Data Ontology for AI Agents: A Practical Tutorial for M&E and Development Professionals

AI agents can now work across spreadsheets, reports, databases, surveys and other information sources. But access to more data does not automatically make an AI agent better at M&E.

The agent also needs to understand what the information means.

Consider an indicator called coverage rate. In one programme it might mean people receiving a service divided by the eligible population. In another, it might mean people reached divided by a programme target. Both calculations may be mathematically correct, but they answer different questions.

This tutorial explains how M&E, MEL/MEAL, development, humanitarian, research and evaluation professionals can create that structure without needing to become ontology engineers.

1. Why This Matters for M&E

Imagine asking an AI agent:

To answer responsibly, the agent may need to know:

• which programme the indicator belongs to;
• which outcome it measures;
• the exact indicator definition;
• the numerator and denominator;
• the target and baseline;
• the reporting period;
• the population included;
• the data source;
• the dataset and its grain;
• data-quality issues;
• relevant qualitative evidence;
• which findings and recommendations are connected to the indicator.

If these relationships are not clear, the agent may retrieve the right-looking number but misunderstand what it represents.

2. What Is a Data Ontology?

In practical terms:

A database tells you how information is stored. A data dictionary tells you what individual fields mean. An ontology goes further by representing concepts and relationships between them.

For example, a database may contain:

indicator_id
target
actual
reporting_period
location

A structured M&E information model can additionally represent:

Indicator
    measures → Outcome
Indicator
    has target → Target
Observation
    reports → Indicator
Observation
    belongs to → Reporting Period
Observation
    measured at → Location

That additional meaning becomes increasingly useful when AI agents need to work across multiple information sources.

3. You Do Not Need a Knowledge Graph to Start

The term ontology can sound highly technical. It does not mean that an M&E team needs to immediately build a sophisticated knowledge graph.

You can begin with a structured:

• indicator registry;
• data dictionary;
• results framework;
• evidence catalogue;
• metadata register;
• Excel or Google Sheet;
• JSON or YAML file.

The first objective is simply to make the meaning and relationships explicit enough for another professional, system or AI agent to use them consistently.

4. The Five Building Blocks

A useful starting structure is to classify information into five categories: entities, events, metrics, relationships and rules.

Entities

Things that exist in your programme or evaluation system.

• Programme
• Project
• Outcome
• Output
• Activity
• Indicator
• Target
• Participant
• Partner
• Location
• Dataset
• Evaluation
• Evaluation question

Events

Things that happen.

• Survey completed
• Training delivered
• Monitoring visit conducted
• Data submitted
• Indicator measured
• Interview conducted
• Evaluation completed
• Recommendation approved

Metrics

Values calculated or measured according to defined rules.

• Indicator achievement rate
• Coverage rate
• Completion rate
• Dropout rate
• Cost per participant
• Outcome score
• Data completeness rate

Relationships

Connections between concepts.

Programme
    contains → Project
Project
    contributes to → Outcome
Outcome
    measured by → Indicator
Indicator
    has → Target
Finding
    supported by → Evidence
Recommendation
    responds to → Finding

Rules

Conditions that should be applied consistently.

IF denominator = 0
THEN achievement rate = "Not calculable"
IF required data source is missing
THEN flag for data-quality review
IF finding has no evidence source
THEN mark as "Unverified"
IF reporting periods differ
THEN do not compare automatically

5. Start With Your Theory of Change or Results Framework

For development and evaluation professionals, one of the best places to start is an existing Theory of Change, logframe or results framework.

These already contain many of the relationships an AI system needs to understand.

You can make the relationships more explicit:

Programme
    ↓
Activity
    ↓ produces
Output
    ↓ contributes to
Outcome
    ↓ measured by
Indicator
    ↓ compared with
Target
    ↓ observed during
Reporting Period

This creates a useful bridge between traditional M&E practice and AI-assisted workflows.

6. Make Indicator Definitions AI-Ready

A label such as “coverage rate” is not enough.

A stronger indicator record might look like this:

Indicator:
Service coverage rate
Definition:
Percentage of eligible participants who received
the defined service during the reporting period.
Numerator:
Number of eligible participants receiving the service.
Denominator:
Number of eligible participants.
Unit:
Percentage.
Reporting frequency:
Quarterly.
Disaggregation:
Sex, age group and location.
Data source:
Programme monitoring database.
Owner:
M&E Manager.
Interpretation rule:
Missing observations must not automatically
be treated as zero.

The more important an indicator is for decisions, the more useful it is to document its definition, calculation, grain, source and ownership explicitly.

7. Define the Grain of Your Data

Grain describes what one observation represents.

This matters when an AI agent combines datasets. A beneficiary-level dataset and a programme-level dataset should not be joined casually.

8. Connect Evidence to Findings

Evaluation findings should not exist as isolated pieces of text.

A useful evidence structure is:

Evaluation Question
        ↓
Finding
        ↓
Evidence
        ↓
Source
        ↓
Dataset / Document
        ↓
Data Collection Method

For example:

This distinction between evidence and interpretation is particularly important when AI is used to synthesize evaluation findings.

9. Add Provenance

Provenance means being able to understand where information came from and what happened to it.

For an M&E finding, provenance might include:

• original source;
• data collection date;
• dataset version;
• cleaning or transformation performed;
• analysis method;
• analyst or responsible team;
• approval status;
• report or publication where the finding was used.

The W3C PROV-O recommendation provides a formal ontology for representing provenance and relationships involving entities, activities and agents. See the W3C PROV-O specification.

Most M&E teams do not need to implement PROV-O directly. The practical lesson is simpler: make important results traceable to their sources.

10. Adapt the Bronze–Silver–Golden Model

A useful data architecture separates raw information from cleaned information and governed meaning.

Bronze — Source Data

• KoboToolbox or ODK exports
• partner spreadsheets
• survey datasets
• administrative records
• monitoring databases
• interview transcripts
• programme reports

Silver — Cleaned and Validated Data

• standardised variables;
• validated dates;
• consistent codes;
• duplicate checks;
• missing-value checks;
• data-quality flags;
• documented transformations.

Golden — Governed M&E Meaning

• indicator definitions;
• approved calculations;
• targets and baselines;
• reporting periods;
• outcome relationships;
• evidence relationships;
• data sources;
• provenance;
• owners;
• version information.

For AI agents, the golden layer is particularly valuable because it provides the definitions and relationships the agent should use when interpreting the underlying data.

11. What an AI Agent Can Do With This Structure

12. Example: An Agent Investigates an Indicator

Suppose an M&E manager asks:

A structured agent could work through:

1. Identify the exact definition of Indicator 4.
2. Confirm numerator and denominator.
3. Check that the two observations have the same grain.
4. Confirm the reporting periods.
5. Check data-quality flags.
6. Compare relevant disaggregations.
7. Review linked programme reports.
8. Review relevant qualitative evidence.
9. Identify evidence-supported explanations.
10. Separate evidence from interpretation.
11. Provide source references.
12. Flag unresolved questions for human review.

This is a much stronger workflow than simply uploading a spreadsheet and asking an AI model to “find out what happened.”

13. Build an AI-Ready Indicator Registry

An indicator registry is one of the most practical places for an M&E team to start.

Consider documenting:

Indicator ID
Indicator name
Definition
Result level
Linked outcome
Numerator
Denominator
Unit
Grain
Reporting frequency
Baseline
Target
Data source
Collection method
Disaggregation
Calculation
Data-quality rules
Indicator owner
Last reviewed
Version
Status

This can become a trusted reference for both human analysts and AI systems.

14. Add Rules Before Adding More Autonomy

An AI agent should not have to discover your M&E rules by trial and error.

Rule 1:
Never treat missing data as zero.
Rule 2:
Never compare indicators with different definitions.
Rule 3:
Do not calculate a percentage when the denominator
is missing or zero.
Rule 4:
Do not infer causality from descriptive monitoring data.
Rule 5:
Every evaluation finding must identify its evidence source.
Rule 6:
Flag conflicting sources instead of silently choosing one.
Rule 7:
Require human review before external publication.

15. Ontology Helps With Data Quality Too

Data quality is not only about missing values or duplicate records.

Example 1 — Technical data-quality problem

A partner did not submit the required dataset.

Example 2 — Definition problem

Two partners submitted complete datasets but calculated the same indicator using different denominators.

The second problem cannot be solved simply by checking whether fields are complete. The underlying definition needs to be governed.

This is one reason structured definitions become increasingly important when AI systems work across multiple datasets and organizations.

16. Build an Evidence Structure for Evaluation

The same approach can be extended from monitoring data to evaluation evidence.

Evaluation
    ↓
has → Evaluation Question
Evaluation Question
    ↓
answered by → Finding
Finding
    ↓
supported by → Evidence
Evidence
    ↓
derived from → Source
Source
    ↓
collected through → Method
Finding
    ↓
leads to → Recommendation
Recommendation
    ↓
assigned to → Responsible Actor

This can help an AI system distinguish a source document from a finding, a finding from an interpretation, and a recommendation from the evidence supporting it.

17. Connect This With FAIR Data Principles

The FAIR principles focus on making data and metadata Findable, Accessible, Interoperable and Reusable.

GO FAIR specifically includes formal and shared knowledge-representation languages, FAIR vocabularies, qualified references, rich metadata, provenance and domain-relevant community standards. See the GO FAIR Principles.

For M&E, ask:

• Can another team understand this indicator without asking its creator?
• Can the dataset be connected to its metadata?
• Can an AI system identify the source of a result?
• Are important terms defined consistently?
• Can findings be traced to evidence?
• Can another system reuse the information?

You do not need to implement every FAIR principle before using AI. The useful starting point is to improve the quality, consistency and traceability of the information your AI workflows depend on.

18. A 60-Minute Exercise for M&E Professionals

0–10 minutes: Choose one workflow

Choose indicator reporting, data-quality review, partner reporting, evidence synthesis, evaluation reporting or donor reporting.

10–20 minutes: List the concepts

Write down the programmes, outcomes, indicators, targets, datasets, reporting periods, evidence sources and other important concepts.

20–30 minutes: Define relationships

Use simple relationships such as measures, has target, reported in, supported by and derived from.

30–40 minutes: Document rules

Write down the conditions your team already applies when calculating and interpreting results.

40–50 minutes: Add provenance

Record where important results and findings came from.

50–60 minutes: Test with AI

Give the structured information to an AI assistant and ask it a real M&E question. Compare the result with what happens when you provide only the raw data.

19. A Prompt for Testing Your M&E Information Model

Try this with an AI assistant:

You are supporting an M&E team.
Use the definitions and relationships provided below.
Do not invent indicator definitions.
Do not treat missing data as zero.
Do not infer causality from descriptive monitoring data.
Do not silently resolve conflicting definitions.
For every important conclusion:
1. Identify the indicator or concept used.
2. State the reporting period.
3. Identify the source.
4. Distinguish evidence from interpretation.
5. Flag uncertainty or conflicting information.
If the information is insufficient, state what is missing.
M&E information model:
[PASTE YOUR DEFINITIONS HERE]
Task:
[INSERT YOUR M&E QUESTION HERE]

The objective is not to make the prompt unnecessarily long. It is to give the AI the structured information it needs to reason responsibly.

20. The AI-Readiness Checklist

• Important concepts have clear definitions.
• Indicators have documented calculations.
• Numerators and denominators are defined.
• Data grain is documented.
• Reporting periods are explicit.
• Important relationships are documented.
• Data sources are identifiable.
• Important findings have evidence references.
• Rules are explicit.
• Important definitions have owners.
• Definitions can be versioned.
• Conflicting information can be identified.
• Human review remains in place for consequential decisions.

21. Practical Applications Across the M&E Cycle

22. Do Not Overbuild It

There is a temptation to create a huge ontology covering every programme, variable, document and relationship.

That is usually not the best starting point.

Start with concepts that are:

• used frequently;
• important for decision-making;
• often misunderstood;
• shared across teams;
• used by multiple datasets or systems;
• likely to be used by future AI agents.

For many teams, that might mean starting with 10–20 critical indicators, major outcomes, important datasets, evaluation questions and core evidence sources.

23. The Bigger Lesson for AI-Assisted Development Work

AI readiness is often framed as a technology question:

For many M&E and development teams, an equally important question is:

This becomes particularly important when information is spread across implementing partners, country offices, donors, government counterparts, programme teams, M&E specialists and evaluators.

Shared definitions and relationships can make collaboration easier for people and create a stronger foundation for AI-assisted workflows.

24. Your Practical Next Step

Choose one recurring M&E workflow this week.

Do not start by building an AI agent.

Start by documenting:

1. What are the important concepts?
2. What does each concept mean?
3. How are the concepts related?
4. How are the indicators calculated?
5. What is the data grain?
6. Where does each result come from?
7. What rules should be applied?
8. Who owns each definition?
9. Where does human judgment remain necessary?

Once these questions are answered, you have a much stronger foundation for designing reliable AI-assisted workflows.

25. Explore More EvalCommunity Resources

If you want to take this further, EvalCommunity already has resources that connect AI with practical M&E work.

• AI Ecosystem for Evaluators — explore AI applications across programme design, data collection, analysis, reporting and learning.
• M&E Framework Library — access frameworks, templates and tools that can provide the structured foundation for AI-assisted workflows.
• Principles for AI Use in Evaluation — review principles around human oversight, data quality, transparency, accountability and appropriate AI use.
• AI Tool Assessment & Implementation — assess AI tools before introducing them into professional M&E workflows.

Learn AI for M&E and Build Practical AI Agents

Understanding data, definitions and relationships is one part of becoming effective with AI. The next step is learning how to apply AI across real evaluation workflows and build practical agents that can support recurring M&E tasks.

The AI for M&E Professional Bundle from EvalCommunity Academy combines the AI in Monitoring & Evaluation Certificate and the AI Agents for Evaluators Certificate. The current bundle price is USD 349, compared with USD 490 when purchased separately. The Academy describes the courses as covering practical AI use in evaluation, evidence analysis, reporting, data quality, indicator tracking, workflow automation and no-code AI agents.

Sources and Further Reading

• W3C — PROV-O: The PROV Ontology
• GO FAIR — FAIR Principles
• EvalCommunity — AI Ecosystem for Evaluators
• EvalCommunity — M&E Framework Library
• EvalCommunity Academy — Principles for AI Use in Evaluation
• EvalCommunity Academy — AI for M&E Professional Bundle