AI-Powered Citizen Engagement in Uganda – Case Study
From Surveys to Conversations: AI-Powered Citizen Engagement in Uganda
What is this case study about?
This case study documents an innovative AI-powered citizen engagement platform developed by Stephan Dietrich (Assistant Professor at UNU-MERIT Maastricht University), together with Yannick Markhof (ETH Zurich), Rose Vincent Camille (Utrecht University), and Firminus Mugumya (Makerere University). The project is implemented in partnership with the Kampala Capital City Authority (KCCA) and has received financial support from the International Growth Center, the Nederlandse Organisatie voor Wetenschappelijk Onderzoek (NWO), and J-PAL.
The platform addresses a fundamental challenge in public administration and evaluation: traditional impact evaluations are costly and produce results months after data collection, making real-time course correction difficult. By enabling continuous, bidirectional communication with citizens via text and WhatsApp, the platform aims to put real-time data at the fingertips of policymakers while testing behavioral interventions at marginal cost.
Why was this approach necessary?
Traditional impact evaluations face two persistent limitations:
- Timing: Evaluators wait months for endline survey results, but policy decisions often require real-time course correction.
- Cost: Impact evaluations are expensive, limiting the number of interventions that can be tested and scaled.
The research team asked: Wouldn't it be great if we could establish a platform that obtains real-time data for decision making at the fingertips of policymakers, tests information treatments and nudges at marginal cost, and connects policymakers more closely with constituents? The answer is the conversational AI platform developed with KCCA.
How does the platform work?
Two-way knowledge exchange architecture
The platform is built on a web interface connected to Twilio, enabling communication via text messages and WhatsApp. Key features include:
- Multi-channel communication: Accepts text and audio messages, supporting multilingual interactions.
- Real-time processing: Incoming messages are transcribed, translated, and classified automatically.
- LLM-powered responses: Different Large Language Models are used for various service modules, with both local and API model options.
- Agentic capabilities: The system can perform actions like changing phone numbers or adjusting default settings autonomously.
- Local deployment: LLMs run locally to protect sensitive citizen data.
The platform enables natural, conversational interactions rather than rigid survey flows.
What AI methods and tools are used?
🗣️ Speech-to-Text
Automatic transcription of voice messages enables citizens to respond in their preferred language and format.
🌍 Machine Translation
Real-time translation supports multilingual communication across Uganda's diverse linguistic landscape.
📋 Message Classification
Incoming messages are automatically categorized to route them to appropriate service modules or trigger specific responses.
🤖 Local LLMs
Different LLMs power service modules, running locally to protect sensitive citizen data while enabling sophisticated interactions.
⚙️ Agentic Workflows
The platform includes agentic capabilities that can perform actions like updating contact information or adjusting user preferences.
🧪 A/B Testing Framework
Built-in experimentation capabilities allow testing of different message treatments and behavioral nudges at scale.
Key applications and use cases
1. Information exchange and service delivery
The platform enables citizens to ask questions and receive immediate, accurate responses about public services. For example:
2. A/B testing and behavioral nudges
The platform allows the public administration to run randomized experiments on message framing and behavioral interventions at marginal cost. Example:
- Version A (Enforcement frame): "Dear esteemed client. Renew your trade license to avoid enforcement action and/or prosecution."
- Version B (Community frame): "Dear esteemed client. Help your community grow! Renew your trade license."
The platform automatically tracks responses to each version, enabling rapid analysis of which messages are most effective for different citizen segments.
3. Complementing traditional surveys
Beyond real-time engagement, the platform can serve as a cost-effective mechanism for gathering follow-up data, complementing traditional baseline and endline surveys with continuous feedback loops.
Research design and evaluation framework
The team has designed a three-pronged research approach to rigorously evaluate and optimize the platform:
📱 1. Uptake optimization
Test different designs and recruitment strategies to maximize platform uptake among diverse citizen populations.
🎯 2. Preference elicitation
Use conversational interactions to elicit citizen preferences and priorities, informing evidence-based policy design.
📊 3. Impact evaluation
Conduct traditional impact evaluations to measure the effects of the platform itself on citizen engagement, service satisfaction, and policy outcomes.
What are the challenges and considerations?
The team has been transparent about the complexities of deploying AI-powered citizen engagement platforms in public administration contexts:
🔒 Data privacy and security
Running LLMs locally rather than in the cloud protects sensitive citizen data, but requires robust infrastructure and technical capacity.
🌐 Linguistic diversity
Uganda is home to over 40 languages. The platform's translation capabilities must handle this diversity accurately and equitably.
📱 Digital access
Not all citizens have access to smartphones or reliable internet. The platform uses SMS as a fallback to maximize inclusivity.
🤖 Conversational quality
Ensuring that LLM responses are accurate, culturally appropriate, and genuinely helpful requires continuous refinement and human oversight.
🏛️ Institutional integration
Integrating the platform into existing public administration workflows requires change management and sustained institutional commitment.
⚖️ Ethical experimentation
A/B testing with citizens raises ethical considerations around informed consent and the responsible use of behavioral insights.
Current status and next steps
Current status
The platform is currently being piloted with the Kampala Capital City Authority. An endline survey is planned for Q4 2026 to evaluate the platform's impact on citizen engagement and service delivery outcomes.
New applications
The team is actively co-developing additional use cases with partners and welcomes collaboration inquiries.
Long-term plans
- Channel integration: Seamlessly connect in-person interviews, text messages, and other communication channels into unified conversation flows without losing context.
- Backend infrastructure: Develop tamper-proof data pipelines that make the entire flow from citizen response to policy results auditable, building trust without compromising privacy.
Key lessons for evaluators and M&E practitioners
📱 Real-time data transforms evaluation
Conversational AI platforms can provide continuous feedback loops, enabling course correction and adaptive management that traditional endline surveys cannot support.
🧪 A/B testing at marginal cost
Digital platforms dramatically reduce the cost of experimenting with different interventions, allowing more rigorous testing of what works.
🔒 Privacy requires intentional design
Running LLMs locally rather than in the cloud demonstrates that AI innovation and data protection can—and should—go hand in hand.
💬 Conversation beats digitized surveys
Moving beyond rigid survey flows to truly conversational interactions increases engagement and reveals insights that structured questions may miss.
🤝 Partnerships are essential
Successful deployment requires deep collaboration between researchers, public administration, and technology partners—not just a technology solution.
📊 Evaluators as platform designers
This case shows how evaluators can move beyond assessing interventions to designing the data infrastructure that enables continuous learning.
Frequently asked questions
What makes this platform different from traditional survey tools?
How does the platform protect citizen privacy?
How can the platform be used for evaluation?
What are the technical requirements for deployment?
What is the current status of the project?
Main reference & original sources
📘 This case study is based on the presentation by Stephan Dietrich, Yannick Markhof, Rose Vincent Camille, and Firminus Mugumya, as documented in:
1. WFP Evaluation (2026, March 23). Four examples of how AI, machine learning and data innovation are reshaping evaluation and research. Medium. https://wfp-evaluation.medium.com/four-examples-of-how-ai-machine-learning-and-data-innovation-are-reshaping-evaluation-and-research-f4d1b3cf4e74
2. Dietrich, S., Markhof, Y., Camille, R. V., & Mugumya, F. (2025, December). From Surveys to Conversations: AI-Powered Citizen Engagement. Presentation slides. WFP Global Impact Evaluation Forum 2025. Download full presentation (PDF)
Resources for further learning
- WFP Evaluation – Medium Blog
- EvalCommunity – AI in M&E Course
- Download: From Surveys to Conversations (PDF)
- EvalCommunity Services & Resources
Advance your skills in AI-enhanced evaluation
Learn how to integrate conversational AI, real-time data collection, and A/B testing into your own M&E practice. Access courses, templates, and expert guidance.
