Development assistance has just suffered the steepest single-year contraction on record, while artificial intelligence (AI) has spread rapidly and unevenly through the organizations that deliver frontline services. Research suggests that the sector’s ability to absorb these financial and technological changes depends less on the technology itself than on organizational capacity and the funding models that shape it.
Development funding has changed abruptly in recent years. Net official development assistance fell by 23% in real terms in 2025, the largest annual decline ever recorded. This followed a drop the previous year, and the OECD projects a further fall in 2026, marking three consecutive years of contraction for only the second time in history.
Aid for core development programs and technical co-operation has fallen by 26%, the steepest drop on record for that component, while financing for health is expected to fall as much as 63% below its 2022 peak. The OECD’s analysis of the 2025 preliminary data and its near-term projections highlight service reductions and organizational closures across affected regions.
At the same time, AI adoption in the social sector has become near-universal. A 2026 benchmark study of 346 organizations found that 92% report using AI tools in some form, although only 7% report major gains in organizational capability — a pattern known as ‘an efficiency plateau’. The temptation is to read this as a technology story. The evidence suggests it is a capacity and financing story.
A capacity deficit that pre-dates AI
The constraint on absorbing shocks is structural. Research on the nonprofit starvation cycle describes a self-reinforcing dynamic: funders hold unrealistic expectations about what running an organization costs, organizations conform by underspending and underreporting administrative costs, and those reports in turn harden funder expectations.
What began as a largely anecdotal claim has since accumulated empirical support. An analysis of 25 years of financial data confirmed a sustained downward trend in reported overhead, driven by deep cuts to administrative expenditure. A study of German nonprofit finances from 2006 to 2015 found falling overhead ratios there too, though through different mechanisms.
The link between overhead and results has now been tested directly. Examining whether administrative spending affects program outcomes, revealed an inverted U-shaped relationship: below an optimal point, cutting overhead degrades program outcomes rather than protecting them. Notably, the optimal levels they identified sit substantially above both the 25% ceiling commonly recommended by charity rating agencies and the 27% average in their sample. The implication is uncomfortable: the ratio most widely used to signal efficiency is, past a point, a measure of self-harm.
Adoption is broad, shallow, and unequal
Survey evidence consistently finds enthusiasm outpacing capability. Drawing on more than 1,300 nonprofit professionals, the State of AI in Nonprofits study found that while 86% were exploring AI tools, only 24% had a formal strategy — and organizations with budgets above $1 million were adopting these tools at roughly twice the rate of smaller ones.
The 2026 benchmark study points to the same pattern. Close to half of the organizations have no formal AI policy and the measurement of AI usage outcome is rare, leaving most unable to tell whether AI is expanding capacity or merely accelerating existing tasks. A systematic review of 65 studies on AI adoption in NGOs similarly finds evidence scattered across engagement, decision-making, prediction, management and optimization use cases, with challenges clustering at the organizational rather than technical level.
This inequality has a geography. Reporting on the adoption gap in the Global South notes that generative AI use reaches roughly a quarter of the working-age population in the Global North against 14% in the Global South – and that the gap is widening.
Where the productivity gains land
The most rigorous field evidence on generative AI at work comes from a study of 5,179 customer support agents, published in the Quarterly Journal of Economics. Access to an AI assistant raised the amount of issues resolved per hour by about 14% on average but the distribution matters far more than the mean. Novice and lower-skilled workers gained roughly 34%, while effects on experienced high performers were minimal.
This finding is unusually well matched to the sector’s workforce needs. Organizations delivering services in low- and middle-income settings routinely cannot match private-sector salaries for senior specialists, so end up operating with junior teams with high turnover. That is precisely the profile for which measured gains are largest, suggesting the sector’s AI opportunity lies less in flagship program innovation and more in the unglamorous operational core: case triage, documentation, regulatory filings, reporting, data pipelines.
Illustrations of this pattern exist. In Brazil, the nonprofit NoHarm.ai developed Portuguese-language clinical models that rank prescriptions by risk so that scarce pharmacist hours reach the highest-risk patients (and subsequently applied the same class of tools to its own support and compliance workload).
The funding model is the binding constraint
If capacity is the bottleneck, the question becomes whether funding practice permits building it. Here the evidence has shifted considerably. A three-year study by the Center for Effective Philanthropy examined the effects of large unrestricted gifts. It found that recipients strengthened long-term financial stability rather than encountering the financial cliff many funders had predicted, with leaders reporting reduced burnout and greater capacity for innovation. Recipients held roughly twice the cash reserves of comparable organizations, and the feared crowding-out of other donors largely failed to materialize.
Taken together, the four literatures converge. Aid financing is contracting sharply. Organizational infrastructure has been systematically underfunded for decades, with measurable costs to program outcomes. AI adoption is widespread but shallow, and is most unequal precisely where fiscal pressure is greatest. And flexible funding demonstrably builds the capability that restricted project funding cannot.
Implications for funders and policymakers
- Retire the overhead ratio as a proxy for efficiency. Continuing to reward low overhead in a contracting funding environment penalizes the investments — systems, skills, data infrastructure — that determine whether organizations survive the shock.
- Fund AI capability, not AI tools. Near-universal tool use alongside minimal capability gains suggests that the missing input is strategy, workflow redesign and measurement, not software licenses. Governance should be a funding requirement rather than a reason for caution.
- Treat the adoption gap as infrastructure policy. Public computer programs, local-language data investment, and support for organizations building on domestic administrative systems address a divide that market diffusion is widening rather than closing.
- Direct flexible capital toward smaller and Southern organizations. Since adoption already correlates with size, undifferentiated funding will reproduce the divide. The unrestricted-funding evidence base was built largely on large recipients in the Global North. Extending and testing it in low- and middle-income settings is an open and urgent research question.
The development sector is being asked to do more with substantially less at the moment a general-purpose technology has arrived whose measured benefits fall most heavily on inexperienced workers in resource-constrained settings. Whether that coincidence becomes a partial cushion or a widened divide is not a question about models. It is a question about how development organizations are financed — and it is being answered now, by default, in the absence of deliberate policy.
