EvalCommunity Academy Tutorial

How to Use AI Safely with Sensitive M&E Data

A risk-based workflow for confidential evaluations, monitoring data, qualitative evidence, and security-sensitive projects.

Artificial intelligence can support evaluators and M&E professionals with qualitative coding, evidence synthesis, report drafting, indicator analysis, knowledge products, and learning summaries.

However, many projects involve information that should not be entered into a public or unapproved AI chatbot. This is especially important when working with security-sector actors, vulnerable populations, confidential evaluations, complaints, protection data, politically sensitive findings, or unpublished institutional information.

Download PDF

Learning objectives

By the end of this tutorial, you will be able to:

Identify sensitive information in M&E documents and datasets.
Distinguish anonymization from pseudonymization.
Classify an AI-supported task by risk level.
Apply data minimization before using AI.
Create a limited AI-ready data packet.
Use synthetic data where appropriate.
Document and review an AI-assisted workflow.
Recognize when AI should not be used.

1. The scenario

Imagine that you are evaluating a project involving security-sector institutions. You have:

Fourteen key-informant interviews
Meeting notes from project partners
An internal monitoring report
A draft evaluation report
References to ranks, institutions, locations, incidents, and operational challenges

You would like AI assistance to identify themes, compare perspectives, organize findings, develop a report structure, and improve the final document.

However, anonymizing all the material may take longer than completing the analysis manually.

2. What makes M&E data sensitive?

Sensitive information is not limited to names, email addresses, or telephone numbers. M&E professionals must also consider contextual, institutional, operational, and political risks.

Personal data

Personal data is information relating to an identified or identifiable person. It can include:

Names and contact details
Identification or staff numbers
Photographs and signatures
Exact professional roles
Precise dates and locations
Combinations of details that indirectly identify someone

Example of indirect identification

“The only female district commander appointed in 2024.”

No name is included, but colleagues familiar with the institution may still identify the person.

Sensitive non-personal data

Information may create harm even when it does not identify an individual. Examples include:

Locations of supplies or programme activities
Internal institutional weaknesses
Unpublished evaluation findings
Allegations involving an organization
Partner-performance assessments
Community tensions or political relationships
Operational or security information

3. Anonymization and pseudonymization are different

Pseudonymization

Pseudonymization replaces identifying information with a code or label. A separate identification key may still connect that code to the original person, institution, or location.

Original information	Pseudonymized information
Colonel Maria D.	Respondent 04
Northern Security Directorate	Institution B
Municipality of Rivona	District 3

Pseudonymization reduces risk, but the information may still be linkable to its original source.

Anonymization

Anonymization aims to prevent people from being identified through the available information and other information that could reasonably be combined with it.

4. Start with the task, not the dataset

A common mistake is to prepare the complete dataset for AI before defining exactly what support is required. This creates unnecessary work and unnecessary risk.

Ask: What is the smallest amount of information the AI needs to complete this task?

Poorly defined task

“Analyze all the interviews and write the evaluation report.”

Better-defined task

“Develop a neutral structure for presenting three already validated evaluation findings.”

Other focused AI tasks

Compare anonymized recommendations and identify overlap.
Suggest headings for a summarized evidence matrix.
Rewrite a generalized paragraph in clearer evaluation language.
Review whether conclusions are supported by listed evidence references.
Generate questions for human evaluator review.

5. The Sensitive M&E Data Decision

Complete the following decision process before providing project information to an AI system.

Question 1

Is the tool and environment approved?

Check:

Organizational AI policy
Data-protection and security policies
Confidentiality clauses
Donor or government restrictions
Evaluation ethics requirements
Commitments made through informed consent

Approved: Continue to Question 2.

Unclear: Obtain guidance before proceeding.

Not approved: Do not upload project information.

Question 2

Does the task require real project data?

Many tasks can use:

Fictional examples
Blank templates
Generalized scenarios
Synthetic evidence matrices
Aggregated summaries
Approved public information

No real data required: Use synthetic or public information.

Some real data required: Continue to Question 3.

Substantial raw data required: Reconsider the task or environment.

Question 3

What could happen if the information were disclosed?

Consider possible harm to:

Participants and communities
Staff and implementing partners
Institutions and programme access
Evaluator independence and credibility
Organizational reputation
Operational security
Future relationships and participation

6. Classify the AI-supported task

Green: Lower-risk task

Creating a generic evaluation-report outline
Improving a blank template
Editing public, non-sensitive text
Analyzing synthetic data
Generating example indicators or questions

Approach: AI use may be reasonable, subject to organizational rules and human review.

Amber: Controlled task

Analyzing an anonymized evidence matrix
Organizing generalized interview themes
Reviewing pseudonymized monitoring information
Comparing non-identifiable stakeholder groups
Summarizing aggregated results

Proceed only when:

The tool is approved.
The information is minimized.
Identifiers have been reviewed.
The output will receive qualified human review.
The activity is documented.

Red: High-risk task

Uploading identifiable transcripts
Uploading protection or safeguarding case files
Sharing allegations about recognizable individuals
Providing precise security locations
Uploading restricted documents
Sharing passwords or access keys
Asking AI to make consequential decisions about individuals

Approach: Do not use a general-purpose external AI tool. Consider manual analysis, synthetic data, or an approved secure environment.

7. Build an AI-ready data packet

An AI-ready data packet is not a cleaned copy of the complete dataset. It is a deliberately limited package containing only the information required for one approved task.

The packet may contain

A precise task instruction
A generalized project description
An anonymized evidence table
Selected extracts
Evidence identifiers
Output instructions

The packet should exclude

Names and contact details
Precise addresses and locations
Unnecessary dates and titles
Identifiable quotations
Document metadata
Access credentials
The identification key

Example transformation

Original record

During an interview on 14 March 2026, Deputy Commander Arben K. from the West Regional Border Police stated that an incident near Kelmara checkpoint had not been documented because the liaison officer feared disciplinary action.

Minimized narrative

A regional security official reported that an operational incident had not been formally documented because a responsible staff member feared disciplinary consequences.

Evidence ID	Stakeholder group	Generalized finding	Confidence
INT-07	Regional security actor	Fear of disciplinary consequences may discourage incident reporting.	Moderate

8. Apply data minimization

Data minimization means limiting the information processed to what is genuinely necessary for the intended purpose.

Level	Action	Example
Records	Reduce the number of records.	Use five evidence entries instead of 30 transcripts.
Variables	Remove irrelevant fields.	Exclude contact details and unrelated demographics.
Detail	Reduce precision.	Replace an exact date with a project phase.
Output	Narrow the requested output.	Ask for themes rather than a final conclusion.

9. Separate sensitive analysis from AI support

AI does not need to participate in every stage. A safer workflow preserves human control over the most sensitive evidence and decisions.

Step 1: Collect information using approved systems.
Step 2: Store original records securely.
Step 3: Review and code sensitive information manually.
Step 4: Create a minimized evidence matrix.
Step 5: Validate the matrix against the sources.
Step 6: Provide only the approved packet to AI.
Step 7: Use AI to organize themes or test structures.
Step 8: Verify every AI-generated interpretation.
Step 9: Approve and take responsibility for final conclusions.

10. Use synthetic data for methodological tasks

Synthetic data is invented information designed to represent the structure of a situation without reproducing real cases.

Design a qualitative coding framework
Test an AI prompt
Develop a reporting template
Compare analytical approaches
Build an evidence-matrix structure
Train team members without exposing project information

Example

Create fictional extracts representing a positive institutional perspective, a critical community perspective, and a mixed implementation perspective. Ask AI to demonstrate a coding method using those examples.

Synthetic information must never be presented as real evidence or used to manufacture findings.

11. Practical prompts

Prompt 1: Develop a report structure without sharing evidence

You are supporting the organization of an evaluation report.
Do not generate findings or assume evidence that has not been provided.
Create a proposed structure for presenting:
1. Institutional capacity findings
2. Implementation constraints
3. Stakeholder trust
4. Sustainability risks
For each section, suggest:
- A neutral heading
- Questions the evaluator should answer
- Types of evidence required
- Possible limitations to disclose
Use professional evaluation language. Do not invent project facts.

Prompt 2: Analyze a minimized evidence table

Analyze only the information in the evidence table.
Requirements:
- Do not infer participant identities.
- Do not infer locations, institutions, or events.
- Do not add external facts.
- Cite the relevant evidence ID.
- Distinguish evidence from interpretation.
- Identify contradictory evidence.
- Flag findings supported by one source.
- State “insufficient evidence” where appropriate.
- Do not draft a final evaluation conclusion.
Output:
1. Emerging themes
2. Supporting evidence IDs
3. Contradictory evidence
4. Evidence gaps
5. Questions for evaluator review

12. Apply the efficiency test

Before investing significant time in data preparation, compare the cost, benefit, and risk of each approach.

How long will safe preparation take?
How long would manual completion take?
What additional value will AI provide?
What new risks and review requirements will AI introduce?

Option	Preparation	Analysis	Review	Total
Manual review	0 hours	3 hours	1 hour	4 hours
AI transcript review	5 hours	1 hour	2 hours	8 hours
AI evidence matrix	1.5 hours	45 minutes	1 hour	3 hours 15 minutes

13. Document the AI-assisted workflow

A short AI-use record improves accountability, traceability, and institutional learning.

Project: [Project name or internal code]

Task: [Specific AI-supported task]

Tool and environment: [Approved environment]

Date: [Date]

Data provided: [Minimized information]

Data excluded: [Original records and identifiers]

Preparation measures: [Anonymization or aggregation]

Output: [Themes, structure, questions, or revision]

Human review: [Reviewer role]

Corrections: [Summary]

Final decision: [Accepted, revised, or rejected]

Retention or deletion: [Required action]

14. Know when not to use AI

Manual analysis may be preferable when:

The dataset is small.
The context is highly specific.
The information is extremely sensitive.
Anonymization removes essential meaning.
The AI output requires extensive verification.
The tool is not approved.
Consent or contractual conditions do not permit processing.
The consequences of disclosure are serious.
The task depends heavily on cultural, political, or interpersonal nuance.

15. Practical exercise

Scenario

You have ten interviews with justice-sector actors containing:

Names and professional roles
Institutions and locations
Descriptions of incidents
Critical observations about management
Direct quotations

Your objective is to identify the main barriers to institutional learning.

Complete these steps

Write the precise AI-supported task.
Classify the information by sensitivity and analytical value.
Choose manual analysis, synthetic data, an evidence matrix, or a secure environment.
Create an AI-ready packet with no more than ten evidence entries.
Test whether a project colleague could identify a person, institution, location, or event.
Document whether the expected benefit is proportionate to the remaining risk.

16. Responsible AI checklist

☐ The AI task is clearly defined.

☐ The tool and environment are approved.

☐ Relevant policies and restrictions have been reviewed.

☐ The information has been classified by sensitivity.

☐ Real project data is genuinely necessary.

☐ Only the minimum necessary information is included.

☐ Direct identifiers have been removed.

☐ Indirect and contextual identifiers have been reviewed.

☐ Sensitive non-personal information has been assessed.

☐ The identification key is stored separately.

☐ Evidence references have been retained.

☐ A qualified person will review the output.

☐ The activity will be documented.

☐ Retention and deletion requirements are understood.

☐ Manual completion has been considered.

☐ The expected benefit justifies the remaining risk.

17. Frequently asked questions

Is replacing names with participant numbers enough?

Usually not. Roles, locations, dates, quotations, incidents, and demographic characteristics may still reveal identities.

Can I upload data when model training is disabled?

Disabling training does not automatically resolve authorization, retention, access, confidentiality, contractual, legal, or ethical requirements.

Is an enterprise AI account automatically appropriate?

No. Enterprise controls may reduce some risks, but the environment must still be approved and correctly configured.

Is AI useful for a small qualitative dataset?

Yes, but it may provide more value for coding frameworks, evidence matrices, alternative interpretations, or writing support than for processing all original records.

What if anonymization removes important context?

Reduce the scope of the AI task and conduct the contextual analysis manually.

Is deciding not to use AI a negative outcome?

No. A documented decision not to use AI can demonstrate professional judgment and responsible data management.