Skip to main content
Free course

Working Smarter with AI: A Practical Guide for Development Professionals

Artificial intelligence is rapidly transforming how international development organisations design programmes, analyse evidence, manage projects and measure results. This practical beginner-level course introduces development professionals to responsible and effective use of AI across the project cycle. Learners explore how AI can support needs assessments, political economy analysis, proposal development, project design, monitoring and evaluation, knowledge management, report writing, stakeholder engagement and organisational learning. The course also examines risks such as bias, misinformation, data privacy, intellectual property concerns and weak governance. Throughout the course, learners are encouraged to use AI as a productivity and decision-support tool while maintaining human judgement, accountability and professional standards.

Lesson 1

Understanding the Role of Artificial Intelligence in Modern Development Programmes

1.1 What AI Means for Development Practice
Module 1 · Lesson 1.1

What AI Means for Development Practice

Artificial intelligence is already changing how development organisations analyse information, design programmes, communicate with stakeholders and manage knowledge. Its value, however, depends less on the technology itself than on how responsibly and professionally it is used.

Learning objectives

By the end of this lesson, you should be able to:

  • explain artificial intelligence in practical terms relevant to development work;
  • identify routine tasks where AI can improve speed, structure or analysis;
  • distinguish between appropriate AI support and decisions requiring human judgement; and
  • apply a simple process for using and verifying AI-generated outputs.

1. What is artificial intelligence?

Artificial intelligence, or AI, refers to digital systems that can perform tasks commonly associated with human intelligence. These tasks may include generating text, summarising documents, translating content, classifying information, recognising patterns, comparing options and supporting analysis.

A practical definition

In development practice, AI is best understood as a digital assistant for information-intensive work. It can help a professional process material more quickly, organise ideas and produce a first draft. It does not understand communities, institutions or political realities in the same way that experienced practitioners do.

Many widely used AI tools are based on large language models. These systems generate responses by identifying patterns in large volumes of existing text. They can produce clear and convincing language, but they do not independently confirm whether the information is accurate, current, politically appropriate or ethically acceptable.

2. What can AI support in development work?

Development professionals often work under time pressure and must process large amounts of information from assessments, partner reports, donor guidance, evaluations, meeting notes, policies and stakeholder consultations. AI can assist with many of these tasks.

Programme management

A project manager receives monthly reports from eight implementing partners. AI can help produce an initial summary of achievements, delays, emerging risks and issues requiring management attention.

Human role: confirm the evidence, assess political and operational implications, and decide what action to take.

Monitoring, Evaluation and Learning

An MEL specialist has 60 interview notes from beneficiaries. AI can help group recurring themes, identify frequently mentioned barriers and suggest categories for deeper qualitative analysis.

Human role: verify the coding, protect personal data, interpret context and avoid treating frequency as proof of importance.

Proposal development

A proposal team is responding to a donor call with a short deadline. AI can help compare the call requirements against a draft concept note and identify missing sections, unclear assumptions or weak links between activities and expected results.

Human role: ensure technical accuracy, donor compliance, realistic costing and ownership by local partners.

Localisation and partner support

A local civil society organisation has strong field knowledge but limited experience with international donor formats. AI can help reorganise its ideas into a clearer problem statement, activity plan or results framework.

Human role: preserve the organisation’s voice, avoid imposing external assumptions and ensure that the final product reflects genuine local priorities.

Climate and economic development

An SME support programme is reviewing business plans from farmers and small enterprises. AI can help identify whether each plan discusses climate hazards, supply-chain risks, energy costs or possible adaptation measures.

Human role: verify the risks against local evidence and avoid rejecting applicants solely on the basis of automated screening.

Knowledge management

A governance programme has accumulated years of reports, evaluations and policy papers. AI can help develop summaries, compare recommendations and locate repeated lessons across documents.

Human role: check source quality, distinguish current guidance from outdated material and retain institutional accountability.

3. AI as an assistant, not an authority

AI can improve productivity, but it should not be treated as an independent expert or decision-maker. Development decisions affect people, institutions and public resources. They require professional accountability, local knowledge, ethical judgement and, in many cases, formal approval.

Important principle

A well-written AI response is not necessarily a correct response. AI can invent facts, misinterpret donor requirements, overlook political sensitivities, reproduce bias or recommend actions that are inappropriate in the local context.

AI should therefore be used mainly to support:

  • first drafts and outlines;
  • summaries and comparisons;
  • idea generation and alternative formulations;
  • initial classification of non-sensitive information;
  • quality checks against a defined checklist; and
  • preparation of questions for further investigation.

AI should not independently determine:

  • which community, partner or applicant receives assistance;
  • whether a person is vulnerable, credible or eligible;
  • the final interpretation of evaluation findings;
  • politically sensitive recommendations;
  • legal, safeguarding or fiduciary decisions; or
  • whether confidential personal information may be shared.

4. A responsible five-step workflow

  1. Define the task. Be clear about what you want AI to support. A specific task, such as “summarise the three main implementation risks”, is safer and more useful than a vague request such as “analyse this project”.
  2. Protect sensitive information. Remove names, personal data, confidential negotiations, security information and restricted organisational documents unless the tool has been formally approved for such use.
  3. Provide context and criteria. Explain the country, sector, target group, donor requirement and intended audience. Ask the tool to use only the information provided when accuracy is essential.
  4. Review and verify the output. Check facts, figures, quotations, references, assumptions and recommendations against reliable sources and professional experience.
  5. Revise and take responsibility. Adapt the output to the local context, organisational standards and intended purpose. The accountable professional—not the AI tool— remains responsible for the final product.

5. Real-life practical exercise

Exercise: Improve a project reporting task

You are supporting a livelihoods project implemented by three local partners. Each partner has submitted a monthly narrative report. The reports are long, use different formats and contain a mixture of achievements, challenges and activity descriptions. Your manager needs a one-page briefing before a donor meeting.

Your task

  1. Select one recent report, meeting note or other non-confidential document from your work.
  2. Remove names, personal data and any information that should not be entered into an external AI tool.
  3. Ask AI to identify:
    • three main achievements;
    • three implementation challenges;
    • two issues requiring management attention; and
    • any claims that are not supported by evidence in the text.
  4. Compare the AI output with the original document.
  5. Correct errors, restore missing context and write the final briefing in your own professional voice.

Suggested prompt

Review the text below as a development project reporting assistant. Using only the information provided, prepare a concise briefing with: (1) three main achievements, (2) three implementation challenges, (3) two issues requiring management attention, and (4) any statements that appear unsupported by evidence in the text. Do not invent facts. Clearly mark any uncertainty.

Verification checklist

  • Did the AI accurately represent the document?
  • Did it confuse completed activities with planned activities?
  • Did it overlook an important political, gender, safeguarding or operational issue?
  • Did it invent a result, number, explanation or recommendation?
  • Would the final briefing be suitable for sharing with the donor?

Apply the lesson to your own role

Identify three routine tasks that involve reading, drafting, summarising, comparing or organising information. Complete the table below.

Routine task How AI could help Human verification needed Information that must not be shared
Example: Review partner reports Summarise progress and identify recurring delays Check evidence and interpret operational causes Names, personal data and confidential partner information
       
       
       

Keep your examples practical. Focus on tasks where AI can support speed, structure or initial analysis without replacing professional responsibility.

Lesson summary

  • AI can support information-intensive tasks across development programmes.
  • Its strongest uses include drafting, summarising, comparing, organising and identifying possible patterns.
  • AI outputs may be inaccurate, biased or poorly adapted to local context.
  • Confidential and personal information must be protected.
  • Development professionals remain accountable for verification, judgement and final decisions.

Resources for this lesson

Web link

Lesson content: 1.1 What AI Means for Development Practice

Open resource
Lesson 2

Applying Artificial Intelligence from Project Identification to Learning and Knowledge Management

1.2 Opportunities Across the Project Cycle
Module 1 · Lesson 1.2

Opportunities Across the Project Cycle

International development projects generate large volumes of information at every stage. Artificial intelligence can help teams analyse, structure, draft, compare and synthesise this information more efficiently. Its value is greatest when it supports clearly defined tasks while professionals retain responsibility for evidence, context, ethics and decisions.

Learning objectives

By the end of this lesson, you should be able to:

  • identify practical AI opportunities across the development project cycle;
  • match AI-supported tasks to different project stages;
  • distinguish productivity support from decisions that must remain human-led;
  • recognise the main risks associated with AI use at each stage; and
  • select one realistic, low-risk AI application for your own work.

1. AI across the development project cycle

Donors and organisations use different terminology, but most development initiatives move through a similar cycle: analysis, design, mobilisation, implementation, monitoring, evaluation, learning and reporting. These stages are connected rather than strictly linear. Monitoring may lead to redesign, evaluation may shape a new programme, and stakeholder feedback may require changes during implementation.

Where AI creates the most value

AI is particularly useful where teams must process large amounts of text, compare documents, organise information, create first drafts or identify possible patterns. It is less suitable for decisions involving rights, eligibility, safeguarding, political judgement, legal interpretation or allocation of resources.

1

Analysis

Understand context and needs

2

Design

Define results and approach

3

Mobilisation

Prepare people and systems

4

Implementation

Deliver and adapt activities

5

Monitoring

Track progress and risks

6

Evaluation

Assess performance and results

7

Learning

Capture and apply lessons

8

Reporting

Communicate evidence and accountability

2. Opportunities by project stage

1. Context and needs analysis

AI can support:

  • summaries of policies, strategies and evaluations;
  • comparison of country, donor and sector priorities;
  • organisation of stakeholder information;
  • initial coding of consultation notes;
  • identification of recurring themes and evidence gaps; and
  • preparation of questions for further research.
Example: A governance team has 18 interview notes, four policy papers and two evaluations. AI produces an initial thematic map showing repeated concerns about institutional coordination, local capacity and access to services.
Human role: validate the themes, examine power relations, identify missing voices and assess whether the evidence reflects national and subnational realities.

2. Project design

AI can assist with:

  • problem and objective analysis;
  • theories of change and results chains;
  • draft outputs, outcomes and assumptions;
  • initial risk registers;
  • possible indicator formulations; and
  • consistency checks between problems, activities and results.
Example: A climate adaptation team provides evidence from a vulnerability assessment. AI helps organise the findings into a draft results chain linking water insecurity, municipal planning and resilient service delivery.
Human role: confirm causality, test assumptions with stakeholders and approve indicators only after checking feasibility, data sources and attribution.

3. Proposal development and resource mobilisation

AI can support:

  • donor opportunity research;
  • call-for-proposal summaries;
  • compliance and eligibility checklists;
  • proposal outlines and section planning;
  • review against evaluation criteria;
  • editing for clarity and consistency; and
  • checking alignment between activities, results and budget narrative.
Example: A consortium preparing an EU proposal asks AI to compare the draft technical narrative against the call requirements and identify missing information on target groups, sustainability and risk management.
Human role: confirm donor interpretation, technical quality, cost realism, partner commitments and accuracy of all claims.

4. Mobilisation and start-up

AI can assist with:

  • staff induction materials;
  • project handbooks and frequently asked questions;
  • workplan templates and task breakdowns;
  • stakeholder and contact databases;
  • meeting schedules and briefing notes;
  • communication plans; and
  • initial training materials.
Example: A new regional programme must onboard staff in five countries. AI helps transform the project document, policies and operating procedures into a concise induction package.
Human role: ensure that procedures are current, roles are correctly assigned and national legal or administrative requirements are reflected.

5. Project implementation

AI can support:

  • meeting notes and action trackers;
  • partner communication and routine correspondence;
  • field visit report drafting;
  • training and presentation materials;
  • risk and issue log updates;
  • translation and plain-language adaptation;
  • communication products and newsletters; and
  • preparation of options for adaptive management.
Example: After a two-day workshop, AI converts an approved transcript into draft minutes, an action matrix, follow-up messages and a short communication brief.
Human role: verify decisions, responsibilities and deadlines, and ensure that politically sensitive discussions are represented accurately.

6. Monitoring

AI can assist with:

  • cleaning non-sensitive text responses;
  • initial coding of open-ended survey answers;
  • summarising monthly monitoring reports;
  • identifying repeated implementation concerns;
  • drafting indicator narratives;
  • reviewing data-quality checklists; and
  • preparing questions for field verification.
Example: A livelihoods programme receives 600 open-ended beneficiary comments. AI groups them into themes such as market access, transport costs, training relevance and payment delays.
Human role: review the coding, protect personal data, assess representativeness and avoid treating frequency alone as evidence of significance.

7. Evaluation

AI can support:

  • document review matrices;
  • initial qualitative coding;
  • comparison of evidence across sources;
  • clustering of recommendations;
  • identification of evidence gaps;
  • draft summaries of findings; and
  • editing of evaluation products.
Example: An evaluation team uses AI to organise evidence from interviews, monitoring reports and previous reviews under the OECD-DAC criteria.
Human role: determine findings, assess contribution and causality, resolve conflicting evidence and approve all conclusions and recommendations.

8. Learning, reporting and knowledge management

AI can assist with:

  • after-action review summaries;
  • lessons learned and good-practice notes;
  • case studies and success stories;
  • knowledge repository tagging;
  • quarterly and annual report outlines;
  • executive summaries;
  • briefing notes and presentations; and
  • adaptation of technical findings for different audiences.
Example: A regional programme uses AI to compare annual reports from six countries and identify repeated lessons on local ownership, procurement delays and staff turnover.
Human role: verify evidence, distinguish lessons from opinions and ensure that reporting accurately reflects both achievements and limitations.

3. Choosing suitable AI tasks

Not every project task should be automated or AI-assisted. A useful starting point is to select tasks that are repetitive, information-heavy, easy to verify and unlikely to expose sensitive data.

  1. Start with a defined task. Choose one activity, such as summarising approved partner reports or checking a proposal against a donor checklist.
  2. Assess information sensitivity. Remove personal, confidential, security-related or commercially restricted information.
  3. Specify the expected output. State the audience, format, length, criteria and evidence base.
  4. Check the output. Compare the result with the original evidence and identify omissions, errors or invented information.
  5. Measure the benefit. Determine whether AI saved time, improved quality or simply created additional review work.

Good practice: begin with low-risk applications

Organisations can build confidence by starting with activities such as meeting summaries, document comparisons, outline development, plain-language editing and preparation of internal checklists. More sensitive applications should only be introduced after appropriate governance, data protection and quality assurance arrangements are in place.

4. Human accountability throughout the cycle

AI must not become the decision-maker

AI may help prepare information, but it should not independently determine who receives funding, which community is prioritised, whether a person is vulnerable, whether evidence is credible, what policy should be adopted or whether a programme has succeeded.

Across all project stages, professionals remain responsible for:

  • accuracy of facts, figures, references and quotations;
  • protection of personal and confidential information;
  • technical quality and methodological soundness;
  • political, cultural and institutional appropriateness;
  • ethical, safeguarding and inclusion considerations;
  • compliance with organisational and donor requirements; and
  • final decisions and formal approvals.

5. Practical exercise

Exercise: Identify AI opportunities in a complex programme

You have joined a €12 million EU-funded climate resilience programme. The programme works with five implementing partners, three ministries and several municipalities. It has 120 indicators, quarterly reporting requirements, annual reviews, regular donor missions and frequent stakeholder workshops.

Your task

  1. Identify five project activities where AI could save significant staff time.
  2. For each activity, describe the specific AI-supported task.
  3. State the expected benefit.
  4. Identify one risk that must be managed.
  5. Explain what a person must verify before the output is used.
Project stage or task How AI could support Expected benefit Risk to manage Human review required
Example: Quarterly partner reporting Summarise reports and identify recurring delays Faster management briefing Important context may be omitted Check evidence, explanations and partner perspectives
     
     
     
     
     

Optional AI prompt

I manage a development programme with multiple partners, government stakeholders, quarterly reporting, donor missions and a large monitoring framework. Suggest five low-risk, practical tasks where AI could save staff time. For each task, identify: the project-cycle stage, the expected benefit, the main risk and the human verification required. Do not recommend automated decisions on funding, eligibility, safeguarding or beneficiary selection.

Reflection

Consider your current role or organisation:

  • Which project-cycle stage consumes the most staff time?
  • Which task is repetitive, information-heavy and easy to verify?
  • What information would need to be removed before using an AI tool?
  • What would success look like: faster delivery, clearer outputs or better use of evidence?

Select one modest application that can be tested safely. A useful pilot should have a clear task, a responsible reviewer and a simple way to measure whether AI added value.

Lesson summary

  • AI can support every stage of the development project cycle.
  • Its strongest applications involve information processing, drafting, comparison and synthesis.
  • Different stages require different safeguards, verification and professional expertise.
  • AI should support project teams, not make decisions on rights, resources, eligibility or policy.
  • Organisations should begin with low-risk, clearly defined tasks and expand only when benefits are demonstrated.

Resources for this lesson

Web link

Lesson content: 1.2 Opportunities Across the Project Cycle

Open resource
Lesson 3

1.3 Limits, Risks and Human Accountability

1.3 Limits, Risks and Human Accountability
Module 1 · Lesson 1.3

Limits, Risks and Human Accountability

Using Artificial Intelligence Responsibly in Development Programmes

Artificial intelligence can improve speed and productivity, but it also introduces serious risks. Development organisations work with public resources, vulnerable groups, sensitive information and politically complex institutions. AI use must therefore be governed by professional judgement, clear safeguards and identifiable human accountability.

Learning objectives

By the end of this lesson, you should be able to:

  • identify the main limitations and risks of AI in development practice;
  • recognise when AI-generated outputs require additional scrutiny;
  • distinguish between information that may and may not be entered into public AI tools;
  • explain what human accountability means in practical terms; and
  • apply a simple risk review before using AI-supported work.

1. Why AI outputs require caution

AI can produce fluent, confident and professional-looking responses. This creates a particular risk: weak or incorrect content may appear more credible than it really is. A polished output can still contain false references, invented data, weak reasoning or assumptions that do not fit the local context.

Key principle

AI should be treated as a tool for preparing, organising or testing ideas—not as an authoritative source. The more important the decision, the stronger the human review must be.

Inaccurate information

AI may invent sources, dates, quotations, statistics or legal provisions. It may also combine correct information in a misleading way.

Example: An AI-generated policy brief cites a national strategy that does not exist and attributes a target to the wrong ministry.

Weak contextual understanding

AI may misunderstand local institutions, informal power structures, conflict dynamics, administrative practice or cultural norms.

Example: AI recommends a stakeholder consultation process that excludes the traditional authority actually responsible for local land decisions.

Bias and exclusion

AI may reflect bias present in the data on which it was trained or in the material provided by the user. This is especially important when working on gender, disability, ethnicity, migration, poverty, conflict and marginalised communities.

Example: A draft beneficiary profile describes women only as vulnerable recipients and overlooks their role as producers, entrepreneurs and community leaders.

Loss of nuance

Summaries may remove disagreement, uncertainty, minority views or politically sensitive detail. Important differences can disappear when complex evidence is compressed.

Example: A consultation summary reports broad support for reform but omits strong opposition from one affected community.

Over-reliance

Teams may stop checking original documents, reduce consultation or accept AI-generated analysis without sufficient challenge.

Example: A proposal team copies AI-generated indicators without checking whether data are available or whether the project can credibly influence the result.

Unclear accountability

Staff may wrongly assume that responsibility is reduced because AI produced the first draft. In practice, the organisation remains responsible for everything it submits, publishes or implements.

Example: A donor receives an inaccurate report. “The AI wrote it” is not an acceptable explanation.

2. Data protection and confidentiality

Development organisations often hold sensitive information about communities, beneficiaries, staff, partners, governments and donors. Entering this information into a public AI tool may expose it to unauthorised processing, storage or reuse.

Do not assume that an AI tool is private

Personal data, confidential proposal content, unpublished evaluation findings, security-sensitive information, partner records and restricted government documents should not be entered into public AI tools unless organisational policy explicitly permits it and appropriate safeguards are in place.

Type of information Public AI tool? Reason
Published donor guidance Generally acceptable Already public, subject to normal accuracy checks.
Approved public project brochure Generally acceptable Public content, provided no restricted details are added.
Beneficiary names and contact details Do not enter Contains personal data and may expose vulnerable individuals.
Unpublished evaluation findings Do not enter without approval May be confidential, politically sensitive or subject to contractual restrictions.
Draft proposal and partner budget Do not enter without approval May contain commercially sensitive information and intellectual property.
Security incident report Do not enter May endanger staff, partners or operations.
Anonymised, non-sensitive training example Usually acceptable Lower risk when all identifying and restricted details are removed.

Good practice: use the minimum necessary information

Even when AI use is permitted, provide only the information required for the task. Remove names, exact locations, personal identifiers, confidential figures and unnecessary background detail. Where possible, use fictionalised or anonymised examples.

3. Bias, representation and do-no-harm

AI can reproduce stereotypes or favour perspectives that are more visible in available data. Groups with limited digital presence, less documentation or lower representation in formal sources may be ignored or misrepresented.

Gender and livelihoods

An AI-generated rural livelihoods analysis may focus on male landowners because formal records underrepresent women’s unpaid labour, informal businesses and shared access to land.

Disability inclusion

A draft training plan may assume that online delivery is automatically inclusive, overlooking accessibility barriers, assistive technology needs or low digital literacy.

Conflict sensitivity

A stakeholder map may use formal institutional categories and fail to identify actors whose influence is informal, contested or politically sensitive.

Migration and vulnerability

AI may use broad categories such as “migrants” or “refugees” without recognising differences in legal status, age, gender, disability, documentation or protection needs.

Development professionals must therefore ask whose perspective is represented, whose voice is missing, what assumptions are being made and whether the output could create harm if acted upon.

4. What human accountability means

Human accountability means that a named person or team remains responsible for the quality, accuracy, ethical appropriateness and consequences of AI-supported work. Responsibility cannot be transferred to a software tool.

  1. A person defines the task. The user decides what AI is asked to do and what information is provided.
  2. A person checks the evidence. Facts, figures, sources, legal references and quotations are verified.
  3. A person applies context. Political, cultural, institutional and operational realities are considered.
  4. A person assesses ethics and risk. Data protection, safeguarding, inclusion and do-no-harm principles are reviewed.
  5. A person approves the final output. A responsible professional signs off before the content is submitted, published or used for a decision.

Decisions that should remain human-led

  • selection of beneficiaries, partners or grant recipients;
  • eligibility, vulnerability or safeguarding assessments;
  • legal and policy recommendations;
  • allocation of public or donor resources;
  • evaluation conclusions and judgements on performance;
  • security decisions and conflict-sensitive actions; and
  • formal commitments made to communities, governments or funders.

5. A practical AI risk review

Before using AI-generated content, ask five questions:

  1. Accuracy: Can every important fact, number and reference be verified?
  2. Context: Does the output reflect local institutions, politics, culture and implementation realities?
  3. Bias: Are any groups stereotyped, excluded or underrepresented?
  4. Confidentiality: Was any personal, restricted or security-sensitive information exposed?
  5. Accountability: Is a named person responsible for approving and using the final output?

6. Practical exercise

Exercise: Review a real document through an accountability lens

Select a document you recently prepared, such as a project report, concept note, briefing, evaluation section, meeting summary or training material. Do not upload confidential content into a public AI tool for this exercise.

Your task

  1. Identify one section where AI could have supported drafting, summarising or organisation.
  2. Identify one section where human expertise, confidentiality or political judgement was essential.
  3. Assess the main risk if AI had been used carelessly.
  4. Define one safeguard that would have reduced that risk.
  5. Name the person or role that should remain accountable for the final output.
Document section Could AI help? Main risk Human judgement required Safeguard
Example: Executive summary Yes, for an initial synthesis Important caveats may be removed Decide which findings are most significant Compare against the full report and evidence matrix
     
     
     

Reflection

Consider your organisation:

  • Does it have a clear policy on approved AI tools?
  • Do staff know what information may not be entered into public systems?
  • Who is responsible for checking AI-supported outputs?
  • How are bias, safeguarding and do-no-harm risks reviewed?

Where policies are unclear, the safest approach is to treat sensitive information as restricted and seek formal guidance before using an AI tool.

Lesson summary

  • AI can produce convincing but inaccurate, biased or contextually weak outputs.
  • Sensitive, personal and confidential information must be protected.
  • Bias can affect how groups, risks and priorities are represented.
  • AI may support preparation, but important decisions must remain human-led.
  • A named person or team must remain accountable for every final output and decision.

Resources for this lesson

Web link

Lesson content: 1.3 Limits, Risks and Human Accountability

Open resource
Lesson 4

2.1 AI-Supported Context and Needs Analysis

2.1 AI-Supported Context and Needs Analysis
Module 2 · Lesson 2.1

AI-Supported Context and Needs Analysis

Using Artificial Intelligence to Organise Evidence, Identify Gaps and Strengthen Programme Understanding

Good development programming begins with a clear understanding of context, needs, stakeholders, institutions and systems. Artificial intelligence can help teams process large volumes of background information quickly, but it cannot replace direct engagement with communities, local experts, partners or official evidence sources.

Learning objectives

By the end of this lesson, you should be able to:

  • identify appropriate AI applications in context and needs analysis;
  • structure source material for more reliable AI-supported analysis;
  • distinguish evidence, interpretation, assumptions and gaps;
  • design prompts that require transparency and uncertainty; and
  • combine AI-supported desk analysis with stakeholder validation.

1. Why context and needs analysis matter

A project can be technically strong and still fail if it misunderstands the operating environment. Context analysis examines the wider political, economic, social, institutional, environmental and security conditions that influence a programme. Needs analysis focuses more directly on the problems, priorities, capacities and barriers affecting a defined population or system.

Context analysis and needs analysis are related but different

Context analysis explains the environment in which change must occur. Needs analysis identifies the gap between the current situation and the desired situation for a specific group, institution or service.

A strong analysis normally considers:

  • who is affected and how different groups experience the problem;
  • which institutions, markets or systems shape the situation;
  • what existing capacities and assets can support change;
  • what political, social, economic or environmental factors may influence results;
  • what evidence is reliable, current and representative; and
  • what information is still missing.

2. Where AI can add value

Document review

AI can summarise policies, evaluations, strategies, assessments and research reports.

Human role: confirm whether sources are current, authoritative and relevant to the target area.

Policy comparison

AI can compare national plans, donor strategies and sector frameworks against common themes or criteria.

Human role: interpret institutional mandates, political commitments and implementation gaps.

Consultation synthesis

AI can organise notes from interviews, focus groups and workshops into recurring themes.

Human role: protect identities, review minority views and ensure that frequency is not confused with importance.

Stakeholder mapping

AI can structure information about mandates, interests, influence, incentives and relationships.

Human role: validate informal influence, hidden interests and politically sensitive relationships.

Barrier identification

AI can identify repeated constraints affecting access, participation, services, markets or institutional performance.

Human role: distinguish symptoms from root causes and examine differences between population groups.

Evidence-gap analysis

AI can identify unanswered questions, weak claims and areas requiring further data collection.

Human role: decide which gaps are material and how they should be investigated.

3. Real-life development examples

Youth employment programme

A team preparing a livelihoods programme provides AI with a labour-market assessment, national employment strategy, employer interviews and youth consultation notes. AI identifies recurring constraints such as limited work experience, weak transport links, skills mismatch and poor access to business finance.

The team then compares these themes with official labour data and conducts validation sessions with young women, rural youth, employers and vocational training providers.

Municipal climate resilience

A climate programme uses AI to compare municipal development plans, flood-risk assessments and community consultation records. AI highlights repeated concerns about drainage, informal housing, water supply and limited maintenance budgets.

Engineers, municipal officials and affected residents verify which risks are current, where impacts are most severe and which groups face the highest exposure.

Governance reform

A governance team supplies laws, institutional mandates, previous assessments and stakeholder interviews. AI produces an initial map of overlapping responsibilities and coordination gaps.

Political-economy analysis is still needed to understand incentives, informal authority, reform resistance and the difference between formal rules and actual practice.

Humanitarian service access

AI helps organise non-identifiable consultation notes from displaced households, local authorities and service providers. It identifies barriers relating to documentation, distance, cost and information.

Protection specialists verify that the analysis does not expose individuals, reinforce stereotypes or overlook groups with limited participation in consultations.

4. Building a reliable AI-supported analysis

AI performs better when the user provides source material, clear boundaries and explicit analytical instructions. A vague request such as “What are the needs of rural youth?” invites generic assumptions. A stronger request defines the context, population, sources and expected output.

  1. Define the analytical question. State exactly what the analysis should explain, such as barriers to women-led SMEs accessing climate finance in a specific region.
  2. Select and assess the evidence. Use official statistics, credible research, consultation notes, programme data and local expert input.
  3. Remove sensitive information. Anonymise personal data, confidential statements and politically sensitive details before using an approved AI tool.
  4. Set clear boundaries. Instruct the AI to use only the supplied material and to avoid adding unsupported external facts.
  5. Separate evidence from interpretation. Ask for direct findings, possible explanations, assumptions and unanswered questions in separate sections.
  6. Validate with people and data. Compare the output with field findings, official evidence and the perspectives of affected groups.

AI cannot confirm whose needs matter most

AI may identify recurring themes, but it cannot decide which need should be prioritised, whether a consultation was representative or whether one group’s preferences should outweigh another’s. Prioritisation requires transparent criteria, participation and accountable human judgement.

5. Prompt structure for context analysis

A useful prompt should include five elements:

  • Role: the type of analytical support required;
  • Context: country, sector, location and programme purpose;
  • Population: the specific target group or institution;
  • Evidence rule: which sources may be used; and
  • Output structure: findings, assumptions, gaps and validation questions.
Act as a development context-analysis assistant. Review only the consultation notes provided below. The analysis concerns unemployed and underemployed young people aged 18–29 in three rural municipalities. Identify: (1) the five most frequently reported needs, (2) differences between women and men where supported by the notes, (3) existing capacities or opportunities, (4) statements that are interpretations rather than direct evidence, and (5) evidence gaps requiring validation. Do not invent facts. Clearly mark uncertainty and list five questions for follow-up consultations.

Good practice: ask AI to challenge the analysis

After receiving an initial summary, ask the tool to identify missing perspectives, contradictory evidence, unsupported claims and alternative explanations. This can strengthen critical review, but the results still require human verification.

6. Common analytical mistakes

  • Using generic AI knowledge instead of supplied evidence. This may introduce outdated or irrelevant assumptions.
  • Treating consultation frequency as statistical representativeness. Repeated comments may reflect who attended, not the wider population.
  • Ignoring differences within a target group. Age, gender, disability, location, income and legal status may shape needs differently.
  • Confusing needs with proposed solutions. A request for training does not automatically mean that lack of training is the root problem.
  • Overlooking existing capacities. Needs analysis should identify assets, institutions and coping strategies, not only deficits.
  • Failing to record uncertainty. Weak evidence should be presented as a gap, not converted into a confident conclusion.

7. Practical exercise

Exercise: Draft an evidence-based needs-analysis prompt

Your organisation has completed six small-group consultations with women-owned microenterprises in two secondary cities. The notes discuss access to finance, energy costs, digital skills, market access and climate-related disruptions. The team wants an initial synthesis before planning further interviews.

Your task

  1. Define the target group and geographic scope.
  2. Instruct AI to use only the supplied consultation notes.
  3. Ask it to separate direct evidence from interpretation.
  4. Require identification of recurring needs, existing capacities and differences between groups.
  5. Ask for evidence gaps and questions for stakeholder validation.

Model prompt

Review only the anonymised consultation notes provided below. The target group is women-owned microenterprises operating in two secondary cities. Summarise the main barriers affecting business continuity and growth. Separate: (1) direct evidence from participants, (2) possible interpretations, (3) existing capacities and coping strategies, and (4) evidence gaps. Identify any differences by sector, business size or location where supported by the notes. Do not invent facts or generalise beyond the evidence. Conclude with six questions for follow-up interviews.
Analysis question Evidence provided AI-supported task Evidence gap Validation method
Example: Why are firms losing customers? Consultation notes and sales observations Group reported market-access barriers No customer or market data Interview buyers and review sales records
     
     
     

Reflection

Consider a recent analysis undertaken by your team:

  • Which sources carried the most weight, and why?
  • Whose perspectives were missing or underrepresented?
  • Which findings were evidence-based and which were assumptions?
  • Where could AI have improved organisation without replacing consultation?

AI-supported analysis is most useful when it makes evidence easier to review and gaps easier to see. It becomes risky when it creates an illusion of certainty.

Lesson summary

  • AI can help organise, compare and summarise evidence for context and needs analysis.
  • Reliable outputs require clear questions, defined populations and supplied source material.
  • Evidence, interpretation, assumptions and gaps should be presented separately.
  • AI cannot replace direct engagement with communities, partners, local experts or official data.
  • All findings must be validated before they inform programme priorities or resource decisions.

Resources for this lesson

Web link

Lesson content: 2.1 AI-Supported Context and Needs Analysis

Open resource
Lesson 5

2.2 Strengthening Theories of Change and Results Frameworks

2.2 Strengthening Theories of Change and Results Frameworks
Module 2 · Lesson 2.2

Strengthening Theories of Change and Results Frameworks

Using Artificial Intelligence to Test Programme Logic, Assumptions and Measurable Results

A theory of change explains how and why a programme is expected to contribute to desired outcomes. Artificial intelligence can help teams structure causal pathways, identify assumptions, draft result statements and test logical consistency. It cannot determine whether the proposed change process is realistic, politically feasible or supported by sufficient evidence.

Learning objectives

By the end of this lesson, you should be able to:

  • explain the purpose of a theory of change and results framework;
  • identify appropriate AI applications in programme logic development;
  • review causal links, assumptions, risks and unintended effects;
  • assess whether indicators are measurable, feasible and useful; and
  • design a prompt that tests a draft results chain for logical gaps.

1. From activities to meaningful change

Development projects often begin with a list of activities: training staff, preparing policies, delivering grants, building infrastructure or supporting coordination. A theory of change goes further. It explains why these activities are expected to influence behaviour, institutions, services or living conditions—and what must be true for that change to occur.

A practical definition

A theory of change describes the causal pathway from intervention to intended change, including the assumptions, external factors and evidence supporting that pathway. A results framework translates this logic into measurable result levels, indicators, baselines, targets and means of verification.

InputsResources used
ActivitiesWhat the project does
OutputsDirect deliverables
OutcomesChanges in behaviour or performance
ImpactLonger-term development change
Level Example Key question
Activity Train municipal engineers in climate-risk assessment. What will the programme do?
Output Municipal engineers complete practical risk-assessment training. What immediate deliverable or capacity is produced?
Outcome Municipal planning teams apply climate-risk evidence in infrastructure decisions. What behaviour, practice or performance changes?
Impact Municipal infrastructure and services are more resilient to climate hazards. What longer-term condition improves?

2. Where AI can support programme logic

Structuring a results chain

AI can organise a problem statement, target groups, intervention areas and donor priorities into a draft sequence of activities, outputs, outcomes and impact.

Human role: confirm whether the causal sequence is realistic and supported by evidence.

Identifying assumptions

AI can suggest conditions that must hold for one result level to lead to the next, such as political commitment, staff retention, budget availability or market demand.

Human role: determine which assumptions are critical, testable and within or outside programme influence.

Testing logical consistency

AI can compare activities, outputs and outcomes to identify gaps, duplication or result statements that are not connected to the intervention.

Human role: assess causality, contribution and whether the proposed change is plausible in the local context.

Exploring alternative pathways

AI can suggest different ways a programme might influence change and identify where complementary interventions may be needed.

Human role: decide which pathway is politically, technically and financially viable.

Drafting result statements

AI can help distinguish outputs from outcomes and improve wording so that results describe change rather than activities.

Human role: ensure statements are specific, realistic and aligned with the programme mandate.

Identifying unintended effects

AI can prompt teams to consider negative, unequal or unexpected effects on stakeholders and systems.

Human role: assess likelihood, severity, safeguarding implications and mitigation measures.

3. Real-life example: strengthening a livelihoods theory of change

A youth employment programme proposes to train 1,000 young people in digital skills. Its initial logic is: “Training leads to jobs and higher income.”

AI can help the team unpack this weak causal statement by asking:

  • Are the skills aligned with actual employer demand?
  • Can participants access devices, connectivity and transport?
  • Do employers recognise the training certificate?
  • Are women and persons with disabilities able to participate safely?
  • Are internships, job-placement services or business support also required?
  • What happens if the local labour market cannot absorb new workers?

A stronger pathway may be:

ActivityDemand-led digital training and career support
OutputParticipants complete relevant training and placements
Short-term outcomeParticipants demonstrate job-ready skills
Intermediate outcomeMore participants obtain and retain suitable work
ImpactImproved economic inclusion and income security

AI may make weak logic look convincing

A fluent narrative does not prove causality. A credible theory of change should draw on evidence, consultation and local knowledge. AI can help expose questions, but it cannot confirm that the proposed pathway will work.

4. Strengthening results frameworks

AI can help draft indicators, means of verification and learning questions. This can save time, but automatically generated indicators are often too broad, too costly, poorly defined or disconnected from management decisions.

A useful indicator should be

  • specific: clear about what is being measured;
  • measurable: supported by a practical method and reliable data;
  • relevant: linked directly to the result;
  • disaggregated where appropriate: able to reveal differences between groups;
  • ethical: collected without exposing or harming participants; and
  • useful: capable of informing programme decisions.

Weak indicator

“Improved youth empowerment.”

This is vague and does not define what empowerment means, how it will be measured or when change is expected.

Stronger indicator

“Percentage of supported participants who report increased influence over employment or business decisions six months after completing the programme, disaggregated by sex, disability and location.”

This is more specific, but the team must still validate the tool, cost and ethical implications.

AI can also help identify:

  • missing baselines and unrealistic targets;
  • indicators that measure outputs rather than outcomes;
  • result statements with no corresponding indicator;
  • indicators with no feasible data source;
  • missing disaggregation; and
  • learning questions needed to test assumptions.

5. A practical AI-supported design workflow

  1. Provide the evidence base. Include the problem analysis, target groups, intervention scope, stakeholder insights and relevant donor priorities.
  2. Ask for a draft, not a final framework. Instruct AI to propose options and identify uncertainty.
  3. Test each causal link. Ask why one result is expected to lead to the next and what evidence supports the connection.
  4. Review assumptions and risks. Identify external conditions, stakeholder behaviour and possible unintended effects.
  5. Assess indicators. Check measurability, feasibility, ethics, cost, data availability and management usefulness.
  6. Validate with stakeholders. Review the logic with technical experts, partners, MEL staff and affected groups.

6. Prompt structure for reviewing a results chain

Act as a programme-design and MEL review assistant. Review the draft results chain provided below. Use only the information supplied. For each link between activities, outputs, outcomes and impact: (1) assess whether the causal connection is clear, (2) identify weak or missing assumptions, (3) identify missing stakeholders, (4) identify possible unintended positive or negative effects, and (5) list evidence needed to validate the logic. Do not rewrite the framework until after completing the diagnostic review. Clearly distinguish evidence, interpretation and uncertainty.

Good practice: ask for challenge, not agreement

AI often follows the framing provided by the user. Ask it explicitly to identify contradictions, missing links, overly optimistic assumptions and alternative explanations. This produces a more useful design review than asking whether the theory of change is “good”.

7. Practical exercise

Exercise: Diagnose a weak results chain

A municipal governance project proposes the following chain:

Train municipal staff ? improve service delivery ? increase citizen trust.

Your task

  1. Draft a prompt asking AI to review the chain for logical gaps.
  2. Require identification of weak assumptions.
  3. Ask which stakeholders are missing.
  4. Ask for possible unintended effects.
  5. Request evidence and indicators needed to test the pathway.

Model prompt

Review this draft results chain for a municipal governance programme: “Train municipal staff ? improve service delivery ? increase citizen trust.” Identify: (1) missing intermediate results, (2) weak assumptions, (3) stakeholders whose behaviour affects the pathway, (4) possible unintended effects, and (5) evidence needed to validate each causal link. Suggest a revised results chain only after presenting the diagnostic findings. For each proposed outcome, suggest one feasible indicator and explain its main measurement limitation.
Results-chain element Current statement Weakness or gap AI-supported improvement Human validation required
Outcome Improve service delivery Too broad; no specific service or behaviour defined Specify service standard, user group and institutional practice Confirm mandate, data source and realistic timeframe
     
     
     

Reflection

Consider a theory of change or logframe you have recently used:

  • Which causal link is least supported by evidence?
  • Which assumption is most important to programme success?
  • Which stakeholder could block or enable the intended change?
  • Which indicator is costly, unclear or not useful for management?

AI is most useful when it helps teams question and refine programme logic. It is least useful when it simply produces a polished diagram without evidence or stakeholder validation.

Lesson summary

  • AI can help structure theories of change, results chains, assumptions and indicators.
  • Strong programme logic requires evidence for each causal link.
  • Indicators must be measurable, feasible, ethical and useful for decision-making.
  • AI-generated frameworks may appear convincing while remaining unrealistic or incomplete.
  • Technical experts, MEL specialists, partners and affected stakeholders must validate the final framework.

Resources for this lesson

Web link

Lesson content: 2.2 Strengthening Theories of Change and Results Frameworks

Open resource
Lesson 6

2.3 Proposal Development and Donor Responsiveness

2.3 Proposal Development and Donor Responsiveness
Module 2 · Lesson 2.3

Proposal Development and Donor Responsiveness

Using Artificial Intelligence to Strengthen Compliance, Clarity and Proposal Quality

Proposal and business development teams can use artificial intelligence to improve efficiency during opportunity analysis, compliance review, outline preparation and drafting. AI can make proposal work faster and more organised, but successful bids still depend on evidence, strategy, local insight, credible partnerships and accountable human judgement.

Learning objectives

By the end of this lesson, you should be able to:

  • identify practical AI applications across the proposal-development process;
  • extract and organise donor requirements into a compliance matrix;
  • use AI to review a draft against donor criteria without inventing evidence;
  • recognise confidentiality, quality and generic-writing risks; and
  • apply a structured human review before proposal submission.

1. Where AI can support proposal development

Opportunity analysis

AI can summarise a funding notice, terms of reference or request for proposals and identify the donor’s objectives, target groups, geographic scope, expected results and eligibility conditions.

Human role: determine whether the opportunity fits organisational strategy, technical capacity, relationships, risk appetite and available resources.

Compliance review

AI can extract submission requirements, page limits, annexes, deadlines, evaluation criteria and mandatory declarations into a structured checklist.

Human role: verify every requirement against the official solicitation and assign responsibility.

Outline development

AI can convert donor instructions into a proposal structure and suggest where evidence, examples, results and partner roles should appear.

Human role: shape the argument, decide emphasis and ensure the outline reflects the win strategy.

Drafting support

AI can help prepare first drafts of non-sensitive sections, improve clarity, shorten repetitive text and adapt technical language for different audiences.

Human role: add evidence, local knowledge, organisational experience and credible commitments.

Criterion-by-criterion review

AI can compare a draft section with evaluation criteria and flag missing responses, weak logic, unsupported claims and unclear wording.

Human role: decide whether the critique is valid and provide the evidence needed to improve the draft.

Final quality control

AI can identify inconsistent terminology, duplicated content, weak transitions and sections that do not clearly answer the donor’s question.

Human role: conduct final compliance, technical, budget, safeguarding and management review.

2. Understanding donor responsiveness

A responsive proposal does more than describe a good project. It answers the specific opportunity, reflects the donor’s language and priorities, addresses every evaluation criterion and provides evidence that the proposed approach is feasible.

Responsive does not mean generic alignment

Statements such as “the project supports inclusion, sustainability and capacity building” are not enough. A strong proposal explains exactly how the intervention responds to the donor’s objectives, who will benefit, what will change, why the approach is credible and how results will be measured.

AI can help test whether a proposal:

  • answers each donor question directly;
  • uses evidence rather than broad assertions;
  • connects the problem, approach, results and budget;
  • explains partner roles and comparative advantages;
  • addresses cross-cutting requirements such as gender, climate, conflict sensitivity or localisation; and
  • distinguishes organisational experience from proposed future activities.

3. Building a compliance matrix

A compliance matrix converts donor instructions into a practical management tool. It helps the team confirm that every requirement has been addressed and identifies who is responsible for providing the response and supporting evidence.

Donor requirement Proposal response Evidence Status
Explain the target population and selection approach. Section 2.1: Target groups and geographic focus Needs assessment, partner records and selection criteria Drafted; validation required
Demonstrate experience in climate-resilient livelihoods. Section 4.2: Organisational capacity Three completed project references and performance data Evidence incomplete
Describe the monitoring and learning approach. Section 3.5: MEL and adaptive management Draft indicators, learning questions and review schedule In progress
Submit signed partner declarations. Mandatory annex Signed forms from each consortium member Not started

Good practice: treat AI extraction as a first pass

AI may miss requirements hidden in annexes, footnotes, templates or linked guidance. Always compare the compliance matrix against the complete official documentation and update it as clarifications are issued.

4. Developing stronger proposal content

AI-generated proposal language often sounds polished but generic. It may repeat common phrases such as “holistic approach”, “inclusive stakeholder engagement” or “sustainable capacity building” without explaining what the team will actually do.

Generic statement

“The project will empower vulnerable communities through an inclusive and sustainable capacity-building approach.”

This does not explain who will be supported, what capacity will change, how inclusion will be achieved or what evidence supports the approach.

Stronger statement

“The programme will support 40 local producer groups to adopt drought-resilient production and financial planning practices through seasonal coaching, demonstration plots and links to participating rural finance institutions. Women-led groups and farmers with disabilities will receive adapted outreach and accessibility support.”

This version defines actors, actions, scale and inclusion measures. The proposal team must still verify feasibility and evidence.

AI can improve drafting by asking it to:

  • replace vague claims with requests for specific evidence;
  • identify where the draft describes activities but not results;
  • flag claims that are not supported by data or examples;
  • suggest clearer topic sentences and logical transitions;
  • shorten repetition without removing required content; and
  • show where donor terminology has been used without explanation.

5. Confidentiality and proprietary information

Proposal information may be commercially sensitive

Do not enter confidential bid information, partner details, proprietary methodologies, internal pricing, staff data, negotiation positions or unpublished technical approaches into AI tools that are not approved for such use.

Before using AI, confirm:

  • whether the organisation has approved the specific tool;
  • whether uploaded content may be stored or reused;
  • whether partner agreements restrict disclosure;
  • whether personal or financial information has been removed; and
  • whether the task can be completed using anonymised or fictionalised content.

6. A practical AI-supported proposal workflow

  1. Review the opportunity. Summarise the donor’s purpose, scope, eligibility, expected results, evaluation criteria and submission rules.
  2. Make the bid decision. Assess strategic fit, technical strengths, partner requirements, budget realism and probability of success.
  3. Create the compliance matrix. Record every requirement, owner, evidence source, deadline and status.
  4. Develop the win strategy. Define the core problem, differentiated approach, evidence, partnerships and reasons the team is credible.
  5. Draft section by section. Use AI for structure, clarity and first drafts while keeping evidence, commitments and sensitive content under human control.
  6. Review against criteria. Test every section for compliance, responsiveness, unsupported claims, feasibility and internal consistency.
  7. Complete human sign-off. Conduct technical, financial, contractual, safeguarding and senior-management approval.

7. Prompt structure for proposal review

Act as a proposal quality reviewer. Compare the proposal section below against the donor criterion provided. Use only the text and evidence supplied. Identify: (1) which parts of the criterion are fully addressed, (2) which parts are missing or unclear, (3) unsupported claims, (4) vague or generic wording, and (5) evidence that should be added. Suggest clearer wording where possible, but do not invent facts, data, experience, partnerships or commitments. Present findings in a table with columns for criterion, finding, risk and recommended action.

Good practice: review before rewriting

Ask AI to diagnose the weaknesses first. If it rewrites immediately, it may conceal gaps with fluent language. A useful review should show what is missing before suggesting improved wording.

8. Practical exercise

Exercise: Review a proposal section against donor criteria

A donor criterion asks the applicant to explain how local organisations will participate in programme governance, implementation and learning. The draft states:

“The project will work closely with local partners and build their capacity throughout implementation.”

Your task

  1. Draft a prompt asking AI to compare the statement with the donor criterion.
  2. Instruct it to identify missing elements and unsupported claims.
  3. Ask it to suggest clearer wording without inventing partner arrangements.
  4. Require a list of evidence or decisions the proposal team must provide.
  5. Identify which parts require partner consultation before revision.

Model prompt

Review the following proposal statement against this donor criterion: “Explain how local organisations will participate in programme governance, implementation and learning.” Draft statement: “The project will work closely with local partners and build their capacity throughout implementation.” Identify which parts of the criterion are not addressed, which claims require evidence, and which terms are too vague. Suggest a clearer structure for the response, but do not invent partner names, governance arrangements, budgets or commitments. Conclude with five questions that the proposal team must answer with local partners before revising the section.
Review criterion Finding Risk Suggested revision Evidence needed
Local participation in governance No decision-making role is described Proposal may appear tokenistic or non-responsive Explain representation, voting or advisory role Agreed governance structure and partner confirmation
     
     
     

Reflection

Consider a recent proposal developed by your organisation:

  • Which donor criterion was hardest to answer with evidence?
  • Which section contained the most generic language?
  • Which information was too sensitive for a public AI tool?
  • Where could AI have improved compliance without replacing strategy?

AI is most useful when it helps teams see gaps, organise requirements and improve clarity. It becomes risky when it creates claims, experience or commitments that the organisation cannot verify.

Lesson summary

  • AI can support opportunity analysis, compliance review, outlining, drafting and quality control.
  • A strong proposal must answer donor criteria directly and provide credible evidence.
  • Compliance matrices help teams manage requirements, responsibilities and missing information.
  • Confidential and proprietary bid information must be protected.
  • Successful proposals still depend on strategy, partnerships, realistic budgets and accountable human expertise.

Resources for this lesson

Web link

Lesson content: 2.3 Proposal Development and Donor Responsiveness

Open resource
Lesson 7

3.1 Project Management and Adaptive Implementation

3.1 Project Management and Adaptive Implementation
Module 3 · Lesson 3.1

Project Management and Adaptive Implementation

Using Artificial Intelligence to Improve Coordination, Reflection and Timely Decision-Making

Project teams manage complex information across workplans, budgets, partner updates, risks, meetings, donor communications and implementation decisions. Artificial intelligence can help organise this information and support reflection, but management decisions must remain accountable, context-aware and aligned with contractual and governance arrangements.

Learning objectives

By the end of this lesson, you should be able to:

  • identify practical AI applications in day-to-day project management;
  • use AI to organise meetings, actions, updates and implementation risks;
  • apply AI to support adaptive management and reflection;
  • recognise decisions that require contractual, political or safeguarding judgement; and
  • draft a prompt that converts meeting notes into an accountable action list.

1. Where AI can support project management

Meeting preparation

AI can draft agendas, organise discussion questions and prepare short briefing notes based on approved project information.

Human role: decide the meeting purpose, participants, sensitivities and decisions required.

Meeting follow-up

AI can turn notes into summaries, decisions, action points, deadlines and unresolved questions.

Human role: verify what was actually agreed and confirm responsibilities with participants.

Partner update synthesis

AI can compare narrative updates, group recurring issues and identify inconsistencies or missing information.

Human role: interpret partner performance, capacity constraints and relationship dynamics.

Risk identification

AI can scan approved narrative updates for delays, dependencies, repeated bottlenecks and emerging implementation risks.

Human role: assess likelihood, impact, ownership and appropriate mitigation.

Briefing and donor communication

AI can prepare first drafts of status notes, mission briefs, decision memos and talking points.

Human role: confirm accuracy, contractual implications, tone and authorised messaging.

Workplan tracking

AI can compare recent updates with milestones and flag tasks that appear delayed, unclear or dependent on unresolved decisions.

Human role: determine whether dates, scope or resources should change.

2. What adaptive implementation means

Adaptive implementation is the disciplined process of using evidence, feedback and changing conditions to adjust how a programme is delivered while protecting its objectives, accountability and compliance. It is not informal improvisation. Changes should be documented, reviewed and approved at the appropriate level.

A simple adaptive-management cycle

Observe ? Interpret ? Decide ? Act ? Review. AI can support observation and organisation, and it can generate questions for interpretation. Accountable managers and governance bodies must make and authorise decisions.

Management task AI support Human decision
Compare progress with milestones Flag delayed activities and inconsistent status updates Decide whether corrective action or formal revision is needed
Review partner feedback Group recurring barriers, concerns and requests Assess credibility, priority and relationship implications
Analyse monitoring findings Summarise patterns and generate reflection questions Determine whether the implementation approach should change
Update the risk register Identify possible emerging risks in reports and meeting notes Rate risks, assign ownership and approve mitigation
Prepare donor communication Draft a clear summary of progress, constraints and next steps Approve disclosure, commitments and contractual interpretation

3. Example: adapting a training programme

A regional livelihoods project plans to train 600 small businesses through in-person workshops. Monitoring updates show low attendance by women-owned businesses, high transport costs and repeated requests for shorter sessions delivered closer to participants.

AI could help the team:

  • compare attendance data with the original participation targets;
  • summarise reasons given for non-attendance;
  • group feedback by location, business type and participant group;
  • identify questions for a reflection meeting; and
  • draft alternative delivery scenarios for discussion.

The project team must still decide whether to change the schedule, delivery model, budget or targets. These decisions may require consultation with partners, donor approval, safeguarding review or formal contract modification.

AI cannot authorise a programme change

AI may suggest alternatives, but it does not understand the full contract, approval hierarchy, political context, safeguarding risks or operational consequences. Changes affecting scope, budget, targets, partner responsibilities or commitments must follow the project’s governance procedures.

4. Turning information into management action

AI-supported summaries are useful only when they lead to clear ownership and follow-up. A strong action list distinguishes decisions already made from proposals, open questions and risks.

  1. Prepare safe source material. Remove personal, confidential and security-sensitive information.
  2. Define the purpose. State whether the output is a summary, action list, risk review or briefing note.
  3. Separate decisions from discussion. Ask AI not to present suggestions as agreed actions.
  4. Assign responsibility carefully. Use only names or roles explicitly stated in the notes.
  5. Flag uncertainty. Mark unclear deadlines, owners and commitments for confirmation.
  6. Validate and circulate. A responsible team member checks the output before it becomes an official record.

Good practice: use status labels

Ask AI to classify items as Decision made, Action agreed, Proposal for review, Risk or Open question. This reduces the danger of turning an informal discussion into an apparent commitment.

5. Prompt structure for meeting notes

Review only the anonymised meeting notes provided below. Produce: (1) a concise summary of the discussion, (2) decisions explicitly agreed, (3) an action list with responsible person or role and deadline only where stated, (4) risks and dependencies mentioned, and (5) follow-up questions for unclear items. Do not invent owners, deadlines, decisions or commitments. Mark any uncertain information as “To be confirmed”. Exclude personal opinions and sensitive personal information.

6. Using AI in reflection and adaptation sessions

AI can help prepare reflection sessions by comparing planned and actual progress, organising feedback and generating questions. It should support discussion rather than predetermine the conclusion.

Useful reflection questions

  • What changed since the workplan was approved?
  • Which assumptions no longer hold?
  • Which groups are benefiting less than expected?
  • What bottlenecks are repeated across partners?
  • Which adjustment can be tested at low risk?

Evidence to review

  • monitoring data and milestone status;
  • partner and participant feedback;
  • risk and issue logs;
  • budget and procurement progress;
  • changes in policy, security or operating context.

7. Practical exercise

Exercise: Convert meeting notes into an accountable action list

A partner coordination meeting discussed delayed training materials, venue confirmation, participant outreach and a possible change to the launch date. Some issues were agreed, while others were left for management approval.

Your task

  1. Draft a prompt asking AI to summarise the meeting.
  2. Require separate lists for decisions, actions, risks and open questions.
  3. Ask it to identify responsible persons and deadlines only where stated.
  4. Instruct it to mark uncertain items as “To be confirmed”.
  5. Ensure that no sensitive personal information is included.

Model prompt

Turn the anonymised meeting notes below into a project-management record. Create separate sections for: decisions made, actions agreed, risks or dependencies, and open questions. For each action, record the responsible person or role and deadline only if explicitly stated. Do not infer agreement from discussion. Do not invent names, dates or commitments. Mark missing details as “To be confirmed”. Conclude with five follow-up questions the project manager should resolve before circulating the record.
Action Responsible person Deadline Risk or dependency Follow-up question
Finalise training materials Training lead To be confirmed Technical review not completed Who approves the final version?
     
     
     

Reflection

Consider your current project-management practices:

  • Which recurring coordination task takes the most time?
  • Where are decisions, actions and proposals sometimes confused?
  • Which project information is too sensitive for a public AI tool?
  • Who has authority to approve implementation changes?

AI is most useful when it reduces administrative workload and makes issues easier to see. It should not replace management judgement, governance or documented approval.

Lesson summary

  • AI can support agendas, meeting summaries, action lists, briefings and risk identification.
  • Adaptive management uses evidence and feedback to improve implementation.
  • AI can organise information and prompt reflection, but it cannot approve programme changes.
  • Actions, owners, deadlines and decisions must be verified before circulation.
  • Project managers remain accountable for context-aware, contractual and ethical decisions.

Resources for this lesson

Web link

Lesson content: 3.1 Project Management and Adaptive Implementation

Open resource
Lesson 8

3.2 Monitoring, Evaluation and Learning Workflows

3.2 Monitoring, Evaluation and Learning Workflows
Module 3 · Lesson 3.2

Monitoring, Evaluation and Learning Workflows

Using Artificial Intelligence to Organise Evidence, Support Qualitative Analysis and Strengthen Learning

Monitoring, evaluation and learning work often involves collecting, cleaning, analysing and interpreting diverse forms of evidence. Artificial intelligence can improve efficiency, especially when working with text-based information, but findings must remain transparent, ethically produced and validated by people who understand the programme and its context.

Learning objectives

By the end of this lesson, you should be able to:

  • identify appropriate AI applications within MEL workflows;
  • use AI to support qualitative coding and thematic synthesis;
  • distinguish patterns in the data from interpretation and claims;
  • document how AI was used and how outputs were checked; and
  • draft a prompt for transparent and cautious qualitative analysis.

1. Where AI can support MEL

Data collection tools

AI can generate draft survey questions, interview guides, observation checklists and focus-group prompts based on evaluation questions or indicators.

Human role: ensure questions are valid, neutral, culturally appropriate, accessible and ethically approved.

Qualitative coding

AI can suggest coding categories and classify anonymised interview notes, open-ended survey responses or learning-session records into themes.

Human role: refine the codebook, review ambiguous cases and preserve minority or contradictory views.

Pattern identification

AI can identify recurring issues, differences between groups and possible relationships across large volumes of text.

Human role: assess whether patterns are meaningful, representative and supported by the data.

Finding synthesis

AI can prepare initial summaries of field reports, interviews, evaluation notes and partner updates.

Human role: verify accuracy, restore missing nuance and distinguish findings from interpretation.

Learning questions

AI can suggest questions for after-action reviews, pause-and-reflect sessions and adaptive-management discussions.

Human role: select questions that matter to programme decisions and affected stakeholders.

Draft reporting

AI can help organise evidence into findings, limitations, lessons and recommendations.

Human role: approve conclusions, assess contribution and prevent overstatement of results.

2. Example: coding participant feedback

An evaluator receives 180 anonymised responses to the question: “What was most useful about the training, and what should be improved?”

AI can help organise the responses into themes such as:

  • relevance to participants’ work;
  • accessibility and language;
  • facilitation quality;
  • practical exercises and examples;
  • training duration and pace; and
  • suggested follow-up support.

The evaluator should review the proposed coding structure, check responses assigned to more than one theme, examine negative and minority views, and verify that the summary does not claim more than the data supports.

Frequency is not the same as importance

The most frequently mentioned issue is not automatically the most significant. A safeguarding concern, accessibility barrier or serious unintended effect may be mentioned only once but still require urgent attention.

3. Building a transparent qualitative workflow

  1. Define the analytical purpose. State the evaluation question, population and intended use of the findings.
  2. Prepare the data safely. Remove names, contact details, case identifiers and other sensitive information.
  3. Develop an initial codebook. Combine evaluation questions, known themes and openness to emerging categories.
  4. Use AI for a first-pass analysis. Ask it to classify responses, explain uncertain coding and preserve contradictory evidence.
  5. Review a sample manually. Compare AI coding with human coding and adjust the instructions or codebook.
  6. Validate the full analysis. Review key themes, minority views, example quotations and exceptions.
  7. Document the process. Record the tool, data used, prompt, checks, limitations and human decisions.

A transparent AI-use record should explain

  • which AI tool was used;
  • what type of data was analysed;
  • how information was anonymised;
  • what prompt or coding instructions were applied;
  • what human quality checks were performed; and
  • how final findings and conclusions were approved.

4. Distinguishing data, analysis and conclusion

MEL task AI support Human validation
Identify recurring themes Group similar responses and count coded references Check coding consistency and whether important minority views were retained
Select example quotations Locate concise statements illustrating a theme Verify exact wording, anonymity and representativeness
Compare participant groups Summarise differences where group labels are available Assess sample size, bias and whether comparison is justified
Draft findings Convert coded data into a structured narrative Confirm that every finding is supported by evidence
Draft conclusions Suggest possible interpretations and implications Apply context, triangulate sources and approve final judgement

Good practice: preserve uncertainty

Ask AI to identify responses that could fit more than one category, themes supported by limited evidence and statements that contradict the dominant pattern. Uncertainty should be visible, not removed for the sake of a cleaner narrative.

5. Ethical and analytical risks

Common risks include:

  • Loss of confidentiality: identifiable interview or case information may be exposed.
  • Biased coding: categories may reflect the analyst’s framing rather than participants’ meaning.
  • Loss of nuance: complex experiences may be reduced to simple labels.
  • False representativeness: findings may be generalised beyond the actual sample.
  • Overstated contribution: programme results may be described as impact without sufficient evidence.
  • Selective reporting: inconvenient, negative or contradictory findings may be excluded.

AI must not be used to improve the appearance of results

AI should not manipulate findings, remove negative evidence, exaggerate impact or turn weak evidence into confident conclusions. MEL products must accurately reflect what the data shows, including limitations and uncertainty.

6. Participatory sense-making remains essential

MEL is not only a technical exercise. Communities, partners, implementers and decision-makers may interpret findings differently. Participatory sense-making helps explain why patterns occurred, whether the interpretation is fair and what action should follow.

AI cannot replace conversations about:

  • why different groups experienced the programme differently;
  • whether findings reflect local realities;
  • which unintended effects matter most;
  • what recommendations are feasible; and
  • who should act on the lessons.

7. Prompt structure for qualitative coding

Act as a qualitative-analysis assistant. Analyse only the anonymised responses provided below. Code the responses into themes related to relevance, accessibility, facilitation quality, practical application and suggested improvements. Allow a response to receive more than one code where appropriate. For each theme, provide: (1) a short definition, (2) the number of responses coded, (3) two brief example quotations copied exactly from the supplied data, (4) contradictory or minority views, and (5) coding uncertainty. Do not infer participant characteristics, causal explanations or programme impact beyond the data.

8. Practical exercise

Exercise: Code anonymised qualitative feedback

Your team has collected anonymised open-ended feedback from participants in a project-management course. The responses discuss course relevance, exercises, language, pacing, accessibility and requests for follow-up support.

Your task

  1. Draft a prompt asking AI to code the responses into themes.
  2. Allow multiple codes for a single response.
  3. Request short example quotations from the supplied text.
  4. Ask AI to mark uncertain or contradictory cases.
  5. Instruct it not to make claims beyond the available data.

Model prompt

Review only the anonymised participant responses below. Develop a concise coding framework and classify each response. Use themes related to relevance, accessibility, facilitation, practical exercises, pacing and follow-up needs, while allowing new themes to emerge from the data. A response may receive multiple codes. For each theme, include a definition, number of coded responses and up to two exact example quotations. Note uncertain classifications, contradictory feedback and themes supported by very few responses. Do not claim that findings represent all participants or prove programme impact.
Response Theme Example quote Uncertainty Validation note
The case exercise was useful, but the session moved too quickly. Practical relevance; pacing “The case exercise was useful” Low Two codes are justified by separate parts of the response
     
     
     

Reflection

Consider a recent MEL activity:

  • Which part of the workflow required the most manual effort?
  • How were qualitative coding decisions documented?
  • Which voices or contradictory findings might have been overlooked?
  • How could stakeholders help interpret the findings?

AI can reduce the time required to organise evidence, but it does not remove the need for methodological judgement, ethical safeguards or participatory interpretation.

Lesson summary

  • AI can support data-collection design, qualitative coding, synthesis and learning questions.
  • Qualitative findings should preserve nuance, uncertainty, minority views and contradictory evidence.
  • Teams should document what AI analysed, how it was prompted and how outputs were checked.
  • AI must not be used to manipulate findings or overstate programme impact.
  • Final interpretation and learning should involve MEL specialists, partners and affected communities.

Resources for this lesson

Web link

Lesson content: 3.2 Monitoring, Evaluation and Learning Workflows

Open resource
Lesson 9

3.3 Reporting, Learning Products and Knowledge Sharing

3.3 Reporting, Learning Products and Knowledge Sharing
Module 3 · Lesson 3.3

Reporting, Learning Products and Knowledge Sharing

Using Artificial Intelligence to Communicate Evidence Clearly, Accurately and Responsibly

Development professionals produce donor reports, learning briefs, case studies, policy notes, presentations and communication materials. Artificial intelligence can help transform technical information into clearer formats for different audiences, but every product still requires fact-checking, evidence review, appropriate permissions and accountable human editing.

Learning objectives

By the end of this lesson, you should be able to:

  • identify appropriate AI applications in reporting and knowledge sharing;
  • adapt technical information for different audiences without losing accuracy;
  • distinguish evidence-based achievements from unsupported or exaggerated claims;
  • apply quality and permission checks to stories, quotations and images; and
  • draft a prompt for simplifying technical text responsibly.

1. Where AI can support reporting and communication

Donor reports

AI can help organise progress by outcome area, improve consistency and identify where claims lack supporting evidence.

Human role: verify results, reporting periods, indicators, contractual language and donor requirements.

Learning briefs

AI can convert detailed monitoring and evaluation findings into concise lessons, implications and recommendations.

Human role: confirm that lessons reflect evidence and include important limitations or contradictory findings.

Case studies

AI can help structure a story around context, challenge, intervention, result and lesson.

Human role: protect dignity, obtain consent and avoid presenting individual experience as representative of everyone.

Policy notes

AI can simplify technical analysis, organise policy options and prepare key messages for decision-makers.

Human role: verify policy accuracy, feasibility, political sensitivity and implications.

Presentations and briefs

AI can create outlines, shorten long text and propose audience-appropriate headings or speaker notes.

Human role: decide what matters most and ensure visuals and messages are not misleading.

Communication materials

AI can adapt content for websites, newsletters, social media and community-facing materials.

Human role: approve public claims, terminology, branding, permissions and safeguarding considerations.

2. Reporting progress without exaggeration

Donor reporting should explain what was achieved, what evidence supports the achievement, what remains incomplete and what affected implementation. AI can improve structure and readability, but it may make modest results sound more significant than they are.

Activity is not the same as result

Completing workshops, producing guidance or holding meetings does not automatically demonstrate a change in behaviour, institutional performance or living conditions. Reports should distinguish activities, outputs, outcomes and longer-term impact.

Product Primary audience AI support Quality check
Narrative donor report Donor programme and contract managers Structure progress by result area and identify evidence gaps Verify reporting period, indicators, figures and contractual terminology
Learning brief Programme teams and partners Summarise lessons, implications and practical recommendations Confirm that lessons are supported and limitations remain visible
Case study Partners, donors and the public Develop a clear narrative and key messages Check consent, dignity, attribution and representativeness
Policy note Government and senior decision-makers Simplify analysis and compare policy options Validate legal, institutional and political assumptions

3. Adapting content for different audiences

The same evidence may need to be communicated differently to donors, government officials, technical specialists, communities or internal management. Audience adaptation should change the format, level of explanation and emphasis—not the underlying facts.

Donor audience

Focus on agreed results, indicators, evidence, value for money, challenges, risk management and compliance.

Government audience

Focus on policy relevance, institutional roles, implementation feasibility, public value and decision points.

Community audience

Use clear language, explain practical implications and avoid unexplained technical or donor terminology.

Internal management

Highlight decisions required, operational risks, resource implications, unresolved issues and next steps.

Good practice: specify what must remain unchanged

When asking AI to adapt content, identify the facts, figures, limitations, approved terminology and level of certainty that must be preserved. This reduces the risk that simplification becomes distortion.

4. From technical text to clear communication

Technical version:

“The programme operationalised a multi-stakeholder capacity-development mechanism to enhance municipal application of climate-risk screening methodologies within subnational capital investment planning processes.”

Clearer version:

“The programme helped municipal teams use climate-risk screening when planning local infrastructure investments.”

The clearer version is shorter and easier to understand. Before using it, the team should confirm that “helped municipal teams use” accurately reflects the evidence. If staff only attended training but did not yet apply the method, the statement would need to be more cautious.

5. Evidence, permissions and responsible storytelling

Before publishing a report, brief, story or communication product, check:

  • whether every factual claim is supported by a reliable source;
  • whether figures match approved monitoring or financial records;
  • whether quotations are accurate and properly attributed;
  • whether informed consent covers the intended use of stories or images;
  • whether personal or sensitive information should be removed;
  • whether partner names, logos or information require approval; and
  • whether the product could create safeguarding, security or reputational risk.

Do not let AI invent a human story

AI may produce realistic-sounding quotations, beneficiary profiles or success stories. These must never be presented as real unless they come from verified records and have the necessary permissions.

6. A practical AI-supported workflow

  1. Define the product and audience. Clarify purpose, format, length, tone and intended use.
  2. Prepare verified source material. Use approved data, reports, quotations and terminology.
  3. Set evidence boundaries. Instruct AI not to invent facts, outcomes, quotations or explanations.
  4. Generate structure or a first draft. Use AI to organise information and improve readability.
  5. Check every claim. Trace statements back to evidence and retain uncertainty or limitations.
  6. Review permissions and risk. Confirm consent, attribution, confidentiality and partner approval.
  7. Complete human editorial approval. Ensure the final product is accurate, useful and appropriate for the audience.

7. Prompt structure for audience adaptation

Rewrite the technical paragraph below for a non-specialist audience. Preserve all verified facts, figures, limitations and approved terminology. Use clear sentences and explain technical terms briefly where needed. Do not add new claims, imply causality, exaggerate achievements or remove uncertainty. After the rewrite, list any statement that still requires evidence or clarification before publication.

Good practice: ask for an evidence check after rewriting

A rewritten paragraph may sound stronger than the source material. Ask AI to identify which sentences describe verified facts, which are interpretations and which require additional evidence.

8. Practical exercise

Exercise: Rewrite technical reporting for a non-specialist audience

Technical paragraph:

“During the reporting period, the project facilitated three subnational coordination workshops involving 74 institutional stakeholders. Preliminary participant feedback indicates improved understanding of the referral protocol, although evidence of consistent institutional application is not yet available.”

Your task

  1. Draft a prompt asking AI to rewrite the paragraph for a general audience.
  2. Require it to preserve the number of workshops and participants.
  3. Ensure that preliminary feedback is not presented as proven institutional change.
  4. Ask it to retain the limitation about application evidence.
  5. Request a final list of claims that should be checked before publication.

Model prompt

Rewrite the paragraph below for a non-specialist audience in no more than 90 words. Keep the verified facts that three workshops were held and 74 stakeholders participated. Explain the early feedback clearly, but do not suggest that institutions have already changed their practice. Preserve the limitation that consistent application has not yet been demonstrated. Do not add quotations, causes or results. After the rewrite, list the facts that should be checked against project records before publication.
Original statement Evidence retained Simplified wording Risk of distortion Final check
Preliminary feedback indicates improved understanding Participant feedback collected after workshops Early feedback suggests participants better understood the protocol Could be mistaken for demonstrated behaviour change Confirm feedback method and number of respondents
     
     
     

Reflection

Consider a recent report or knowledge product:

  • Which claims were strongest, and what evidence supported them?
  • Where could technical language have been clearer?
  • Were limitations and incomplete results sufficiently visible?
  • Were all stories, quotations, images and partner references properly approved?

AI can improve communication, but clarity should never come at the cost of accuracy, consent or responsible representation.

Lesson summary

  • AI can support donor reports, learning briefs, case studies, policy notes and communication materials.
  • Audience adaptation should change presentation without changing the underlying facts.
  • Activities, outputs, outcomes and impact should be reported accurately and distinctly.
  • Facts, figures, quotations, stories, images and permissions must be checked before publication.
  • All final products require evidence review, audience-sensitive editing and accountable human approval.

Resources for this lesson

Web link

Lesson content: 3.3 Reporting, Learning Products and Knowledge Sharing

Open resource
Lesson 10

4.1 Bias, Misinformation and Contextual Harm

4.1 Bias, Misinformation and Contextual Harm
Module 4 · Lesson 4.1

Bias, Misinformation and Contextual Harm

Recognising Distorted Assumptions, Unsupported Claims and Harmful Development Narratives

Artificial intelligence systems can reproduce bias because they are shaped by the data, design choices and assumptions behind them. In international development, this can influence how communities are described, whose knowledge is prioritised and which solutions appear credible. Responsible practitioners must review AI outputs for bias, misinformation, missing perspectives and contextual harm.

Learning objectives

By the end of this lesson, you should be able to:

  • recognise common forms of bias in AI-generated development content;
  • identify misinformation, invented references and unsupported claims;
  • assess whether an output is respectful and contextually appropriate;
  • apply practical questions to detect missing perspectives and assumptions; and
  • revise harmful or misleading language without inventing evidence.

1. How bias appears in development work

Bias is not always obvious. It may appear through language, framing, omissions, assumptions or the prioritisation of one form of knowledge over another.

Deficit-based language

Communities may be described mainly through weakness, vulnerability or need, while their knowledge, institutions and coping strategies are ignored.

Review question: Does the text recognise capabilities, agency and existing solutions?

Imported assumptions

AI may suggest that an approach from one country or institution will work elsewhere without considering legal, cultural, political or resource differences.

Review question: What local evidence supports the proposed transfer?

Missing informal institutions

AI may focus only on formal government or donor structures and overlook community leaders, customary systems, informal markets or mutual-support networks.

Review question: Which actors influence outcomes but are absent from the analysis?

Gender-blind analysis

Outputs may assume that women and men have equal access to time, finance, mobility, technology or decision-making.

Review question: Are barriers and impacts disaggregated by gender and other relevant factors?

Passive-beneficiary framing

Local people may be described as recipients of outside solutions rather than rights-holders, partners and decision-makers.

Review question: Who is shown as having knowledge, authority and agency?

Dominant-language bias

AI may prioritise evidence available in widely used languages and overlook local-language research, oral knowledge or community documentation.

Review question: Which sources or perspectives may be absent because they are less digitally visible?

2. Misinformation and invented evidence

AI can generate fluent but incorrect information. It may invent statistics, references, policy provisions, institutional names, quotations or summaries. The language can sound confident even when the content is false.

Confidence is not evidence

A polished answer is not necessarily a correct answer. Any statistic, quotation, legal statement, policy claim, institutional reference or research finding should be checked against a reliable source before use.

Risk Example Review question Corrective action
Invented statistic “Seventy percent of rural households lack access to formal finance.” What source, year, location and population support this figure? Remove or replace with a verified statistic
Fabricated reference A report title or author that cannot be located Does the source exist and say what the text claims? Verify the original publication before citation
Incorrect policy summary A law is described as requiring something it does not require Has a legal or policy specialist checked the wording? Use the official text and qualified review
Unsupported causal claim Training is said to have reduced unemployment Does the available evidence establish this relationship? Use cautious wording and state the evidence limits

3. Contextual harm in apparently neutral language

An output may be factually plausible and still create harm. Language can reinforce stereotypes, exclude particular groups, ignore conflict dynamics or recommend actions that are unsafe in the local context.

Potentially harmful wording:

“Traditional communities resist modern agricultural methods and require intensive awareness raising.”

This framing may be harmful because it:

  • labels communities as resistant without examining their reasons;
  • assumes the promoted method is automatically superior;
  • ignores local knowledge and previous experience;
  • places responsibility for failure only on the community; and
  • suggests persuasion before consultation or evidence gathering.

More respectful and analytical wording:

“Some farmers have not adopted the proposed methods. Further consultation is needed to understand concerns related to cost, risk, local experience, labour requirements and suitability to local conditions.”

Good practice: replace judgement with inquiry

When an AI output uses labels such as resistant, uneducated, backward, dependent or unwilling, ask what evidence supports the description and what structural, historical or practical factors may be missing.

4. Who is represented and who is missing?

Use the following questions when reviewing sensitive outputs:

  • Whose perspective is presented as authoritative?
  • Whose knowledge, language or experience is absent?
  • Are local actors described as partners or merely beneficiaries?
  • Are differences within the target population visible?
  • Does the output recognise power relations and unequal access?
  • Could the wording reinforce stigma, discrimination or stereotypes?
  • Would affected people recognise the description as fair?

5. A practical review workflow

  1. Identify the intended use. Consider the audience, decision and possible consequences of the output.
  2. Separate facts from interpretation. Mark statistics, observations, assumptions and recommendations.
  3. Check representation. Ask whose perspectives are included, excluded or treated as more credible.
  4. Review language. Flag stereotypes, deficit framing, generalisations and passive-beneficiary language.
  5. Verify evidence. Check every important claim, reference, figure and policy statement.
  6. Request alternatives. Ask for competing interpretations and context-specific explanations.
  7. Involve contextual expertise. Use local specialists, partners and affected groups to review sensitive content.

Contextual expertise includes more than technical knowledge

Useful review may require local-language skills, lived experience, institutional knowledge, gender analysis, political-economy understanding, safeguarding expertise and familiarity with community norms.

6. Prompt structure for bias and misinformation review

Review the paragraph below as a critical development-practice editor. Identify: (1) deficit-based or stereotypical language, (2) unsupported claims or statistics, (3) missing stakeholders or perspectives, (4) assumptions transferred from other contexts, (5) language that reduces local actors to passive beneficiaries, and (6) possible contextual or safeguarding harm. Do not invent evidence. Suggest more respectful and precise wording, and list the questions that require local validation before the text is used.

Good practice: request alternative interpretations

Ask AI to provide at least two plausible explanations for a challenge and to state what evidence would be needed to distinguish between them. This can reduce overconfident, single-cause conclusions.

7. Practical exercise

Exercise: Review an AI-generated development paragraph

AI-generated paragraph:

“Poor rural communities are slow to adopt digital financial services because they lack awareness and trust in modern banking. The project should conduct awareness campaigns and encourage local leaders to persuade households to open accounts.”

Your task

  1. Identify biased, deficit-based or overly general language.
  2. Flag claims that are not supported by evidence.
  3. List stakeholders and perspectives missing from the paragraph.
  4. Identify possible barriers not considered by the AI.
  5. Rewrite the paragraph as a neutral problem statement that calls for local evidence.

Model review prompt

Review the paragraph below for bias, missing perspectives, unsupported claims and inappropriate language. Identify where the text assumes that lack of awareness or trust is the main cause without evidence. Suggest additional factors that should be investigated, such as fees, documentation, connectivity, accessibility, previous experience, gender barriers and service quality. Do not assume these factors are present; present them as questions for local research. Rewrite the paragraph using respectful, neutral language.
Text or claim Problem identified Missing perspective or evidence Suggested revision Human validation needed
“Poor rural communities are slow to adopt…” Generalising and deficit-based framing Differences across households, locations and user groups “Use of digital financial services varies across rural communities.” Local usage data and stakeholder consultation
     
     
     

Reflection

Consider an AI output you recently reviewed:

  • Which assumptions were presented as facts?
  • Whose perspective was missing?
  • Did the language recognise local agency and knowledge?
  • Which claims required verification from trusted sources?

Responsible use of AI requires more than checking grammar or factual accuracy. It requires attention to power, representation, dignity and the consequences of how development challenges are framed.

Lesson summary

  • AI outputs can reproduce bias through language, assumptions, omissions and source selection.
  • AI may generate false statistics, references, summaries and causal claims.
  • Development practitioners should check who is represented, who is missing and who is given agency.
  • Trusted evidence and contextual expertise are essential for sensitive outputs.
  • Respectful, accurate and contextually appropriate communication remains a human responsibility.

Resources for this lesson

Web link

Lesson content: 4.1 Bias, Misinformation and Contextual Harm

Open resource
Lesson 11

4.2 Data Protection, Confidentiality and Intellectual Property

4.2 Data Protection, Confidentiality and Intellectual Property
Module 4 · Lesson 4.2

Data Protection, Confidentiality and Intellectual Property

Using Artificial Intelligence Without Exposing Sensitive Information or Organisational Assets

Development work often involves personal data, partner documents, community feedback, safeguarding reports, political analysis, security information and donor-funded intellectual property. Using AI without appropriate safeguards can create privacy, confidentiality, contractual and reputational risks.

Learning objectives

By the end of this lesson, you should be able to:

  • recognise personal, confidential, sensitive and proprietary information;
  • assess whether project information is appropriate for use in an AI tool;
  • apply practical safeguards such as minimisation and anonymisation;
  • recognise intellectual-property and contractual concerns; and
  • use a simple pre-upload checklist before sharing any project content.

1. Why data protection matters

Information entered into an AI tool may be stored, processed in another jurisdiction, reviewed by service providers or used according to the platform’s terms. Even when a tool appears private, users should not assume that it meets organisational, donor or government requirements.

Do not treat a public AI tool as a secure project workspace

Information should not be uploaded simply because it is convenient. If content is personal, confidential, sensitive, proprietary or restricted, use it only when organisational policy permits the tool and the required protections are in place.

2. Types of information that require caution

Personal data

Names, contact details, identification numbers, photographs, employment records, demographic details or information that can identify an individual.

Risk: privacy breach, unauthorised processing or harm to the individual.

Sensitive personal data

Health information, disability status, ethnicity, religion, political opinions, safeguarding information or other data requiring stronger protection.

Risk: discrimination, stigma, retaliation or serious personal harm.

Confidential project information

Internal reports, partner assessments, negotiation positions, unpublished donor communications, draft budgets or management concerns.

Risk: contractual breach, reputational damage or loss of trust.

Security and political information

Security incidents, travel plans, location information, political-economy analysis or sensitive stakeholder mapping.

Risk: physical, political or operational harm.

Proprietary information

Organisational methodologies, proposal strategies, software designs, commercial models or unpublished research.

Risk: loss of competitive advantage or disputed ownership.

Restricted donor or government data

Information subject to grant conditions, government classification, contractual restrictions or data-sharing agreements.

Risk: non-compliance, legal exposure or loss of access.

3. Classifying information before use

Information type Example Main risk Safe response
Public Published policy, public website text or approved public report Low, but accuracy and copyright still matter Verify source and permitted use
Internal Draft workplan or internal meeting summary Unauthorised disclosure or misunderstanding Use only approved tools and remove unnecessary details
Confidential Partner performance concerns or unpublished donor feedback Contractual and reputational harm Do not upload without explicit authorisation
Sensitive or restricted Safeguarding case, political analysis or security information Serious personal, legal or operational harm Do not use in a public AI tool

Good practice: use the least information necessary

Most AI tasks do not require full documents or real names. Replace identifiable content with neutral labels, use short excerpts, remove metadata and provide only the information needed for the specific task.

4. Anonymisation is more than removing names

Removing a person’s name may not be enough. A combination of location, job title, age, project role, incident date or family details may still identify the individual.

Insufficient anonymisation:

“The only female health officer in District X reported a safeguarding concern during the March mission.”

Safer version:

“A staff member reported a safeguarding concern during a recent activity.”

Even the safer version may still be inappropriate for an AI tool if the underlying case is sensitive. Anonymisation reduces risk but does not automatically make all information safe to use.

5. Intellectual property and ownership

Intellectual property includes documents, designs, training materials, methodologies, code, research, branding and other original work. AI use can create uncertainty about ownership, reuse and attribution.

  • Tool terms may affect how uploaded or generated content can be used.
  • Contracts may assign ownership to a donor, government, partner or consortium.
  • AI-generated text may resemble existing material or reproduce protected wording.
  • Third-party documents may have licence or citation requirements.
  • Generated content still requires originality, plagiarism and attribution review.

Do not assume that generated means unrestricted

AI-generated content should not automatically be treated as original, exclusive or free of third-party rights. Check the tool’s terms, applicable contracts, organisational policy and the sources used to create the final product.

6. Organisational safeguards

Organisations should establish clear rules covering:

  • approved AI tools and account types;
  • categories of data that may or may not be uploaded;
  • required anonymisation and redaction procedures;
  • human review and approval responsibilities;
  • citation, attribution and intellectual-property checks;
  • record keeping for important AI-assisted outputs; and
  • incident reporting when information is shared incorrectly.

Approval should match the level of risk

Low-risk use of public information may need only routine review. Confidential, personal, legal, safeguarding, security or proprietary information may require specialist approval—or may be prohibited entirely.

7. A safe pre-upload workflow

  1. Identify the information. Determine what data, text or files the task would require.
  2. Classify the content. Decide whether it is public, internal, confidential, sensitive or proprietary.
  3. Check policy and contract rules. Confirm whether the selected AI tool is approved for this information.
  4. Minimise and anonymise. Remove all content that is not essential to the task.
  5. Consider intellectual property. Confirm ownership, licence, attribution and reuse conditions.
  6. Document the decision. Record important approvals and safeguards for high-risk or official outputs.
  7. Stop when uncertain. Seek guidance from data-protection, legal, IT-security or management staff.

8. Practical exercise

Exercise: Create a five-question pre-upload checklist

Before uploading any project information to an AI tool, ask the following five questions:

1. Does this content contain personal, sensitive, confidential, security-related or proprietary information?
2. Is this AI tool approved by my organisation for this type of information?
3. Do donor, government, partner or contractual rules restrict how the information may be processed or shared?
4. Can I complete the task with less information, anonymised data or a fictional example?
5. Do I have the necessary ownership, permission and approval to upload and reuse this content?
Checklist question Yes / No Required action Approval or evidence
Does the content include confidential partner information? Yes Do not upload; prepare a fictionalised version Project manager confirmation
    
    
    
    

Reflection

Consider how your organisation currently uses AI:

  • Which tools are officially approved?
  • Which categories of information are prohibited?
  • Who provides advice when staff are uncertain?
  • How are important AI-assisted outputs documented and reviewed?

Convenience is not a sufficient reason to expose project information. Responsible AI use begins with data minimisation, appropriate tools and clear accountability.

Lesson summary

  • Development information may be personal, confidential, sensitive, proprietary or restricted.
  • Public AI tools should not be used for protected information unless policy clearly permits it.
  • Anonymisation, minimisation and approved tools reduce risk but do not remove all obligations.
  • Intellectual-property ownership, licensing, attribution and contract terms must be checked.
  • When uncertain, stop and seek authorised data-protection, legal, security or management advice.

Resources for this lesson

Web link

Lesson content: 4.2 Data Protection, Confidentiality and Intellectual Property

Open resource
Lesson 12

4.3 Prompting, Quality Assurance and Responsible Workflows

4.3 Prompting, Quality Assurance and Responsible Workflows
Module 4 · Lesson 4.3

Prompting, Quality Assurance and Responsible Workflows

Designing Clear Instructions and Reviewing AI Outputs with Professional Judgement

Effective AI use depends on clear prompts and strong quality assurance. A well-designed prompt helps the tool understand the task, context, audience, evidence and limits. A responsible workflow ensures that the resulting output is reviewed for accuracy, relevance, bias, tone, compliance and accountability before use.

Learning objectives

By the end of this lesson, you should be able to:

  • structure a clear and reusable professional prompt;
  • apply a prepare–generate–review workflow;
  • identify the main quality checks required for AI-assisted outputs;
  • document significant AI use and approval responsibilities; and
  • create a reusable prompt template for a recurring task.

1. What makes a prompt effective?

A good prompt gives the AI enough structure to produce a useful first draft. It does not guarantee accuracy, but it reduces ambiguity and makes the output easier to review.

Role

Explain the perspective or function the AI should adopt, such as editor, analyst, facilitator or MEL assistant.

Example: “Act as a development-programme reporting assistant.”

Task

State exactly what the AI should do and what it should not do.

Example: “Summarise progress by outcome area without adding new claims.”

Context

Provide the minimum background needed to interpret the task correctly.

Example: “This is a quarterly report for a donor-funded local governance project.”

Audience

Identify who will read or use the output and what they need to understand.

Example: “Write for non-technical donor programme managers.”

Sources

State which information the AI may use and whether outside knowledge is prohibited.

Example: “Use only the supplied text and figures.”

Format and limits

Specify length, headings, table structure, tone, terminology and required cautions.

Example: “Use four headings, remain under 500 words and mark uncertainty clearly.”

A practical prompt formula

Role + Task + Context + Audience + Sources + Output format + Limitations + Quality checks

2. Weak and stronger prompting

Weak prompt:

“Write a good project report.”

This prompt is unclear because it does not define:

  • the reporting period;
  • the intended audience;
  • the source material;
  • the required structure;
  • the level of detail; or
  • the limits on claims and interpretation.

Stronger prompt:

“Act as a donor-reporting editor. Using only the approved progress notes below, draft a 400-word quarterly summary for a non-technical donor audience. Organise the text under progress, challenges, corrective actions and next steps. Preserve all figures and reporting dates exactly. Do not invent achievements, causes or commitments. Mark missing evidence as ‘verification required’.”

Good practice: define the evidence boundary

Tell the AI what it may use, what it must not infer and how to handle missing information. This is especially important for reporting, evaluation, legal, policy and safeguarding tasks.

3. The prepare–generate–review workflow

  1. Prepare: define the task. Clarify the purpose, decision, audience and level of risk.
  2. Prepare: gather reliable sources. Use approved documents, verified figures and agreed terminology.
  3. Prepare: protect information. Remove personal, confidential, sensitive or unnecessary content.
  4. Generate: write a structured prompt. Specify the role, task, context, sources, format and limitations.
  5. Generate: request transparent output. Ask for assumptions, uncertainty, evidence gaps and items requiring confirmation.
  6. Review: test the output. Check accuracy, relevance, completeness, tone, bias, citations and compliance.
  7. Review: approve and document. Ensure the responsible person authorises the final version and records significant AI use where required.

Iteration does not replace verification

Asking AI to “check its answer” may improve wording, but it does not provide independent validation. Important claims must still be checked against trusted records, sources and qualified human judgement.

4. A practical quality-assurance checklist

Quality area Review question Evidence or check Action if weak
Accuracy Are facts, figures, dates and names correct? Compare with approved source records Correct, remove or mark for verification
Evidence Does each important claim have support? Trace statements to data or documentation Reduce certainty or add evidence
Relevance Does the output answer the actual task? Compare with purpose and audience needs Refocus or shorten the output
Bias and context Are assumptions, stereotypes or missing perspectives present? Use contextual and stakeholder review Revise framing and add missing perspectives
Tone and terminology Is the language appropriate and consistent? Compare with organisational and donor standards Edit wording and restore approved terminology
Compliance Does the output meet policy, contract and data requirements? Check procedures and required approvals Escalate, revise or stop use

5. Asking AI to expose uncertainty

AI outputs become easier to review when uncertainty and missing information are visible. Useful instructions include:

  • “Do not invent missing facts.”
  • “Mark unsupported claims as verification required.”
  • “Separate evidence from interpretation.”
  • “List assumptions used in the analysis.”
  • “Identify conflicting information in the source material.”
  • “State where the available evidence is insufficient.”
Use only the supplied information. If the evidence does not support a conclusion, state that clearly. Separate verified facts, interpretations, assumptions and recommendations. Do not invent references, statistics, quotations, decisions or commitments.

6. Human responsibility and documentation

AI is most useful when integrated into professional practice rather than used as a shortcut. The responsible professional remains accountable for the final output, even when AI helped draft or analyse it.

Significant AI use may need to be documented when:

  • the output informs a management or funding decision;
  • AI supported analysis of monitoring or evaluation evidence;
  • the task involved sensitive, legal, policy or reputational considerations;
  • an organisation, donor or client requires disclosure; or
  • the prompt, tool and review process may need to be reproduced later.

Good practice: maintain a simple AI-use record

Record the purpose, tool, date, source material, prompt version, main human checks, revisions and approving person. The level of documentation should reflect the importance and risk of the task.

7. Reusable prompt template

Role: Act as [professional role or function].

Context: This task relates to [project, sector, reporting period or situation].

Task: [Describe exactly what should be produced or analysed].

Audience: The output is for [audience and intended use].

Sources: Use only [approved documents, data or text].

Output format: Present the result as [headings, table, bullets, word limit or template].

Limitations: Do not [invent facts, infer causality, expose sensitive information or alter approved terminology].

Quality checks: Identify [uncertainty, evidence gaps, assumptions, contradictions and items requiring verification].

8. Practical exercise

Exercise: Build a reusable prompt for your role

Choose one recurring task, such as preparing meeting summaries, reviewing partner reports, drafting learning briefs, developing interview guides or adapting technical content.

Your task

  1. Define the professional role the AI should perform.
  2. Describe the task and its practical context.
  3. Identify the audience and purpose.
  4. Specify the approved sources and output format.
  5. Add limitations and quality checks.

Example: partner progress update

Act as a project-management reporting assistant. Review the anonymised partner update below for a quarterly internal review. Use only the supplied text. Produce a table with: progress against milestone, evidence provided, delay or risk, action required, responsible role and issue requiring management decision. Do not invent dates, achievements, explanations or commitments. Mark missing evidence as “not provided”. After the table, list any contradictions, unclear statements or claims requiring verification.
Prompt section What to include Your wording Quality check
Context Project, task and purpose   Is enough context provided without exposing sensitive information?
Task Exact action AI should perform   Could the instruction be interpreted in more than one way?
Audience Reader, user and intended decision   Is the tone and level appropriate?
Sources Approved material and evidence boundaries   Is the AI prohibited from inventing missing content?
Format Structure, length and presentation   Will the output be easy to review?
Limitations Prohibited claims, data and assumptions   Are major risks addressed?
Quality checks Verification, uncertainty and human review   Does the prompt expose evidence gaps?

Reflection

Consider your current AI practice:

  • Which recurring task would benefit most from a standard prompt?
  • Which quality check is most often overlooked?
  • When should AI use be documented or disclosed?
  • Who remains accountable for the final output?

Strong prompts improve efficiency, but professional value comes from combining AI assistance with reliable evidence, contextual knowledge and accountable human review.

Lesson summary

  • Effective prompts define the role, task, context, audience, sources, format and limitations.
  • A responsible workflow follows three stages: prepare, generate and review.
  • Quality assurance should cover accuracy, evidence, relevance, bias, tone and compliance.
  • Significant AI use may need documentation, disclosure and approval.
  • Human professionals remain responsible for every final output and decision.

Resources for this lesson

Web link

Lesson content: 4.3 Prompting, Quality Assurance and Responsible Workflows

Open resource