Agentic AI is arriving in development cooperation before the sector has the means to govern it. Unlike the chatbots most practitioners have used, agentic systems take actions and coordinate on their own by calling other tools and stringing together steps to complete a task. When the OECD DAC Governance Network meets this month to discuss AI, the hard question to answer is how to adopt and govern such a rapidly evolving and increasingly autonomous technology. As the top donor among DAC countries, Germany should play a leadership role in the development sector’s technological transformation. The Federal Ministry for Economic Cooperation and Development (BMZ) can host a pooled technical capacity that experiments with AI tools and sets accountability standards. What it needs is policy willingness and an institutional framework that treats AI as a moving technological frontier.

The development cooperation sector is not ready for agentic AI

The pace of change is the first problem. In only eighteen months, the technological frontier has moved substantively from chatbots that retrieve information to systems that can generate data and act across multiple tools without supervision. Stanford’s AI Index 2025 documents that the cost of running these models has also fallen more than 280-fold over the same period, with lowered barriers to advanced AI. With innovation cycles measured in weeks, the development cooperation sector is at odds with this pace, with long procurement timelines and rigid bureaucracy. 

This mismatch creates a capacity problem. Development organisations are struggling to simultaneously track the AI innovations and build tools for their programming, while closing internal skills gaps. The likely outcome would be individual staff experimenting on the side, along with small in-house units using whichever commercial model is cheapest.

There are two implications of this capacity constraint. First, it will impede scalability. An OECD overview of existing AI use cases suggests that most pilots stay in the exploratory phase and rarely scale. Second, haphazard AI applications will lead to fragmentation with incompatible standards and create an accountability gap across the sector.

The deeper constraint is data. AI is only as capable as the information environment it draws on. In fragile states and digitally under-connected communities, where, for example, many under-resourced African languages barely appear in training data, agentic systems will produce misleading outputs, with grave implications for development programming.

BMZ’s leadership role

BMZ has done some groundwork that can be leveraged for sectoral reform. The BMZ Data Lab has developed pipelines and a ministerial chatbot (KIEZ). FAIR Forward has worked with seven partner countries on open AI data and policy, and produced insightful resources. The BMZ’s new reform plan also cites AI in three places, including the establishment of a new working hub for strategic foresight. But the plan still treats AI as an instrument to be deployed, not as a moving frontier that needs continuous institutional capacity to track.

BMZ could establish a team that serves as a sector-wide AI capability. Consider an agile, fast-moving unit of engineers and development practitioners working to pool resources across different stakeholders, develop protocols, and stress-test cutting-edge AI tools for donors and development organisations for safe yet faster adoption. Such an AI team would have a broader mandate than just serving the ministry internally. The team would work on shorter contracting and procurement cycles than the sector typically allows, with clauses that let tools evolve as the technology does. Smaller donors and other development organisations would tap the same resource through co-development, joint training, and open access to the tools and protocols it produces.

This kind of unit answers two problems at once. BMZ would be a central node for pooling resources together for joint capacity within the sector. And it can generate sector-specific algorithmic accountability standards, for example, mandatory decision-audit trails for any agentic AI system. These tools, in turn, will strengthen the BMZ to actively shape the normative discussions for the use of AI in development cooperation globally.

The same unit can also tackle the data problem. It could create guidelines for “data readiness assessments”, as a standard prerequisite for any AI-assisted programming (akin to environmental impact assessments). It can also help turn decades of tacit institutional knowledge within the sector into publicly available, AI-assisted, and actionable knowledge (going beyond KIEZ) that could, first, give more visibility to digitally under-connected communities, and second, empower practitioners globally. Instead of a 100-page static PDF toolkit, imagine an interactive system that a practitioner can ask: “What worked in contexts similar to mine?” and receive synthesised, sourced answers.

Germany should not miss this narrow window of opportunity to be the leading voice in AI transformation for development cooperation.

Artificial intelligence (AI) presents numerous challenges and opportunities for development cooperation. This column focuses on how AI can help development organisations become learning organisations. Whether it does so, however, depends on how AI-supported analysis and learning are embedded in organisations. Paradoxically, the US development agency USAID, which was closed in July 2025, is an interesting example of how and what can be learned for development cooperation with AI – and what its limits may be.

USAID’s aborted AI learning initiative

After many years as a USAID staff member, the second author of this column left the organisation in 2021 to study how development organisations can learn from their own experiences more systematically. This work resulted in the Development Evidence Large Language Model (DELLM), an AI system developed specifically for development cooperation. USAID became one of the first clients. To train DELLM, experts first manually coded countless text segments from the approximately 100,000 USAID evaluation reports, spanning more than 60 years. Increasingly automated but guided and supervised by experts, the model learned to ‘read’ these reports and distinguish between sectors, cooperation approaches, outcomes, positive and negative lessons and other aspects.

Once training was complete, USAID staff had the organisation’s vast institutional knowledge at their fingertips. DELLM could be used for operational planning as well as for organisational learning. In terms of operational planning, staff were now able to use the model to access subject-specific and regional insights from over six decades within minutes. Previously, each of these queries would have required months of research and incurred high costs. Due to the sudden closure of USAID in 2025, however, the model could not be further used for organisational learning. As long as evaluation reports were still internally accessible, DELLM was instead used to analyse and preserve insights from the organisation’s past. The results from this search for the most important overall success factors of USAID interventions are insightful in themselves. But they also suggest why translating AI-supported analysis into organisation learning is not straightforward.

Success factors for development cooperation

The search for the most important overarching success factors of USAID development interventions yielded the following insights: (1) Decision-making should take place as closely to the ground as possible in order to receive rapid implementation feedback and be able to adjust course. (2) Reforms should be practical, i.e. build on existing systems and seek to adapt them. (3) Effective solutions should be rapidly embedded locally so that they continue to work even after the end of funding. (4) Reform processes should be led by local actors rather than merely involving them in a consultative capacity. Finally, (5) collaboration with the middle tier is crucial – that is, with partners who are responsible for the practical implementation of measures. What do these five lessons tell us about the potential of AI-supported organisational learning for development cooperation and about its limits?

First, these insights are not new to development effectiveness experts. At the same time, approaches to development cooperation that are based on them are still rarely used. One example are approaches to adaptive development cooperation. Despite having been successfully implemented in some cases, including in German development cooperation, they remain exceptions. At the political level, systematic organisational learning is hampered by development policy’s focus on furthering national interests and by accountability concerns. But it is also held back by the structures and incentives within development ministries and agencies. Organisational learning is unlikely to happen unless evidence-based work is valued and encouraged, generalist career tracks are complemented by specialist ones, objectives and shortcomings are openly discussed and the organisation’s culture supports all that. 

Embedding AI-supported learning for development

AI has the potential to provide access to an organisation’s institutional knowledge in a comprehensive, intuitive and needs-based manner. This is particularly useful for operational planning. But in terms of organisational learning and reform, the specific advantages of AI and the necessary enabling conditions should be kept in mind. First, like for all AI-tools, the quality of their output is only as good as the quality of the underlying data and their training. Second, to translate insights into organisational learning, they have to be embedded in an open, self-critical and evidence-based organisational culture. Third, while AI-supported insights may not be entirely new, together with other sources of knowledge such as development research, they can serve to review and complement existing knowledge. In a conducive organisational environment, AI can therefore be an important tool among others to help transform development organisations into learning organisations and increase their effectiveness and efficiency.

Dr Michael Roll is a sociologist and senior researcher in the research department “Transformation of Political (Dis-)Order” at the German Institute of Development and Sustainability (IDOS) in Bonn.

Lindsey Moore is the CEO and Founder of DevelopMetrics and adjunct professor of AI and Policy at Georgetown University. Her work focuses on responsible artificial intelligence, evidence synthesis, and knowledge management for international development.

The transfer of development finance rests, primarily, on two pillars: first, human and organisational expertise, and, second, financial flows intermediated through trusted organisations. Today, both elements are being profoundly reshaped simultaneously by disruptive innovations - expertise by AI and financial flows by blockchain-based payment and smart contract solutions.

Traditionally, development finance depends on the expertise of professionals in development banks, international organisations, ministries, and other local partners who design and evaluate projects, assess risks, and allocate know-how and capital, based on accumulated knowledge and (often imperfect) data.

First, by processing vast datasets, ranging from satellite imagery to real-time economic indicators, AI systems can generate insights that surpass human analytical capacity in speed and scale. For development finance institutions (DFIs), this creates both an opportunity and a challenge: they must transition from being repositories of expertise to becoming orchestrators of AI-enabled knowledge systems.

At the same time, the second pillar of development finance, financial flows, is undergoing an equally profound transformation. Up to now, capital flows in development finance have largely been intermediated through centralised institutions. Blockchain technology disrupts this paradigm by enabling decentralised, transparent, and immutable financial infrastructures. Through smart contracts, funds can be disbursed automatically when predefined conditions are met.

The most profound transformation emerges from the interaction between these two disruptions. AI reshapes how decisions are made, while blockchain reshapes how those decisions can be executed. Together, they create a new paradigm of development finance, where capital allocation and disbursement become - when the right checks and balances are in place - automated, data-driven and verifiable.

From pilots to scale: Successful good practices around the globe

Highly successful pilots around the globe are allowing a glimpse into this near future: For instance, the Central Bank of Brazil has been testing decarbonisation tokens. When satellite data confirms a reduction in CO₂ emissions, these are automatically paid out to smallholder farmers.

In Syria, German international cooperation has launched a successful pilot initiative to digitise salary payments for healthcare workers. Years of political instability have significantly weakened the country’s banking system. Traditional transfer methods are frequently linked to corruption risks, burdensome administrative procedures, lengthy delays and high transaction costs. Payments to more than 900 health professionals via a stable digital currency have proven to be a highly efficient alternative. Moreover, between 2022-2025, UNHCR has supported over 240,000 people, who were forced to flee, with blockchain-based payments, thereby increasing the speed, efficiency and transparency of aid delivery.

How development finance institutions can adapt

Crucially, the dual disruption does not diminish the importance of DFIs, it elevates it. DFIs must govern data, ensure accountability and regulate decentralised systems. This will fundamentally alter the development finance architecture.

Against this backdrop, we make three observations. In terms of organisational design, DFIs will evolve towards platform-based orchestrators connecting data ecosystems, stakeholder and partnership management and decentralised finance infrastructures. Second, human resource management in DFIs is likely to focus on a new skill profile that combines AI for routine tasks with human experience for nuanced-decision-making skills and multidisciplinary connective capacity. Third, DFIs will need to invest in partnership ecosystems and turn increasingly into platform hubs to co-create solutions with tech firms and data providers.

To conclude, development finance is entering a new era defined by a double disruption and this comes, of course, with risks. To name only some: AI can produce biases and hallucinations. This makes human-centered AI (or augmented intelligence) all the more important. When evaluating any blockchain solution (e.g. crypto currency, stable coin, central bank digital currency, smart contract), it is essential to look closely at the entire creation and governance chain: who designed and developed the protocol, who issues the assets built on it, who operates and maintains the network, what externalities are (e.g. energy use) and, ultimately, who uses it.

But we believe the outlook of the digital disruption is positive – if we prepare well: There are substantial gains in terms of efficiency and transparency to be made.

Prof. Clara Brandi is Head of Department at the German Institute of Development and Sustainability (IDOS) and Professor of International Economics / Development Economics at the University of Bonn.

Dr Max Büge is a leading expert for AI and blockchain in finance. At GIZ, he manages an international project on digital finance for sustainability.

The opinions expressed are those of the authors and do not necessarily reflect those of their organisations.

Dr Tomy Ncube is a postdoctoral researcher affiliated with the Centre for International Development Innovation at the Ryan Institute, University of Galway, and the School of Geography, Archaeology and Irish Studies.

This week, representatives from academia, government and industry seek to find solutions to decarbonise the built environment at the Sustainable Buildings and Construction Summit in Lausanne. The heat is on: the buildings sector emits more than a third of global CO₂ annually and urbanisation in many low- and middle-income countries is soaring, especially in Sub-Saharan Africa (SSA). The key question is whether solutions will be fast, feasible and deliver economic co-benefits. Much of the international debate highlights low-carbon concrete such as LC3 as a quick fix requiring minimal adjustments to business as usual. Yet its development potential is limited, which might be different when adding bio-based materials to the construction mix.

Many countries in SSA face a triple challenge: First, providing adequate housing for a growing urban population, many with low incomes. Africa already faces a deficit of about 50 million housing units, expected to reach 130 million by 2030. Second, creating jobs and increasing firm productivity are key priorities for policymakers. In Kenya, for example, 1 million youth enter the job market every year. Third, pursuing economic development along a greener pathway than the carbon-intensive trajectories of today’s advanced economies. Infrastructure and buildings make this choice especially consequential, as carbon-intensive construction locks in emissions, raises urban heat and limits re-use of materials. As much of SSA’s built environment is yet to be built, there is a chance to avoid costly retrofits, which needs to go hand in hand with needed shifts in the construction sector in the Global North.

The switch to low-carbon concrete may meet only two of the three challenges in a fast and feasible way: providing housing and decarbonising construction LC3 requires minimal changes in cement factories and is applied the same by masons. Modelling shows it is particularly suitable for multi-storey buildings, where its carbon footprint is lower than for a mix of bio-based materials and fired clay bricks. For these buildings, LC3 makes sense. Yet, many homes in SSA are and will remain single-storey, where emissions depend heavily on material composition and transport distances.

Bio-based materials as a solution to SSA’s triple challenge have so far been largely neglected. This is significant, as improved adobe blocks, timber or bamboo offer not only lower-cost but also more labour-intensive alternatives to LC3, generating stronger local economic benefits and supporting firm development. Improved adobe or interlocking soil blocks can often be produced at or near the construction sites, reducing transportation needs and enabling local SMEs to participate. Both are well suited for affordable housing: research in Rwanda shows wall costs can fall by 60% or more, while improved production standards ensure comparable strength and durability. Our research also finds that bio-based materials can create greener, better-quality jobs in Kigali. However, wider adoption faces supply- and demand-side barriers, including limited awareness and perceptions of low quality that restrict financing. 

This points to a hybrid approach that must be clearly defined: LC3 for multi-storey buildings and structural uses, and bio-based materials for single-storey housing and non-load-bearing interior walls, where they are often more affordable and locally appropriate. In Rwanda, clear government direction is needed for business adoption. Across Sub-Saharan Africa, similar hybrid approaches are promising, but scaling bio-based materials requires locally grounded research. Suitable materials depend on a) type of soil and materials locally available, b) import dependence on cement, steel, etc. c) transport and other lifecycle costs. Sustainable timber and mycelium may suit Europe and bamboo Asia, while in SSA stabilised adobe blocks, soil-cement mixes and plants such as typha offer strong potential.

German development cooperation should bring the economic co-benefits and development potential for local firms working with bio-based more strongly to the international table. In its own development cooperation projects, greening the construction sector including bio-based material options for green job creation and firm upgrading should become an explicit goal. The exact type of bio-based material and its value chain to be supported needs to be analysed ex ante in the country regarding: Soil (availability and quality), import-export structure and tariffs of construction materials, market conditions and bottlenecks on supply and demand side. At the same time, scaling bio-based materials requires local political willingness to confront entrenched interests in the construction sector, alongside clear government regulations that set direction and create a level playing field for sustainable alternatives. This may be somewhat slower than other solutions, but just as feasible and it likely increases economic co-benefits for partner countries.