The case for AI in mission-driven contexts is no longer an argument. The relevant question for funders is which deployments produce outcomes and which produce harm — and what specifically distinguishes one from the other.
This page draws on a recent body of research documenting twenty government-supported AI deployments in India between 2023 and 2026, and synthesises the patterns into six representative cases. The notes on each case are short by design. The point is not the individual deployment; it is the pattern.
Three deployments that are working
Kisan-eMitra — AI grievance redress for PM Kisan
Ministry of Agriculture and Farmers' Welfare + Wadhwani AI (nonprofit AI implementer) + Google.org (philanthropic funding) + BHASHINI / IndicTrans2 + NIC (hosting). Implementation support from Samagra.
A multilingual chatbot integrated with the PM Kisan database, using retrieval-augmented generation across 49 query categories in 11 Indian languages. Farmers query through existing PM Kisan digital channels — no new app, no new data architecture.
Hundreds of thousands of queries resolved within months. The Ministry expanded scope to PM Fasal Bima Yojana, Kisan Credit Card, and Soil Health Card without re-tendering — a strong vote of confidence in the deployment model.
Multi-funder alignment with a clear government mandate and a nonprofit AI implementer (rather than a commercial vendor) created a deployment the government chose to expand. Wadhwani AI's nonprofit status aligned its incentives with public-good outcomes rather than commercial scale. Open-source language infrastructure compressed development time.
Qure.ai — AI chest X-ray screening for TB in Chhattisgarh and Tamil Nadu
Qure.ai (technology) + Piramal Swasthya (nonprofit field implementer) + State NHM health departments + Global Fund (TB funding umbrella) + WHO (validation and endorsement).
qXR, a deep-learning chest X-ray analysis tool deployed on portable devices, operated by community health workers without radiology training. Output is a probability score with an explainability overlay; presumptive positives proceed to sputum testing.
A 2026 study in Open Forum Infectious Diseases showed AI-CAD met WHO's minimal target product profile for TB sensitivity and specificity in an Indian tribal population — the first peer-reviewed validation of its kind. Tamil Nadu programmes reported higher sputum-positive yield than conventional screening.
The most underrated input was independent peer-reviewed validation before scale. Combined with WHO endorsement, this gave state health departments a defensible basis for adoption. Piramal Swasthya — a nonprofit field implementer — provided the operational infrastructure that the technology vendor alone could not.
Sarvam AI + EkStep — "Listen at Scale" voice AI for government welfare
EkStep Foundation (nonprofit DPG host, ₹1 crore grant pool to 20 grantee teams) + Sarvam AI (sovereign 22-language voice platform) + AI4Bharat (IIT Madras research validation) + National Health Authority + Department of Empowerment of Persons with Disabilities.
Multilingual voice AI agents conducting two-way conversations in local languages, capturing verbal consent in-flow, writing structured data directly into government registries. Queries are transactional — not aggregating surveillance profiles.
1.4 million senior citizens contacted via NHA, with a documented 42% increase in daily Ayushman Vay Vandana scheme enrolments. 414,000 persons with disabilities contacted via DEPwD, generating 51,000+ actionable profiles in a single wave.
The cleanest example of the structural pattern: nonprofit DPG infrastructure, sovereign AI technology, independent research validation, government data ownership. Consent was captured verbally, in-conversation, in local language — not buried in a text terms-of-service that low-literacy populations cannot read.
Three deployments that caused documented harm
Samagra Vedika — AI welfare exclusion
Government of Telangana (deployer) + Posidex Technologies (commercial entity-resolution vendor). No nonprofit, philanthropic, or social-sector partner in design or oversight.
An entity-resolution system cross-referencing government databases to flag welfare ineligibility. Originally designed for criminal identification by Telangana Police, then repurposed for welfare assessment without recalibration.
1.86 million food security cards cancelled between 2014 and 2019; 142,086 fresh applications rejected. Amnesty International documented systematic wrongful exclusions including a 67-year-old widow tagged as a car owner due to name confusion — a case that reached the Supreme Court. The vendor obstructed independent algorithmic audits.
The system has been replicated by Haryana, Tamil Nadu, J&K, and Karnataka without addressing any of its failures — a pattern of replication without evidence that funders should treat as a red flag in any state-level AI proposal. Procurement-driven AI without civil-society involvement, deployed against vulnerable populations, with an opaque commercial vendor and no grievance mechanism, causes harm at scale.
POSHAN Facial Recognition — biometric exclusion from nutrition entitlements
Ministry of Women and Child Development; state governments (Jharkhand and Bihar); UIDAI infrastructure (Aadhaar). No nonprofit involvement in design, pilots, or monitoring.
Smartphone-based facial recognition mandated for beneficiary authentication before ration distribution, matching faces against Aadhaar reference photos taken years earlier.
Documented by Decode and the Pulitzer Center in 2026: pregnant women denied rations because pregnancy changed their faces. Post-partum women excluded as their appearance changed. Younger women whose Aadhaar photos were taken at 18 rejected at 26. Anganwadi centres in Jharkhand reported no food supplies for two months, yet still required women to attend for biometric verification as a condition of staying on the rolls.
A fundamental design error: biometric authentication is unreliable for populations whose appearance changes — exactly the population this scheme exists to serve. Any AI system that becomes a mandatory gate to a basic entitlement must be funded with offline fallback as a non-negotiable design standard. The intervention solved a low-magnitude problem (fraud) while creating a high-magnitude one (exclusion of vulnerable women from nutrition entitlements).
PMFBY AI Crop Insurance — algorithmic claim denial at smallholder scale
Ministry of Agriculture; state governments; public and private insurers; ISRO and commercial remote-sensing providers; weather data partners. No nonprofit or beneficiary-side organisation in the algorithmic decision loop.
Satellite imagery and weather data processed through ML yield-estimation models, replacing manual crop cutting experiments for claim assessment. Claims evaluated at taluk or district level rather than individual farm.
Widespread farmer reports of claim rejection for visible crop losses that satellite models failed to detect on plots smaller than half a hectare — the standard size in Indian smallholder agriculture. Gujarat, Andhra Pradesh, and Bihar paused or withdrew from PMFBY citing implementation failures. Farmer trust collapsed; voluntary enrolment fell sharply in states with opt-out.
Models trained on uniform large-plot agriculture were misapplied to Indian smallholders without recalibration. Area-averaging at district level made individual farm losses invisible to the AI. Any AI proposal involving small-plot, mixed-crop, low-data-density agriculture should be interrogated against ground-truth conditions before scaled funding. Insurer transparency was absent — affected farmers had no way to understand or contest algorithmic decisions.
What the patterns tell funders
Successful deployments shared four conditions
- A nonprofit or mission-driven AI implementer central to technical design
- Multiple funders aligned around a clear mandate
- AI additive to a human service-delivery layer, not a replacement for it
- Independent validation — peer-reviewed, WHO-endorsed, or research-lab-led — before scale
Failed deployments shared four conditions
- Procurement-driven contracting with a commercial vendor, no nonprofit accountability layer
- Mandatory authentication of basic entitlements without offline fallback
- Replication across jurisdictions before evidence of effectiveness was established
- No civil-society or affected-community voice in design
The failure pattern that should most concern funders is the speed of replication without evidence. Telangana's Samagra Vedika has already been copied by four other states. Funders are one of the few institutional forces with leverage to slow this down.
What funders should fund — and what they should question
Fund nonprofit AI implementers as central technical partners, not commercial vendors as principals. Fund the wraparound — evaluation frameworks, consent architecture, grievance mechanisms — as much as the model itself. Require offline fallback as a non-negotiable design standard for any AI that gates access to entitlements. Fund peer-reviewed independent validation before scaling. Fund civil-society AI accountability — the investigative journalism and algorithmic audits that surfaced every one of the failure cases above.
Question proposals that scale before independent validation. Question proposals that replicate a model from another jurisdiction without auditing the original. Question proposals where the only accountability layer is the deploying government itself. Question proposals where biometric or algorithmic authentication is the gate to a basic entitlement.
The technical core of AI is becoming commodity. The architecture of partnerships, accountability, and design discipline is what distinguishes a deployment that reaches a population from one that excludes it.
Sources
Drawn from primary investigations including Amnesty International (2024–2026), Al Jazeera (January 2024), Pulitzer Center AI Accountability Network (March 2026), Decode and The Wire investigations, Open Forum Infectious Diseases (Qure.ai / Piramal validation study, 2026), WHO endorsement of AI-CAD as a TB triage tool, Wadhwani AI program reporting, and Sarvam AI's "Listen at Scale" documentation (March 2026).
Tilted Ground companion frameworks: AI Readiness for Funders, Five Questions for Funders, and Three Kinds of Scale.