To conceptualise and demonstrate an AI-driven approach for systematically integrating climate change considerations into evaluation frameworks in Africa.

Traditional development evaluation has historically operated on linear logic models – assuming that if project inputs are delivered (e.g. building a school), outcomes will follow (e.g. increased attendance). However, foundational authors in the field argue that this linearity fails to account for complexity. Patton (2019) notes that in the Anthropocene, environmental externalities are no longer ‘background noise’ but central drivers of success or failure.

In the context of Africa, Tamasiga et al. (2023) argue that the global climate finance architecture often misinterprets vulnerability because of a lack of granular data. Standard evaluation cycles, which often occur years after a project’s conclusion, are too slow to capture the rapid onset of climate shocks like cyclones or flash droughts. Consequently, evaluators often struggle to distinguish between a project failing because of poor design versus failing as a result of an overwhelming climate shock – a distinction that is vital for accountability.

The Fourth Industrial Revolution has introduced new tools to address these analytical gaps. Nalubega and Uwizeyimana (2019) provide a foundational analysis of how the Fourth Industrial Revolution impacts public sector M&E in Africa. They contend that while AI offers immense potential for real-time feedback, its adoption remains limited by infrastructure and capacity constraints.

Scholarship distinguishes between two primary applications of AI in evaluation:

Olawade et al. (2024) demonstrate that ML models can process complex environmental datasets to identify pathways to net-zero sustainability that manual analysis would miss. This suggests that AI is not merely a tool for efficiency, but a mechanism for uncovering non-linear relationships between environmental variables and development outcomes.

Specific to climate change, ML has emerged as a potent tool for meteorological prediction and impact assessment. Kenne et al. (2024) utilised AI to predict sub-seasonal summer temperatures in West Africa, demonstrating that algorithmic models often outperform traditional meteorological forecasts in data-scarce regions.

However, a gap remains in linking these meteorological insights to social indicators. While meteorologists use AI to predict storms, and social scientists use surveys to measure education or health, few studies integrate these domains. York and Bamberger (2020) argue for the use of ‘Big Data’ in evaluation to bridge this divide, suggesting that satellite imagery and remote sensing can serve as proxies for ground-truth data in conflict or disaster zones. By triangulating climate data (e.g. rainfall anomalies) with social data (e.g. school attendance), evaluators can construct ‘counterfactuals’ that isolate the specific impact of climate shocks on project performance.

The application of these advanced technologies in Africa presents a unique paradox. On one hand, the continent faces significant ‘data gaps’. On the other hand, the rapid digitalisation of services (e.g. mobile money, Short Messaging Service [SMS] surveys) generates novel datasets ripe for AI analysis.

Recent studies highlight the necessity of localised algorithms. Generic global models often fail to account for the specific socio-economic dynamics of African rural communities. For instance, Global Pulse (2023) highlighted how AI models trained on Western data failed to predict food security trends in East Africa accurately. This validates the argument by Tamasiga et al. (2023) that African-led data architectures are essential for redefining climate vulnerability.

To operationalise these insights, this study grounds its methodology in design science research (DSR). Unlike natural science, which seeks to explain phenomena, DSR seeks to create artefacts that solve human problems (Hevner et al. 2004). The study conceptualises the AI algorithm not just as a statistical tool, but as a ‘socio-technical artefact’ designed to enhance the adaptive capacity of evaluation systems.

While the literature establishes the separate value of AI in evaluation (Nalubega & Uwizeyimana 2019) and AI in climate modelling (Kenne et al. 2024), there is a paucity of research that combines them into a unified evaluation workflow. Few studies explicitly demonstrate how an evaluator can use ML to ‘climate-adjust’ the performance ratings of social projects. This research aims to fill that gap by proposing and simulating a specific algorithmic framework for this purpose.

The study uses a DSR and proof-of-concept approach to develop and refine an AI-driven evaluation algorithm. Design science research is appropriate because the purpose is not only analytical but also solution-oriented, producing an artefact (the conceptual algorithm) intended for practical application in real-world development evaluations.

The study is structured into three components:

The algorithm is intended for development programmes in Africa, where climate sensitivity is high and data ecosystems vary widely. The design emphasises regional climate conditions, programmatic data structures and evaluation needs relevant to the continent. The integration operates through a Climate-Evaluation Data Fusion Layer, which merges evaluation indicators and climate variables into a unified analytic pipeline.

This study employed a mixed computational approach combining simulated climate data and ML modelling to empirically demonstrate how AI can strengthen climate-responsive evaluation systems. The methodology was designed to approximate Kenya’s real climate–education dynamics using authoritative distributions from the Copernicus Climate Change Service (2023) rainfall climatology, ECMWF Reanalysis 5 (European Centre for Medium-Range Weather Forecasts Reanalysis, version 5) (ERA5) reanalysis, World Bank EdStats (2023) and United Nations Children’s Fund (UNICEF) Multiple Indicator Cluster Surveys (MICS) patterns (UNICEF 2015).

The proposed Climate-Informed Early Action (CIEA) framework operates on a three-tier processing logic designed to overcome the static nature of traditional evaluation methods (see Figure 1).

A monthly dataset for counties in Kenya (Kenya Ministry of Education 2021) over 24 months. Each row represents a county–month observation and contains:

Climate shocks were simulated using probability distributions calibrated to East African rainfall seasonality (long rains March–May, short rains October–December), ensuring realism consistent with CHIRPS (Copernicus Climate Change Service 2023) and ERA5 spatial structure (CHIRPS 2023):

To capture short-term and cumulative climate effects, the following engineered variables were added:

Machine learning modelling is the process of using algorithms to automatically learn patterns from data, then building and refining a predictive or decision-making model that can generalise those patterns to make accurate predictions or classifications on new, unseen data (see Figure 2).

Two models were trained:

The Random Forest captures non-linear relationships between climate shocks and attendance.

Contrary to ‘black box’ deep learning approaches, this study selects Ensemble Learning techniques for their interpretability and robustness with small datasets, which are common in development projects:

To quantify climate-attributable education loss, a counterfactual scenario was generated by resetting all climate-shock variables to neutral (zero-anomaly):

The difference between predicted actual attendance and counterfactual attendance represents the climate-attributable impact.

This study introduces a novel evaluation innovation: the Counterfactual Climate Impact Model (CCIM).

Unlike traditional evaluations – which rely on historical comparisons or non-experimental designs – CCIM uses ML to generate synthetic counterfactual programme outcomes under climate-neutral conditions. This allows evaluators to:

This positions AI as a direct contributor to Sustainable Development Goal (SDG) 4 (education) and SDG 13 (climate action) by enabling real-time, climate-aware evaluation frameworks for governments in the Global South.

This study involved the development of an AI algorithm. As such, it did not involve direct interaction with human subjects or the collection of primary data from individuals. However, the study did consider several ethical considerations related to the potential development and deployment of such an algorithm in the future:

While this study focused on the conceptual development of the algorithm, these ethical considerations were central to its design and will need to be thoroughly addressed in any future implementation.

The Random Forest demonstrated strong predictive performance:

This substantially outperformed the linear regression baseline:

These results indicate that climate–attendance relationships are non-linear, validating the use of ML for climate-responsive evaluation.

The top predictors of attendance were:

The dominance of drought and flood signals confirms the high sensitivity of educational participation to weather extremes, aligning with climate vulnerability patterns in ASAL (Arid and Semi-Arid Lands) counties.

Using the CCIM:

This shows that climate variability significantly depresses school attendance in roughly half of all months nationally.

A K-Means clustering model identified three climate-risk county groups:

In Table 1, traditional evaluations typically rely on static indicators, periodic data collection and linear analyses. These methods often overlook non-linear climate interactions and struggle to account for rapidly changing environmental conditions.