Navigating Eligibility Criteria for Welfare: An Offline-First Approach to AI-Assisted Access to Social Protection in Rural India

Access to social protection in LMICs is often framed as whether benefits exist or whether citizens know about them; but a harder problem lies in between. Households eligible for multiple schemes must navigate fragmented programs, scheme-specific rules, and many documents to prove eligibility, and small errors can stay invisible until an application is rejected months later. Exclusion errors of this kind are well documented in India's welfare system (Muralidharan et al., 2020; Drèze et al., 2019). This means that citizens often bear much of the work of discovering, interpreting, proving, and maintaining their eligibility. Herd and Moynihan (2018) describe these learning and compliance costs as administrative burden. In practice, citizens rarely navigate these systems entirely on their own. Where scheme-level systems are fragmented and connectivity is unreliable, private agents, frontline workers and non-profits help bridge the gap. This is particularly important in rural India, where internet access remains limited (TRAI, 2025) and many people rely on shared devices. These touch points are essential, but they also make access slower, repetitive, and hard to scale. Workers may need to interpret similar requirements across different schemes, collect and check the same documents multiple times, and identify missing information manually. This raises the question of whether some of this work could instead be supported by a common system. We are exploring whether this workflow can instead be system-assisted. The idea is to create a common layer that identifies relevant schemes, interprets their rules, assembles evidence from household documents, and surfaces gaps before an application is submitted. Rather than replacing frontline workers, the system is intended to reduce repetitive work and help them identify errors earlier. As a first building block, we are developing an offline-first document intelligence module within a mobile app used by frontline workers. The worker photographs each document; the module performs OCR on-device, extracts fields, checks quality, and flags missing or inconsistent information for resolution on the spot. The workflow runs offline and syncs to the cloud once internet connectivity is available. We also account for the difficulty of processing Indic scripts, where OCR remains less reliable than for Latin scripts (Sarkar et al., 2024). The frontline worker remains the final verifier rather than treating the system's output as an automated decision. Our contribution is not any single component. On-device OCR and rule-based eligibility checking are well established. Instead, we are integrating these capabilities into an offline-first, human-in-the-loop pipeline for cross-scheme eligibility determination under real field conditions. We are piloting the app with a nonprofit partner across two blocks of Jhabua district, Madhya Pradesh, India, covering 14 welfare schemes across housing, food security, health insurance, pensions, and employment guarantee. Fieldwork is underway as of late September 2026 and runs for approximately four weeks. We evaluate the system across three dimensions. For document intelligence, we measure OCR accuracy; field-level extraction accuracy; failure rates by document type and script; and the frequency of manual corrections. For eligibility assessment, we measure false-positive and false-negative rates; frontline-worker overrides; and the number of eligibility checks completed. We also examine human and operational factors, including worker-reported usability, trust, and confidence, as well as successful completion of the workflow under offline conditions. Frontline-worker and supervisor back-checks provide an additional reference layer. With approximately 10 frontline workers and 100-120 households, the study is designed to generate feasibility and case-study evidence rather than precise per-scheme accuracy estimates. The deployment obtains informed consent from each household before any data is collected, with consent recorded and stored. Participants can request withdrawal, deletion, or correction of their data and can raise grievances with a designated point of contact. Formal ethics review and fuller alignment with India's Digital Personal Data Protection (DPDP) Act are next steps as the pilot scales. We present this as a case study of an ongoing deployment, with findings on accuracy, usability, and field performance expected by November 2026. The study will provide early evidence on whether AI can help shift social protection from citizen-led navigation towards system-assisted eligibility determination, even where schemes, documents, and digital infrastructure remain fragmented.