Renting Intelligence

Within roughly thirty months, generative artificial intelligence has moved from experimental novelty to operational dependency inside international development institutions. Multilateral development banks, United Nations agencies and large international non-governmental organizations are now committing measurable shares of their administrative budgets to foundation-model tooling, yet they do so under three intersecting uncertainties for which no replicable, sector-specific evidence exists: an upstream supply of model capability concentrated in a handful of foreign providers, internal governance configurations that vary widely with unknown performance consequences, and operational outcomes that have never been measured with the quasi-experimental rigour the sector routinely demands of its programme work.

This thesis addresses that gap with a design built entirely on public data, and it is organized around a single claim: that generative AI does not raise institutional performance by itself, but only where absorptive capacity and governance maturity convert a rented, externally controlled capability into usable organizational capability. The study assembles a census frame of approximately eighty development organizations spanning 2020 to 2026, codes their AI sourcing strategies and governance configurations from protocol-coded public disclosures, and constructs a quarterly institution-quarter outcome panel of forty-five institutions with machine-readable activity data.

The causal impact of generative-AI adoption on observable outcomes — project and procurement cycle time, document throughput and independent evaluation ratings — is approached through staggered difference-in-differences using the Callaway and Sant'Anna (2021) estimator, supplemented by synthetic-control methods (Abadie et al., 2010; Arkhangelsky et al., 2021) and guarded against selection by an adoption-propensity model with inverse-probability weighting. The framework integrates four load-bearing lenses: Resource-Dependence Theory at its core, the supplier-side oligopoly of concentrated upstream markets as the environment, the Technology-Organization-Environment framework as the governance scaffold, and Absorptive Capacity Theory as the conversion mechanism — set within strategic-management and behavioural foundations.

Consistent with a strict zero-fabrication standard, the thesis reports what the public evidence currently establishes and reserves the confirmatory estimates of the primary evaluation outcome until the treatment variable has been validated by independent coders and the long-lagged ratings have accrued. What is established now is substantial. The upstream market is concentrated at the capability frontier — an effective two to five suppliers by training compute — confirming the supplier-side oligopoly and the buyer-side dependence it imposes on development institutions as a measured feature of the market rather than a stylization.

Most consequentially for the central claim, adoption is shown to be non-random: the larger, more transparent and better-resourced institutions adopt sooner and at higher intensity — a positive selection the design measures directly and corrects for, and the empirical signature of the very capability gradient the mechanism predicts. The provisional proxy estimates remain uninformative, by the battery's own placebo and sensitivity results, which is reported as a constraint of post-adoption time and outcome lag rather than as evidence of no effect. That non-interpretability is itself informative: it demonstrates that the present public record is sufficient to detect adoption selection and run the estimator, but not yet sufficient to support a stable performance-effect claim.