Build an AI Agent for Real M&E Work

Select a case.

REMEMBER — How to do this
Choose a task you actually perform. Good candidates are repetitive, evidence-heavy, time-consuming or require structured comparison. Do not choose a task simply because it sounds impressive.

REMEMBER — How to decide
If a formula, spreadsheet, statistical procedure, deterministic automation or single AI interaction solves the task adequately, you may not need an agent.

INPUT → CHECK → RETRIEVE → ANALYSE → COMPARE → VALIDATE → ESCALATE → HUMAN DECISION

REMEMBER — How to do this
You can map the process in a document, spreadsheet, whiteboard, Miro/Lucidchart or directly in ChatGPT/Claude. Ask the model to identify steps, decision points, evidence requirements and existing quality controls.

Choose an architecture.

REMEMBER — Where to build it

• ChatGPT / Claude: prototype the reasoning and instructions.
• Make.com: connect triggers, applications, routing and actions.
• Retrieval: connect approved evidence sources.
• Tools: use spreadsheets, databases, calculators or code when necessary.
• Human-in-the-loop: insert approval before consequential outputs.

Choose the M&E process.

REMEMBER — How to translate methodology into agent steps

Evidence: Retrieve → Extract → Verify → Compare → Assess → Synthesise → Triangulate
Qualitative: Prepare → Preserve context → Segment → Code → Review codebook → Re-code → Compare cases → Contradictions/negative cases → Themes → Synthesis
Quantitative: Inspect → Define variables → Quality check → Plan → Calculate → Disaggregate → Visualise → Interpret → Assess uncertainty → Triangulate → Judge
Data quality: Profile → Detect → Diagnose → Prioritise → Investigate → Recommend → Human resolution → Track

Critical distinction: CALCULATION is not INTERPRETATION, and interpretation is not EVALUATIVE JUDGEMENT.

SOURCE → RETRIEVE → EXTRACT → VERIFY → COMPARE → SYNTHESISE

REMEMBER — How to do this
Start with a small set of approved sample documents. Explicitly identify which sources the agent may use. Preserve the source identifier with important claims. Do not assume that information in a ChatGPT conversation automatically becomes available to an API or Make.com workflow.

REMEMBER — How to implement this
State the boundary explicitly in the agent instructions. In an automated workflow, route consequential outputs to human approval. The agent should escalate uncertainty rather than hide it.

GOAL → CONTEXT → INSTRUCTIONS → RETRIEVAL → TOOLS → LOGIC → VALIDATION → REVIEW

REMEMBER — How to build it

1. Prototype the core task in ChatGPT or Claude.
2. Provide the M&E context and sample evidence.
3. Define the output structure.
4. Test manually.
5. Automate repeated steps in Make.com only after the prototype behaves acceptably.
6. Add retrieval, tools, routing and approval progressively.

Security: never put API keys, passwords or secrets inside this Elementor HTML.

REMEMBER — How to do this
Do not ask only “Write a report.” Define fields that expose evidence and uncertainty.

Finding | Evidence | Source | Interpretation | Contradictory Evidence | Evidence Gap | Uncertainty | Reviewer Action

OUTPUT → VALIDATE → PASS / FAIL → HUMAN REVIEW

REMEMBER — How to implement quality gates
In ChatGPT/Claude, put explicit checks into the instructions. In Make.com, use filters or routers after the AI step.

IF evidence source missing → STOP
IF contradiction detected → ESCALATE
IF required field missing → RETURN FOR REVISION
IF consequential judgement → HUMAN APPROVAL

Evidence traceability Calculation check Required fields Contradictions Uncertainty Human approval

REMEMBER — How to test
Create the benchmark before changing the agent. For every case record: Input → Expected behaviour → Actual output → Pass/Fail → Explanation. Aim for at least 8–10 cases initially.

Accuracy Groundedness Completeness Consistency Robustness Safety

REMEMBER — How to observe
For ChatGPT/Claude, record inputs, evidence and outputs. In Make.com, inspect the scenario execution history and each module's inputs, outputs, routing decisions and errors. Record tool calls, retrieval behaviour, latency and cost where available.

REMEMBER: Replace “AI made a mistake” with: condition → behaviour → impact → probable cause → intervention.

DATA → EVIDENCE → FINDING → INTERPRETATION → JUDGEMENT → RECOMMENDATION

REMEMBER — How to do this
Ask the agent to separate supporting evidence from contradictory evidence and to identify alternative explanations. The agent structures the evidence landscape; the evaluator determines the defensible judgement.

BUILD → TEST → OBSERVE → DIAGNOSE → IMPROVE → REGRESSION TEST

REMEMBER — How to improve
Change one important component at a time where possible. Then rerun the same benchmark. Otherwise you cannot reliably attribute the change in performance.

REMEMBER — How to decide
Base the decision on evidence from testing—not on how impressive the output looks. Consider accuracy, groundedness, completeness, consistency, robustness, safety, usability, efficiency and human-review burden as appropriate.