M&E work, a single prompt may be asked to retrieve information, understand multiple documents, apply methodology, analyze evidence, compare sources, identify contradictions, make judgments, write recommendations and format the final output.

You do not need four separate technical AI agents to benefit from this principle. You can simply structure your prompt and workflow into stages.

The Mental Model to Remember