NOVISSI Togo: How AI Transformed Social Protection Monitoring and Evaluation
NOVISSI Togo: Harnessing Artificial Intelligence to Deliver Shock-Responsive Social Protection
During the COVID-19 pandemic, Togo faced a rapid increase in poverty and vulnerability. Traditional social protection systems — slow, limited in coverage, and dependent on outdated survey data — failed to reach newly vulnerable populations. In response, the Government of Togo launched NOVISSI, a fully digital cash transfer system powered by machine learning. Using satellite imagery, mobile phone data, and geospatial analysis, the system reached over 920,000 beneficiaries (25% of the adult population) in record time, demonstrating how AI can transform Monitoring and Evaluation from a retrospective reporting function into a real-time, predictive, and decision-driven system.
Context: A Crisis That Required Real-Time Monitoring
The COVID-19 pandemic drove reversals in global poverty reduction in 2020, with an estimated 40 million new poor in Sub-Saharan Africa alone. Togo, a country of 8.3 million people in West Africa, was no exception. With a poverty incidence of 45 percent and more than 85 percent of the labor force working in the informal sector, the country was exceptionally vulnerable to the economic shocks of lockdowns and reduced economic activity.
The pre-existing gap:
Only 38 percent of the poor in Togo were covered by social protection before the pandemic. The existing systems were slow, limited in coverage, and dependent on outdated survey data that could not capture the rapidly changing reality of newly vulnerable populations. The informal sector — smallholder farmers, street vendors, tailors, hairdressers, and casual laborers — fell through the cracks of existing programs.
Traditional social protection systems — designed for stable, predictable conditions — failed to adapt to the shock. They could not identify which households had fallen into poverty overnight due to job losses, business closures, or illness. This created a classic Monitoring and Evaluation failure: decisions were being made without timely or accurate data.
The Core Problem: A Monitoring and Evaluation Failure
From a Monitoring and Evaluation perspective, the key challenges facing Togo's social protection system were severe and systemic:
Data on household conditions was collected through periodic surveys that could not capture rapid changes during the crisis.
The existing registry was static, unable to add newly vulnerable households or remove those whose circumstances had improved.
Most data was aggregated at regional or national levels, insufficient for targeting individual households.
Traditional enumeration and verification processes were too slow for emergency response.
The result was a system that could not answer the most basic M&E questions: Who is poor right now? Where are they? How many of them are there? And how can we reach them quickly?
The AI Solution: NOVISSI — A Data-Driven Targeting System
To respond to the crisis, the Government of Togo launched NOVISSI (which means "solidarity" in the local Éwé language) on April 8, 2020 — just one month after the country's first reported case of COVID-19. NOVISSI is a large-scale unconditional emergency cash transfer program designed from scratch in ten days and delivered digitally from end to end.
Two Complementary Delivery Models
NOVISSI's deployment consisted of two complementary delivery models. Model 1 (low-tech approach) initially prioritized informal workers in urban areas under a declared state of emergency, leveraging the voter list to verify applicants' uniqueness, location, and occupation. Model 2 (high-tech approach) was developed for rural expansion, leveraging satellite imagery, mobile phone data, and machine learning to prioritize the poorest cantons and the poorest individuals within them.
AI Component 1: Machine Learning for Poverty Mapping (Geographic Targeting)
The first challenge was identifying which geographic areas had the highest concentration of poverty. Traditional survey data was too slow and too sparse. Instead, the NOVISSI team turned to satellite imagery and geospatial data.
How it works
Machine learning models were trained to analyse satellite images and detect features correlated with wealth and poverty: housing materials (roofing materials, road surfaces), infrastructure (proximity to water, electricity access), and land use. Using the EHCVM 2018-2019 household survey as ground truth, the model generated consumption estimates at 2.4 km² grid cell resolution. These micro-estimates were combined with population density data to rank Togo's 397 cantons from poorest to richest. The 200 poorest cantons were selected for NOVISSI's rural expansion.
AI Component 2: Mobile Phone Data for Individual Targeting
Geographic targeting identified poor areas, but individual households within those areas still needed to be assessed. To achieve individual-level targeting at scale, NOVISSI analysed Call Detail Records (CDR) from mobile phone networks.
How it works
A phone-based survey was administered to a representative sample of mobile phone subscribers in the poorest cantons. Their responses (including consumption data) were matched to their historical mobile phone usage patterns — call frequency, SMS usage, mobile money activity, social network characteristics, and mobility patterns. A supervised machine learning algorithm (gradient boosted regression tree) was trained to recognize patterns of poverty in CDR data. The resulting model estimated average daily consumption for each of Togo's 5.83 million mobile phone subscribers.
AI Component 3: Multi-Source Data Integration
NOVISSI's power comes from integrating multiple data sources into a unified AI-driven decision system:
Voter registry for identity verification (86% coverage of adults)
EHCVM 2018-2019 for ground truth and model training
Satellite imagery and mobile phone CDRs
Mobile USSD system (*855#) for self-enrollment
M&E Integration: NOVISSI as a Full AI-Powered M&E System
NOVISSI is not just a social protection program — it is a complete, AI-powered Monitoring and Evaluation system that integrates monitoring, evaluation, learning, and decision-making into a single, real-time platform.
Monitoring (Real-Time)
Real-time registration via the *855# mobile system. Continuous data collection from mobile phones, administrative systems, and big data sources. Instant tracking of enrollment, payments, and beneficiary demographics. A dashboard provided real-time analytics for program administrators.
Evaluation (Continuous)
AI models are continuously validated using survey data (ground truth) and consumption estimates. Phone-based surveys were conducted to evaluate the impact of both Model 1 and Model 2. The system constantly assesses who is poor, who should receive support, and whether targeting is accurate.
Learning (Adaptive)
Machine learning models improve over time as new data becomes available. The system refines its targeting algorithms based on past performance and updated information. The methodology was replicated for an additional 100 cantons in the second phase, demonstrating adaptive learning.
Decision-Making (Automated)
Automated eligibility determination based on AI predictions. Instant enrollment and payment delivery. Real-time adjustments to targeting as conditions change. The delivery chain is seamless: Outreach → Intake and Registration → Assessment → Enrollment → Provision → Management.
Outcomes and Impact
The NOVISSI system achieved remarkable results in a short period:
M&E Transformation: Before vs. After
Challenges and Risks (Critical for Credibility)
The World Bank case study acknowledges several important limitations and risks of the NOVISSI model:
Risk of exclusion
The system relies heavily on mobile phone ownership and network connectivity. While 85% of households have at least one mobile phone, people without phones — often the very poorest and most marginalized — risk being excluded. Additionally, 22% of SIM cards and 7% of SIM slots are shared, complicating individual identification.
Data privacy concerns
The use of Call Detail Records (CDR) and mobile money transaction data raises significant privacy and surveillance concerns. To mitigate this, researchers hash-encoded each phone number into a unique identifier, stored all data on secure servers, and never transmitted CDR data to the Government of Togo. Only specified researchers accessed the data under non-disclosure and data use agreements.
Individual vs. household assistance unit
NOVISSI's assistance unit was individual informal sector workers, not households. In the absence of a dynamic social registry with up-to-date household data and unique identifiers with universal coverage, some households may have received multiple transfers. This was a pragmatic policy choice given data constraints.
Model limitations
The machine learning models were trained on data from 2018-2019 and may not fully capture the rapid changes in poverty during the pandemic. Additionally, the approach required significant technical expertise and partnerships with academics and MNOs that may not be replicable in all contexts.
Key Lessons for M&E Professionals
AI can replace slow, survey-based M&E
Traditional surveys cannot keep pace with rapidly changing conditions during shocks. AI models using satellite and mobile data can provide near real-time poverty estimates at scale, enabling faster targeting and response.
Monitoring can become real-time and automated
NOVISSI demonstrates that continuous, automated monitoring is possible even in low-resource settings, using data already generated by mobile networks, satellites, and administrative systems.
Evaluation can become predictive
Instead of evaluating outcomes after programs end, AI enables continuous predictive assessment of who is poor and who should receive support — shifting from retrospective to prospective evaluation.
Decision-making can be instant and data-driven
Automated eligibility determination and payment delivery are possible when M&E systems are integrated with program operations — collapsing the gap between data collection and action.
From measuring the past to shaping the future
NOVISSI demonstrates how AI can transform M&E from a retrospective reporting function into a real-time, predictive, and decision-driven system. The traditional M&E model — measure outcomes after the fact, report them months later, and hope that lessons inform future programs — is no longer the only option. AI enables a new paradigm: continuous monitoring, predictive evaluation, and instant, data-driven action. As the World Bank case study concludes, digital approaches such as the one introduced by NOVISSI have the potential to introduce significant gains in terms of efficiencies for governments and financial inclusion.
Frequently Asked Questions
How accurate was the AI poverty prediction compared to traditional surveys?
The AI model was trained and validated using ground truth data from the EHCVM household survey. The methodology had been previously tested theoretically and was replicated for the first time in a real cash transfer delivery context in Togo. The model performed favourably against other feasible targeting approaches, with significant gains in reaching the poorest populations.
What happened to people without mobile phones?
This remains a critical limitation. While 85% of households have at least one mobile phone, those without phones — often the very poorest — risk exclusion. The World Bank case study acknowledges this as a key risk of the model. Complementary mechanisms would be needed to reach these populations.
How was data privacy protected?
To protect confidentiality, researchers hash-encoded each phone number into a unique identifier before analysis and stored all data on secure servers. Call-detail records were never transmitted to the Government of Togo. Only specified researchers accessed them under non-disclosure and data use agreements signed with each mobile network operator.
Can this model be replicated in other countries?
Yes. The methodology has been replicated in other contexts. However, replication requires strong data governance frameworks, mobile network infrastructure, political commitment, and partnerships with academics and MNOs. The World Bank case study notes that the approach was possible in Togo due to high mobile phone penetration (83.6% SIM card penetration) and near-universal mobile coverage.
A New Paradigm for M&E in Social Protection
NOVISSI Togo represents a breakthrough in how Monitoring and Evaluation can be integrated with social protection delivery. By embedding AI and real-time data into the core of program operations, Togo demonstrated that it is possible to identify vulnerable populations, target assistance, and deliver support at unprecedented speed and scale.
For M&E professionals, the lesson is clear: AI can transform M&E from a retrospective reporting function into a real-time, predictive, and decision-driven system. The question is no longer whether we can collect data, but whether we can use it fast and intelligently enough to change outcomes.
From measuring the past to shaping the future — that is the promise of AI in Monitoring and Evaluation.
