AI for ALL Challenge & Panel on Leveraging AI for Development in the Global South

Contents

Executive Summary

This conference featured multiple AI solution presentations addressing critical development challenges across agriculture, healthcare, education, and mental health—followed by a high-level Reserve Bank of India panel on AI governance in financial services. The overarching themes emphasized trust as foundational, human-in-the-loop decision-making as essential, and responsible innovation as the path forward for embedding AI in both regulated and underserved sectors across the Global South.

Key Takeaways

Trust is Built Through Transparency + Verification + Human Accountability: Whether in agriculture, healthcare, or finance, AI adoption hinges on verifiable outputs (digital report cards, explainable decisions, audit trails) and clear human responsibility for outcomes.
Agentic AI is Coming Faster Than Governance Frameworks: Banks and regulators must immediately design contracts, liability models, and consent mechanisms for autonomous agents—not wait for perfect frameworks—using principles-based, risk-tiered approaches.
India's Demographic & Data Assets Position it as a Global AI Innovation Laboratory: Diverse languages, massive unbanked populations, mobile-first users, and granular transaction data create real-world testing grounds for inclusive AI that the Global North cannot replicate.
Responsible Innovation ≠ Restraint: Policy consensus favors experimentation within guardrails (sandboxes, controls, compensatory checks) over blanket caution—reflecting recognition that incremental field learning outperforms theoretical prediction in AI governance.
The Next Frontier is Operationalizing Fairness: Model governance, bias detection, explainability, and independent oversight are no longer optional—they are competitive differentiators and regulatory expectations for institutions deploying AI in credit, healthcare, and public services.

India Impact AI Summit - Financial Services & Development Focus

Key Topics Covered

Sectoral AI Applications

Agricultural technology: Quality digitization, price linkage, market transparency
Healthcare diagnostics: Diabetic retinopathy screening, clinical decision support, health data management
Education & accessibility: Braille literacy, assistive tech for special needs, inclusive learning platforms
Mental health: AI companion chatbots, therapy delivery, multilingual support
Cyber security: Attack surface monitoring, asset management, threat prioritization

Governance & Policy

AI liability frameworks and shared accountability models
Trust mechanisms in agentic AI systems
Regulatory sandboxes and innovation frameworks
Data governance and privacy (HIPAA, GDPR, DPDP compliance)
Bias mitigation and model explainability
Global regulatory fragmentation vs. harmonization

Financial Services Applications

Credit underwriting using alternative data sources
Fraud detection and real-time risk management
Customer service automation (chatbots, virtual relationship managers)
Inclusion and financial accessibility

Key Points & Insights

Quality as Data, Data as Trust: Digital certification of agricultural produce eliminates subjective human negotiation, creating verifiable supply chains and enabling 8–20% income uplift for farmers while building consumer trust across the value chain.
Agentic AI Requires New Accountability Frameworks: Trust in autonomous agents demands more than trustworthy models—it requires contractual mechanisms, digitally signed consent, and clearer liability boundaries between regulators, institutions, developers, and users.
Underserved Populations Benefit from Accessible AI: Voice-based banking in Indian languages, AI-powered braille literacy devices, and autism screening games bypass traditional digital divides and reach rural/disabled populations excluded from conventional fintech.
Human-in-Loop Remains Non-Negotiable for High-Stakes Decisions: Banks report 40% correction rates when AI recommends prescriptions/treatments, and policy consensus holds that human oversight must remain for high-risk, high-ticket transactions and mental health scenarios.
India's Data Advantage is Structural, Not Accidental: UPI transaction granularity, mobile penetration, demographic diversity, and e-governance infrastructure create a unique testing ground for alternate credit assessment models—unavailable elsewhere and potentially globally applicable.
Model Governance Requires Institutional Separation: Independent model risk management teams (structurally separate from product teams) and continuous monitoring for model degradation are essential to prevent drift, bias creep, and undetected failures over time.
Bias Prevention Starts at Variable Selection: Technical debiasing techniques fail if contaminated input variables already embed historical unfairness; upstream ethical design is prerequisite for downstream explainability.
Regulatory Clarity Across Boundaries (Internal & Cross-Border) Accelerates Adoption: Banks cite regulatory sandboxes, data-sharing frameworks (account aggregator), and consistent multi-jurisdictional guidelines as critical enablers for responsible scale.
Deep Fakes & Spoofing in Indian Languages Remain Nascent Challenges: Defect detection tools exist for obvious attacks, but sophisticated deep fakes—especially in code-mixed Hindi-English, Tamil-English contexts—lack sufficient training data and detection capability.
Global Baseline Principles Should Be Technology-Neutral & Risk-Based: Rather than prescriptive AI-specific rules, frameworks should define outcomes (fairness, explainability, audit trails) and let institutions choose implementation paths—mirroring successful banking regulation models.

Notable Quotes or Statements

On Trust & Transparency

"When quality becomes data, data becomes trust, and trust becomes a market."
— Agricultural AI Presenter (digitizing produce grading)

"Trust is the foundation. There are two questions: How would you trust the AI model itself? And second, how do you know the agent is acting legally on behalf of a human?"
— Dr. Ravindran (IIT Madras, RBI Committee member)

On Human-in-the-Loop

"AI can assist you, AI can support you, but the decision would be that of the human being. That's very clear."
— Dr. Debda Chan (MD & CEO, Bank of Baroda)

"For high-risk cases like psychosis or suicidal ideation, we don't think AI should be answering those queries. Humans should be part of the entire loop."
— Hilo AI Mental Health Founder

On India's Structural Advantages

"We are at a point now where we can start running pretty good Indian language models thanks to recent releases that can actually run on the edge—even on feature phones."
— Dr. Ravindran

"With our UPI infrastructure there is so much very small level, microlevel transactions being recorded. We have data at granularity and scale that's usually not possible elsewhere."
— Dr. Ravindran

On Governance Philosophy

"Innovation over restraint. The idea is to nudge institutions to experiment, build controls and compensatory checks within product approval, then roll out responsibly."
— RBI Panel Moderator (Suvendu Pati, CGM Fintech RBI)

"Bank regulators should defend their territory as fit-for-purpose sector regulators as the governance space gets increasingly crowded."
— Terrell Leons (JP Morgan Chase Global AI Policy)

On Data Quality & Inclusion

"All AI ultimately stands on data. We need to make sure our data is diverse, accessible, clean, and governed. And then the people element—we need better STEM education to grow the next generation of AI creators."
— Dr. Debra Banerjee (Chief Data Scientist, L&T Finance)

Speakers & Organizations Mentioned

Government & Regulatory Bodies

Reserve Bank of India (RBI) – Fintech Department, constituted committee on AI governance (Aug 2024 report)
Ministry of Health – Supporting Madu Netra AI diabetic retinopathy screening
Ministry of Social Justice & Empowerment – MOU signed with Hilo Health for mental health AI
Government of India – Adopted RBI's "seven sutras" as national AI governance framework; NEP (National Education Policy) implementation support

Financial Institutions

Bank of Baroda – 50+ AI use cases; ~7 petabytes data lake; Aditi (virtual relationship manager); Bob Sambad (multilingual communication AI); 400+ facility partnerships
L&T Finance (NBFC) – Cyclops credit underwriting platform; ~15–20 ML scoring models; 40% rate of AI recommendation rejection by human underwriters
JP Morgan Chase – Global AI policy; operates in 100+ countries; shared accountability framework

Research & Academic Institutions

IIT Madras – Vadwani School of Data Sciences; Dr. Ravindran (RBI Committee member); AI tutors on feature phones
AIIMS Delhi – Madu Netra AI clinical validation
Stanford University – Infuse Health's clinical trial digital twins; Dr. Debra Banerjee (PhD in AI)
IT Bombay – Hilo Health's clinical psychology & psychiatry team
Xavier's Institute of Visually Challenged – Mumbai (potential partnership for assistive tech)

AI Solution Providers (Showcase)

Agritech AI – Fruit quality digitization (70K metric tons processed; exports to Brazil, Chile, Spain)
Madu Netra AI – Diabetic retinopathy screening (10K+ patients screened; 97% binary accuracy; 0.9 AUC for severity staging)
Tinkerbell Labs – Annie (braille literacy device); Helios 2.0 (content adaptation middleware for inclusive education; 300 schools, 50 centers)
Helium Health / IDUM Health – Care Copilot (speech transcription, differential diagnosis, clinical decision support; 1.7M patient EMR records; 7.9M prescriptions analyzed; 400 facilities)
CropScan Technologies – IoT device for pre/post-harvest pesticide/fertilizer testing; 8K+ farmers in Kenya; $1.8M credit unlocked; 4K farmers accessing export markets
Kedora Innovations – Autism/ADHD early identification via games + AI (83% sensitivity; 92% accuracy in pilots; 700+ children screened; 40K–150K deployment agreements)
Infuse Health – Digital patient twins for remote clinical trial enrollment (23 studies; 5 indications; $1.3M ARR on track)
Square Tech / Percept CEM – Cyber security (asset discovery, gap analysis, threat prioritization; 23 AI models; agentic AI for automation)
Torched Electronics – Joti Smart Glass (multilingual OCR, braille reading; 85+ PhD candidates; 4.5K+ professionals; 40K+ lives empowered)
Infy Health Tech / Hilo – AI mental health companion (1M+ users; 2M+ conversations; 93 languages; 165 countries; 30% MoM growth; clinically validated at AIIMS Delhi; HIPAA, GDPR, DPDP compliant)

Support Organizations

Google Launchpad, Nvidia Inception, NASSCOM AI for Good – Agricultural AI tech support
Wadwani AI – Madu Netra AI support
Nvidia – Hilo Health partnership; Jensen Huang mentioned Hilo at Nvidia summit
World Economic Forum – Hilo as Link Startup; Davos delegation
All Tech Accelerators – Google for Startups, Nvidia Inception (multiple startups mentioned)

International Regulators & Bodies

EU AI Act – Referenced fragmentation challenge
Financial Stability Board (FSB) – Implied role in global baseline-setting
BIS (Bank for International Settlements) – Implied role in harmonization
IMF, World Bank – Implied stakeholders in global governance
WHO – 1 billion PWDs need 2B assistive solutions; 80% from developing countries

Technical Concepts & Resources

AI/ML Architectures & Models

EfficientNet – Used for diabetic retinopathy classification
Transformer-based models – Referenced for explainability in financial services; face challenges with hallucination rates
Large Language Models (LLMs) – For speech-to-text, diagnostic suggestions, mental health conversation
Small Language Models (SLMs) – Edge-deployable Indian language models
Deep learning for game-based assessment – 50+ behavioral parameters extracted (swipe speed, acceleration, timing)
Account Aggregator framework – Data aggregation for credit underwriting
Supervised & unsupervised learning, contextual creation – Cyber security (23 models total)

Datasets & Data Sources

80,000 retinal images – Open-source for retinopathy model training
Retrospective + prospective data – AIIMS Delhi patient cohorts; annotation by retina specialists + adjudication
1.7 million EMR records – Helium Health across African facilities
7.9 million prescriptions – Analyzed for Care Copilot
UPI transaction data – India's micro-transaction fingerprint for credit assessment
Geospatial data – Living circumstances, rural area characterization
Device metadata – Smartphone usage patterns for behavioral signals
Satellite/GIS data – Unstructured rural information now processable via LLMs

Benchmarks, Standards & Compliance

97% sensitivity (diabetic retinopathy binary classification)
0.9 AUC (multi-class retinopathy severity)
83% sensitivity (autism screening via gameplay)
92% overall accuracy (autism screening pilots)
80–90% transcription + diagnosis accuracy (Care Copilot real-world testing)
40% AI recommendation override rate (L&T Finance Cyclops underwriting)
HIPAA, GDPR, DPDP compliance – Hilo mental health
HEPA compliance – Health data privacy
MChat questionnaire baseline – 70% threshold for screening tool validity
Gold standard comparison studies – Blind witness studies (Hilo vs. competitors)

Governance Frameworks & Methodologies

RBI's "Seven Sutras" – Principles adopted by Government of India as national AI governance baseline
Shared accountability framework – JP Morgan Chase's liability model (distinguishing roles: users, deployers, vendors, model developers)
Principles-based regulation – Technology-neutral, risk-tiered approach
Safe & Trusted AI Working Group – Trusted AI Commons initiative (benchmarks, toolkits for Global South scenarios)
Model Risk Management (MRM) – Independent teams monitoring for drift, bias, statistical error
In-the-loop human governance – Mandatory for high-risk, high-ticket decisions
Explainability standards – SHAP values, importance scores, natural language reasoning
Audit at multiple levels – Data audit, model audit, rollout audit, post-outcome audit
Regulatory sandboxes – Experimentation with anonymized, governed data
Consent & contract mechanisms – For agentic AI (new requirement)

Monitoring & Testing Tools

RBI Mule Hunter – Real-time fraud detection (implemented at Bank of Baroda)
Dashboard monitoring – Risk prioritization, remediation tracking (Percept CEM)
Continuous validation of transcription & diagnosis – Care Copilot (80–90% accuracy in low-resource African settings)
Behavioral gameplay parameters – Autism detection (swipe velocity, inter-tap intervals, etc.)

Data Privacy & Governance

Patient consent mechanisms – Mandatory for medical AI
De-identification & anonymization – For sandbox testing
Data lineage tracking – Source-to-inference transparency
Disability Aadhaar / UD ID integration (future)** – Assistive tech needs mapping

Languages Supported

Hilo Health: 93 languages (22 covered fully; 71 supported via translation)
Joti Smart Glass: Multilingual OCR
Bob Sambad (Bank of Baroda): Any language customer input/output
Indian language models (IIT Madras): Hindi, Tamil, Tamil-English code-mix, Hindi-English code-mix
Madu Netra AI: Globally deployed; language-agnostic retinal images

Patents & Intellectual Property

Agritech AI: 5 granted patents (AI-powered produce grading)
CropScan Technologies: Patent for pre/post-harvest pesticide testing
Torched Electronics: 5 patents filed (assistive tech for visually impaired)

Clinical Validation & Publications

Madu Netra: Published on PubMed-indexed journals; BBC, Economic Times, Indian Express coverage
Hilo Health: Clinically validated at AIIMS Delhi; Harvard MRS review mentioned therapy as #1 ChatGPT use case (2025)
Care Copilot: Prospective study design with retina specialist adjudication

Business Models Mentioned

Rental + Sales Model – Agricultural AI infrastructure