What M&E professionals need to know

Working well with AI agents is less about writing longer prompts and more about designing a controlled workflow. Define what the agent may decide, require evidence traceability, stop the workflow when risk thresholds are crossed, validate important outputs, and keep humans responsible for interpretation and external use.

Download PDF

What is an AI agent in an M&E context?

An AI assistant usually responds to one request. An AI agent may complete a sequence of tasks, consult files, use tools, create deliverables or continue until a defined objective is reached.

Human oversight is not a delay. It prevents speed from replacing accuracy, ethics, contextual understanding and professional judgment.

Lesson 1

Define decision boundaries before assigning work

A broad request such as “Review the programme data and prepare the donor report” describes the output but not the agent’s authority. It does not explain whether the agent may change calculations, remove records, interpret causality, use beneficiary data or communicate externally.

Classify decisions before the workflow begins:

Compact decision-boundary template

# me-agent-boundaries.yml
default_rule: stop_and_ask
allowed:
  - flag missing or duplicate data
  - organise evidence by evaluation question
  - draft text clearly labelled for review
approval_required:
  - change indicators, targets or methods
  - exclude data or assign performance ratings
  - use sensitive data or communicate externally
prohibited:
  - invent evidence, quotations or references
  - alter or delete original data
  - submit or publish final outputs
completion_check:
  - list sources, assumptions and limitations
  - separate evidence from interpretation
  - flag every item needing human validation

Key principle: A useful instruction tells the agent where to work and where it must stop.

Lesson 2

Review the output without assuming the process was correct

AI-generated reports, tables and summaries can appear polished while containing incorrect calculations, weak causal claims, missing perspectives, altered quotations or unsupported recommendations. Presentation quality is not evidence quality.

Apply a cold review

Examine the output as if you had not watched it being created. Ask:

• Does it answer the original evaluation or management question?
• Can every important claim be traced to a source?
• Did the agent change a formula, category, quotation or assumption?
• Were contradictory evidence and alternative explanations considered?
• Are uncertainty and evidence gaps clearly visible?

Do not confuse fluency with correctness. A confident answer can still contain weak evidence or invented detail.

Lesson 3

Plan capabilities, not only people or tools

The relevant question is not simply whether an AI tool can perform a task. Ask what combination of AI capability, human expertise and review is needed to complete the work responsibly.

Capability map

• AI-led, human-checked: identify missing indicator values.
• Human-led, AI-supported: develop evaluation questions.
• Human approval required: assign programme performance ratings.
• Human-only: obtain informed consent.

Better planning question: Under what conditions, with what evidence and under whose supervision should AI support this task?

Lesson 4

Establish tripwires before risks become problems

A tripwire is a predefined condition that causes the agent to stop, pause or request human review. It prevents one uncertain assumption or data problem from spreading through an entire workflow.

Useful M&E tripwires

• Conflicting indicator definitions are detected.
• Missing data exceed the approved threshold.
• Personally identifiable information appears.
• A conclusion cannot be linked to evidence.
• Material contradictions are found across sources.
• The sample is too small for the requested interpretation.
• A quotation cannot be verified.
• The workflow attempts an external submission or message.

Compact tripwire template

# me-agent-tripwires.yml
default_action: halt_and_request_review
tripwires:
  - condition: indicator_definition_missing
    action: stop_and_ask
  - condition: missing_data_above_threshold
    action: pause_for_data_quality_review
  - condition: personal_data_detected
    action: halt_immediately
  - condition: finding_has_no_source
    action: remove_and_flag
  - condition: material_contradiction_detected
    action: request_evaluator_review
  - condition: external_submission_requested
    action: require_explicit_approval

Configure thresholds to the context. The appropriate trigger depends on the method, decision risk, donor requirements and data-protection obligations.

Lesson 5

Take responsibility for the system, not only the output

Responsibility does not disappear when an AI agent contributes to an evaluation product. The professional or organisation still selects the tool, controls the inputs, defines the instructions, protects confidential data, validates the evidence and approves the final deliverable.

“The AI generated it” is not a quality-assurance process.

A trustworthy workflow should record

• the tool, purpose and approved sources;
• the prompts, boundaries and tripwires;
• the outputs produced and samples validated;
• the human reviewer and corrections made;
• the remaining limitations and final approval.

A nine-step workflow for managing an M&E agent

1. Define the purpose: state the precise evaluation or management need.
2. Approve the sources: list the files, datasets and periods the agent may use.
3. Specify outputs: name the expected tables, logs, draft sections or matrices.
4. Set boundaries: define allowed, approval-required and prohibited actions.
5. Add tripwires: identify conditions that require the workflow to stop.
6. Require traceability: connect important claims to the source file and location.
7. Validate a sample: compare representative outputs with original evidence.
8. Record human decisions: document accepted, changed, rejected and unresolved suggestions.
9. Approve external use separately: treat drafting and submission as different actions.

Worked example: reviewing a donor report

Weak instruction: “Check this report and improve it.”

Stronger instruction: Review the quarterly report against the approved results framework, indicator table and donor template. Identify inconsistencies, unsupported claims, missing evidence and calculation differences. Do not change indicator values or convert interpretations into facts. Present proposed revisions separately and stop if confidential data, contradictory definitions or unsupported outcome claims are detected.

Self-assessment checklist

Rate each practice from 1 to 5, then improve the two lowest-scoring areas.

• I define what the agent may and may not decide.
• I require original-source traceability.
• I validate important outputs using a representative sample.
• I use tripwires for sensitive or high-risk situations.
• I separate draft generation from external submission.
• I document human review and final approval.

Frequently asked questions

Final takeaway

AI agents can organise evidence, process information, identify inconsistencies and reduce repetitive work. They cannot replace accountability.

The most valuable AI skill in monitoring and evaluation may be judgment: knowing what to delegate, what to verify, what to protect and when to stop the agent from continuing.