The latest World Development Report (WDR) calls artificial intelligence (AI) “a rare path to prosperity” for developing economies. AI is now part of most conversations about health systems and development finance. Yet little has been said about what it could mean for health systems financing: how health aid is pooled, allocated, and disbursed to low- and middle-income countries (LMICs). AI could be a “Fixer,” solving discrete, operational problems while leaving the underlying power structures intact, or it can be a “Shaper,” changing who decides what in and on what basis.
I argue that only a Shaper role can deliver the upsides of AI, but these upsides should not be taken for granted, nor will they come about automatically.
Three AI capabilities are at play in determining which scenario materialises. First, predictive tools which estimate what is not yet observed e.g., nowcasting health aid. Second, generative tools drafting and synthesising proposals, board papers, and reviews e.g., analysing programmes and projects from the IATI registry. And third, agentic tools acting: querying systems, triggering workflows, negotiating parameters between institutions. The first two mostly make existing processes faster—this is where the Fixer role lies. The third is where the Shaper role lies—that is, the question of who decides how money is allocated and spent.
The best of AI in health systems financing won’t happen on its own
That the best of AI won’t happen on its own is not a new message in development—the latest WDR and the Gates Foundation’s Goalkeepers report say the same. It applies to health systems financing, too. We don’t have a clear picture of the extent to which global health institutions are using AI, how AI-ready they are (Gavi doesn’t yet have an application programming interface [API]), whether they are using it in decision-making at all, or how they see the role of AI in their own mission. Coefficient Giving has published use cases of how they use AI to support grant-making. Wellcome asks applicants to disclose the use of generative AI in grant submissions while stating they do not use generative AI tools to assess grant applications or to inform funding decisions. These seem to be exceptions—other major global health financiers, such as Gavi and the Global Fund—do not have publicly available statements of their internal use of AI.
Nor do we know what LMIC governments are doing with the same tools in their dealings with global health partners—preparing applications, negotiating conditions, assembling reports. This knowledge gap matters: if global health as a system cannot see how AI is used on either side of the funding relationship, it cannot steer where AI adoption will go and who will reap the benefits.
In theory, AI can level the playing field of data, knowledge, and analysis firepower. In practice, many LMIC governments are unlikely to absorb and implement AI solutions at a scale and pace matching those of donors and global health institutions. If the likes of Gavi and the Global Fund, or the big philanthropies, adopt faster and more comprehensively than the governments they finance, the status quo is further entrenched: institutions will make more, better, and quicker allocation decisions with increasingly little cooperation from those governments. The asymmetry compounds the more that financing and performance data are read as signals—the stronger the incentive to produce them for the reader rather than the record. How adoption happens matters as much as how fast: buying capability rather than building it swaps dependency on technical consultants for dependency on AI solution providers, a class of actor answerable to no board, no replenishment, and no agreed principle in global health. Where would be the levelling in that?
If, on the other hand, LMIC governments built in-house capacity and used AI tools to their full potential when dealing with health donors, from applications to reporting, and coordinating across a wide range of partners, this would strengthen their sovereignty, improve alignment with national priorities, improve coordination, cut transaction costs, and help hold funders to account—a strong pivot towards the vision of the Lusaka Agenda. The same technology, but very different consequences stemming from who adopts it first and how deeply.
Three speculative scenarios for AI use
How beneficial AI turns out to be and to whom will depend on two things: how transparent its use is across the global health ecosystem, and how evenly and deeply it is adopted between LMIC governments and global health funders. Table 1 sets out deliberately provocative scenarios for what AI use could amount to in health systems financing.
Table 1. AI use scenarios in health systems financing
Scenario | Adoption and transparency | AI capabilities used | Likely consequence |
|---|
The research assistant scenario | Low adoption | Predictive and generative | AI as a Fixer. AI influences the power structures in health systems financing a little. Its potential is underused, but few substantial new harms are introduced. |
The integration scenario | High adoption, high transparency | Predictive, generative, and agentic | AI as an explicit Shaper. Global health actors, particularly LMICs, use AI agents to express and coordinate priorities, and hold each other to account. Transparency rises, decision-making improves, and political processes evolve towards shared objectives. |
The inward retreat scenario | High adoption, low transparency | Predictive, generative, and agentic | AI as an implicit Shaper. Use of AI agents is intense but guarded. Information-sharing falls for fear of being probed, trust and accountability weaken, and performative transparency masks real decision-making behind closed doors. |
The sharpest contrast is between the integration and inward retreat scenarios. Both assume AI becomes powerful enough to shape decisions; the difference is whether that power is visible, shared, and contestable, i.e., the Shaping is explicit, or hidden, uneven, and self-protective, i.e.,the Shaping is rather implicit. These futures are not equally available: integration requires many actors to agree, something global health has not always excelled in, while retreat requires only that each actor pursues its own interest.
Towards the integration scenario
No single authority can decide which scenario will materialise. The integration scenario is worth going for, but work needs to start now. Global health funders, LMIC governments, philanthropies and technical partners are already making choices that will push health systems financing in one direction or the other. The task ahead is to make AI adoption visible, contestable, and useful to countries as well as funders. Three agendas should form the bedrock of getting the best out of AI in this area.
1. Radical transparency of data and use of AI tools
Donors, global health institutions, and governments should disclose how and what they use AI for in analysis and decision-making, to foster credibility and alignment. One way to achieve this is to establish and maintain a register of AI use cases, including pilots. This costs little and, unlike most reforms in this space, requires no one else’s agreement.
The default in global health should be public data for everything. Much as the pre-print revolution reset expectations in science, global health funders’ boards should require grant agreements and individual transactions to be published by default and accessible programmatically, through APIs and other machine-readable interfaces—defaulting to oversharing and correcting in the open, rather than releasing a carefully curated selection of indicators at arbitrary time points.
Decisions in health systems financing should also rely increasingly, and transparently, on signals that are costly to fabricate or dress up with AI tools. This applies, for example, to co-financing and verification arrangements such as those run by Gavi and the Global Fund.
2. Capacity equalisation with support for LMICs
Global health funders should explicitly fund the uptake of AI tools by LMIC governments up to agentic capabilities, particularly for the purposes of health planning, coordination, and advancing priority-setting in health. These can build on existing partnerships and initiatives aiming to boost AI capability. Collaborations in health financing, such as those fostered by the Joint Learning Network and the former International Decision Support Initiative, should include AI uses and expand from domestic health financing to health systems financing use cases.
3. Shared vision for AI in global health and health systems financing
AI should be acknowledged as a potential Shaper of power structures in global health, particularly health systems financing, and included explicitly in ongoing discussions on reimagining the global health architecture so that a coherent vision of AI use can emerge, in line with agreed principles, such as those in the Lusaka Agenda. This shared vision should cover AI in health systems and AI in the financing arrangements, accountability practices, and institutional incentives that shape global health itself.
The most likely risk if these three shifts don’t happen is that of quiet divergence, where better-resourced institutions adopt AI capabilities faster, deeper, ask less of their partners, and disclose less of their own processes, while less resourced partners are left responding to increasingly sophisticated requests with weaker tools. Avoiding that future will require treating AI not as a back-office efficiency agenda, but as part of the political economy of global health itself.
With thanks to Pete Baker and Han Sheng Chia for insightful comments on an earlier draft.
