Balasubramanyam Pattath (Balu)

We extend Kremer’s O-ring theory by modeling artificial intelligence as a failure-reduction technology. AI raises task reliability by eliminating a fraction of residual failure risk, which makes AI and human quality substitutes within a task but complements across complementary tasks. Endogenous deployment generates three allocation regions in the exact two-task model, can attenuate or reverse sorting complementarity under sufficient asymmetry, compresses marginal skill premia without destroying O-ring wage convexity, and raises the value of non-automatable bottlenecks. We then add an extensive-margin adoption problem in which technical AI capability, economic stakes, baseline reliability, complementary non-AI reliability, and deployment costs enter separately. The model yields a sign-changing capability–reliability cross-partial and sharp predictions for capability interactions with stakes, complementary reliability, variable deployment costs, and fixed costs. The empirical design follows these theoretical objects sequentially. Radiology identifies the primitive failure-reduction mechanism using repeated measurements to correct the mechanical errors-in-variables problem created by latent baseline quality. Across detailed occupations, direct measures of technical AI capability strongly predict usage, while conditional wage gradients are weak. We use this occupation-level capability measure and pre-generative-AI May 2022 U.S. occupational staffing patterns to construct a Rajan–Zingales-style external industry AIopportunity index. Linking that index to the 2026 India World Bank Enterprise Survey AI follow-up yields a firm-level adoption design with industry fixed effects. The positive level association between AI opportunity and adoption is primarily between broad sectors. Within industries, external AI capability interacts negatively with predetermined worker schooling. We then add an independently measured local-infrastructure margin using BharatNet service-ready Gram Panchayats. Because the released WBES suppress district identifiers, district BharatNet data are aggregated to the survey’s released state/state-group regions. The capability–BharatNet interaction is negative, stable in sign across leave-one-region-out exercises, and concentrated toward the high-readiness end of the infrastructure distribution. When schooling and BharatNet interactions enter jointly with region fixed effects, the BharatNet interaction remains negative while the schooling interaction becomes imprecise. We treat both patterns as reduced-form evidence rather than structural parameters and conclude with a roadmap for distinguishing improvements in the AI frontier from reductions in complementary deployment constraints.

Artificial intelligence (AI) and digital public infrastructure (DPI) are increasingly discussed as complementary ingredients of digital development, but they are usually financed, deployed, and evaluated separately. This paper develops a simple model in which publicly supplied DPI changes the profitability of privately chosen AI deployment. Production has an O-ring structure: AI reduces failure on digitally enabled tasks, while residual analog tasks remain complementary bottlenecks. DPI has a direct friction-reducing effect and an enabling effect that raises the productivity of AI. Private agents face fixed and convex AI-deployment costs and capture only part of the social return. The resulting economy exhibits three wedges: underprovision of DPI, underdeployment of AI conditional on adoption, and delayed adoption. A Stackelberg government chooses DPI anticipating private AI responses and can supplement infrastructure with targeted adoption grants and intensive-use subsidies. The optimal policy mix changes with digital maturity: DPI is built first, intensive-use support becomes valuable once adoption begins, and targeted grants are most useful near the extensive-margin frontier. With continuous heterogeneity, private AI adoption can occur over an interior capability band rather than through a monotone cutoff. DPI expands this band in both directions, while residual analog bottlenecks delay diffusion and tilt first adoption toward high-capability users. AI-use inequality and adopter– non-adopter polarization can therefore be sharply transitional even when output inequality falls. The framework links the emerging policy discussion on AI–DPI convergence to phased AI evaluation and yields a practical question for adaptive development finance: for this application, country, and stage of digital maturity, where should the next public dollar go?

I study the role of relative position in informal risk sharing in Indian village economies. Using household panel data from three ICRISAT villages, I first document a robust pattern: a household’s rank in the within-group distribution, whether measured through consumption or income, is a powerful predictor of consumption dynamics, significant at the 1% level across all villages and castes. Controlling for rank substantially reduces or eliminates rejections of efficient risk sharing. The finding holds with both consumption rank (motivated by a preference-based derivation) and income rank, which is free of mechanical correlation with contemporaneous consumption changes and provides a direct bridge to the structural model. To quantify the structural importance of this finding, I embed position-dependent enforcement, where social sanctions upon deviation depend on a household’s income-state rank, into Bold and Broer’s (2021) limited commitment model with coalitional deviations. Estimating via simulated method of moments, I find the data consistently prefer position-dependent enforcement (ω = 0.95 across all three villages), with the SMM criterion improving by 46–56% relative to the baseline. The improvement is driven primarily by cross-sectional consumption dispersion: the model matches 51–82% of observed dispersion across villages, compared to 14–30% in the baseline, while preserving the core findings of small stable groups and approximately symmetric insurance.

India’s 1975–77 Emergency subjected districts to a mass sterilization campaign of sharply varying intensity, and a persistent deficit in public health-care utilization followed in the most affected districts, consistent with durable distrust of the public health system. Three decades later, the National Rural Health Mission (NRHM) deployed nearly one million community health workers (ASHAs) as household-level intermediaries between communities and the formal health system. While NRHM was designed as a supply-side health system reform, its scope could have created unintended positive externalities, precisely along the dimension of (re)building trust. Using a difference-in-differences design that interacts district-level sterilization intensity with the national reform, and individual-level data from six survey rounds spanning 1998–2021 (approximately 2.1 million observations; 402 matched districts, 379 in the balanced sample), this paper asks whether the reform narrowed the deficit, and how fast. In the short run (two to three years) there is no detectable narrowing, and the design had the power to detect effects well below those that later emerged, so the null is informative. Over a decade to fifteen years the deficit closes: by 2019–21, a one-standard-deviation increase in sterilization intensity is associated with differential gains of 3.0 percentage points in DPT vaccination and 3.2 points in antenatal care, outcomes with clean pre-reform trends, and comparable gains in institutional and public-facility delivery. The convergence matches or exceeds the measured pre-reform deficit, worse-off districts shift differentially toward public providers, and the path of recovery implies a half-life of roughly eight to ten years. Direct trust measures from 2012–13 and the timing of recovery are consistent with trust rebuilt through sustained community-embedded intermediation, plausibly via ASHAs; a differential rise in voluntary sterilization adoption provides corroborating but not dispositive evidence. Historically rooted distrust, these results suggest, is reversible, but recovery may take a decade or more.