India AI Impact Summit Insights
All sessions

Inclusive AI at Scale: Turning Social Impact into Sustainable Markets

Contents

Executive Summary

This AI summit panel discussion explores how to scale AI technology equitably across the Global South, with emphasis on India's unique digital public infrastructure (DPI) model as a template. The speakers argue that inclusive AI requires governance frameworks that center developing nations, locally-adaptable yet globally interoperable standards, development-centric policy (not just risk mitigation), and strategic investment in foundational models, infrastructure, and human skills at population scale.

Key Takeaways

  1. India is positioned to set the global template for population-scale, equitable AI: Its DPI model (open, interoperable, secure) and willingness to share it ("digital diplomacy") offer a replicable blueprint for the Global South that diverges from Western models centered on large corporations and risk mitigation.

  2. Inclusive AI requires three simultaneous shifts: governance (include Global South voices), policy (development-centric, not just risk), and architecture (interoperable globally, adaptable locally). No single lever is sufficient.

  3. Frugal AI is not "inferior AI"—it's AI designed for purpose and ROI: Optimization at the hardware, model, and prompting layers enables small businesses, startups, and governments to deploy AI at 1-2 cents per transaction, unlocking markets invisible to Western enterprise AI.

  4. Voice and language are non-negotiable for the next billion: English-centric AI excludes 500+ million Indians and billions globally. Multilingual, voice-first interfaces are the bridge; they must be built into foundational models, not layered on top.

  5. The human element is non-negotiable: Reskilling at nation scale, human-centered design, and augmentation (not replacement) determine whether AI serves as a tool for inclusive growth or exacerbates inequality. Technology alone—without governance, skills, and empowerment—fails.

Key Topics Covered

  • AI Governance & Diplomacy: Fragmentation in current AI governance; need for Global South representation in decision-making bodies
  • Digital Public Infrastructure (DPI): India's DPI model (Aadhaar, UPI, Digilocker) as replicable template for other developing nations
  • Scaling from Pilots to Production: Moving beyond proof-of-concepts to enterprise and government-scale AI deployments across financial services, healthcare, agriculture, and labor sectors
  • Infrastructure as Utility: Cloud compute, foundational models, and hardware optimization for frugal/affordable AI
  • Data Governance & Sovereignty: Balancing local data residency requirements with interoperable frameworks; hybrid cloud-edge architectures
  • Secure AI by Design: Integrating security throughout the AI development lifecycle (DevSecOps, MLSecOps) rather than retrofitting
  • Language & Voice AI: Multilingual AI access, particularly voice interfaces, as the friction-free pathway to reach 500+ million non-English speakers
  • Skills & Talent Development: Reskilling millions of workers at nation scale; microcertifications and learning-by-doing approaches
  • Frugal Innovation: Building smaller, optimized models and architectures for low-cost deployment; quantization, mixture-of-experts, task-specific optimization
  • Human-AI Augmentation: Empowering workers rather than replacing them; addressing displacement fears through upskilling and work-life balance improvements
  • Ethical AI & Humane Design: Trust-building through transparent, usable, accessible technology; designing for intended and unintended consequences

Key Points & Insights

  1. Global South Must Be Co-Creator, Not Recipient: Current AI governance bodies (forums, standards-setting bodies) underrepresent developing nations. The panel emphasizes that Global South countries should be architects of data governance, algorithm accountability, safety standards, and cross-border data flow frameworks—not passive participants.

  2. Development-Centric Governance Over Risk-Only Focus: AI governance cannot focus exclusively on risks (bias, safety, alignment). The panel argues it must simultaneously address development needs: healthcare, education, financial inclusion. India's AI Summit in the Global South is positioned as an inflection point for this rebalancing.

  3. Globally Interoperable, Locally Adaptable Framework is Essential: Trust is built through frameworks that maintain global standards while remaining adaptable to local contexts, regulations, and problems. This principle should apply to data residency, compliance, and deployment models.

  4. India's DPI Model as Exportable Template: Aadhaar, UPI, Digilocker, and Digilocker demonstrate:

    • Open-source architecture
    • Interoperability
    • Security by design
    • Population-scale impact
    • These principles can be packaged (e.g., "DPI in a Box") for adoption by other Global South nations.

  5. Three Ingredients for Scale: Organizations successfully scaling AI implement:

    • Top-down ambition: Leadership commitment and bold bets on AI
    • ROI focus: Real metrics (topline, bottomline, customer experience, employee productivity)
    • Massive skilling: Scale upskilling by an order of magnitude; "add a zero" to your current skilling numbers

  6. Infrastructure as Utility Requires Three Pillars:

    • Rich, multimodal foundational models (accessible like "renting a super brain")
    • Optimized hardware (TPUs, efficient chips for inference) with transparent cost metrics (price per token to watts)
    • Open orchestration platforms (single pane of glass) with customer choice and flexibility

  7. Frugal AI Innovation Spans the Stack: Optimization occurs at multiple layers—chip design (int8 quantization, inference-optimized hardware), model architecture (mixture-of-experts, structured state-space models, domain-specific pruning), and prompting/programming (agents don't need poetry capabilities; they need tool-calling and planning).

  8. Voice & Language are the Inclusion Gateway: With 500 million non-English speakers and 2.5 billion without internet access, voice AI and multilingual models (e.g., Project Vani with 150,000+ hours across 100 languages) are critical. Linguistic diversity must be embedded in foundational models, not treated as an afterthought.

  9. Security Must Be Engineered, Not Retrofitted: AI applications require "DevSecOps" and "MLSecOps" mindsets. Security must be integrated at every phase: data integrity, model integrity, runtime protection (prompt injection defense), and continuous governance testing. As deployment scales to billions of users, breach costs far exceed prevention costs.

  10. Humans Are Amplified, Not Replaced: Evidence from education interventions (Azim Preji Foundation) shows that technology without teacher empowerment delivers zero impact. AI should augment workforce capabilities, reduce burnout, and create an "inspired workforce" that handles higher-value work. Reskilling (microcertifications, learning-by-doing, on-the-job upskilling) is essential to prevent displacement fears and worker anxiety.

Notable Quotes or Statements

  • On AI Governance: "Unless we have [Global South countries] on the board, I think you can talk of any sort of AI governance or it is not going to serve the purpose." (Unnamed panelist on representation)

  • On Development-Centric Policy: "AI governance must also be development-centric. You know it cannot be only the risk which needs to be taken into consideration when we talk of AI governance... We need to resolve those problems at population scale, for example, healthcare, education." (Governance panelist)

  • On Global-Local Balance: "We should have a globally interoperable framework and locally adaptable framework." (Emphasized as "very important")

  • On Scaling Ambition: "I have more pilots than Lufthansa in my organization... those days are over. Now we're talking about scale." (Punit, Microsoft)

  • On Skilling: "Whatever number they give me, I add a zero behind it. I say even that is not enough. Like everybody has to be scaled." (Punit, Microsoft, on upskilling scale)

  • On Frugal Models: "Rufus is a generative AI shopping assistant... int 8 quantized model. What it means is a very very small model and it does a job and it does a great work at that job." (Satendra, AWS, on optimized deployment)

  • On Trust via Design: "When you're going to democratize an AI based application, you expect billions of people to be banking with you... You break that trust if the data that they're sharing with you is somehow leaked out because remember for every good person using AI there's also a threat actor using AI." (Sapna, on Secure AI by Design)

  • On Human-AI Dynamics: "How do we take it to the mass market? Unless we bring up use cases where somebody on the street can let's say access credit. A farmer—he's able to get or what kind of irrigation do I need to take care of, how do I take care of my pest control—he gets that capability in languages..." (Akil Chadri, Highpring India)

  • On Bold Vision for 2030: "AI is going to be mainstream. What I mean is like electricity, internet, it's going to be part of our day-to-day lives... South Asia is going to play a leading role." (Akil Chadri, on mainstreaming AI)

  • On the Future of Work: "Every single person in this room will have a 24 by 7 AI helper living inside your phone who will do all your mundane tasks for you... so that all of us have more time to do quality stuff at work for society and with family and friends." (Aditya Swami, Google, on personal AI agents by 2030)

  • On People-Centric Design: "Put people at the center... the people that are using it in their day-to-day work have the best insights for how to use it. So empower them." (Paula Goldman, Salesforce)

  • On Affordable Fintech: "We would like all small businesses, very very small businesses, to be able to get a full suite of business services through which they can manage their businesses in say $1 or $2 a month." (Narendra Singh Yadav, PayTM)

Speakers & Organizations Mentioned

Direct Speakers/Panelists

  • Paula Goldman – Chief Ethical and Humane Use Officer, Salesforce
  • Narendra Singh Yadav – Chief Business Officer, PayTM
  • Kartik Rajaram – Country Manager India, Eleven Labs
  • Punit – Microsoft (specific role/surname not fully clear in transcript)
  • Satendra Pal Singh – CTO, AWS India
  • Shashi – Panelist on infrastructure (specific affiliation unclear)
  • Sapna – Panelist on Secure AI by Design (affiliation unclear)
  • Subadra – Panelist on architectural shifts (affiliation unclear)
  • Amit Zing – Director CTO, IBM Research (partial name, role inferred)
  • Aditya Swami – MD Partnerships, Google
  • Akil Chadri – CEO, Managing Partner, Highpring India
  • Professor M. P. Gupta – Director, IIT Lakhnau (alternate name given as "M.P. Gupta")
  • Sunil Ibrahim – Session Moderator, Co-Chairman (organization unclear)

Organizations/Institutions Referenced

  • Microsoft – Digital India foundation; AI pilots, co-pilots (healthcare, finance, legal)
  • Salesforce – Ethical AI; 5 million learners community in India; 75,000 trained on AI (2024)
  • PayTM – Digital payments at scale (50 million merchants); data-driven SME insights
  • AWS – Two India regions (Mumbai, Hyderabad); 240+ services; Rufus (AI shopping assistant); 6.2 million users skilled in cloud/AI
  • Google – Project Vani (150,000+ hours of speech data, 100 languages); Gemma (open foundational models, 22 versions); Google career certifications; Android ecosystem; DPI in a Box initiative
  • Eleven Labs – Voice AI; 500–$10k startup grants program
  • HCL Technologies – Data analytics, AI foundry, responsible AI framework; agriculture digitalization use cases
  • IBM Research – Frugal AI; model optimization; mixture-of-experts; quantization research
  • Apollo Hospitals – Clinician co-pilots; AI-integrated healthcare
  • Bajaj Finance – AI at scale in financial services
  • Ministry of Labor – AI-driven CV and mock interview platform for 300 million unorganized workers
  • UN Scientific Panel – Developing nations' representation in AI governance discussions
  • Azim Preji Foundation – Education interventions; computer lab study (zero impact without teacher empowerment)
  • IIT Madras, IIT Delhi, IIT Lakhnau – New BS and MBA programs in AI; talent pipeline initiatives
  • Government of India – $70 billion investment in computing infrastructure; DPDP Act (data protection); Ministry of Electronics and IT

Technical Concepts & Resources

AI Models & Architectures

  • Foundational Models – Large-scale base models (multimodal: text, voice, video, gestures, graphics)
  • Gemma – Google's open foundational models (22 versions, including smaller-weight options)
  • Mixture of Experts (MoE) – Model architecture optimization for efficiency
  • Structured State-Space Models – Alternative neural architectures for frugal AI
  • Int8 Quantization – Compression technique reducing model size/compute (e.g., Rufus example)
  • Rufus – Amazon's generative AI shopping assistant; 30 million interactions in 2 days; int8 quantized

Infrastructure & Hardware

  • Tensor Processing Units (TPUs) – Optimized chips for AI workloads; decade-long development; price per token to watts optimization
  • Inference-Optimized Hardware – Chips designed for deployed models (lower power than training hardware)
  • Cloud Compute Infrastructure – Utilities model; local deployment (India regions: Mumbai, Hyderabad)
  • Vortex Platform – Mentioned as orchestration platform for AI at scale (specifics limited)

Data & Governance Frameworks

  • Aadhaar – Foundational digital identity system; 1+ billion population coverage
  • UPI (Unified Payments Interface) – Real-time payments infrastructure; 300+ million transactions; benchmark for reliability at scale
  • Digilocker – Digital document storage; digitally signed QR codes with embedded facial images (innovation noted)
  • Project Vani – Google's multilingual speech data initiative; 150,000+ hours across 100 languages
  • DPDP Act – India's data protection regulation
  • DPI in a Box – Google's toolkit packaging India's DPI principles for other nations

Architectural/Development Approaches

  • DevSecOps – Integrating security throughout development lifecycle
  • MLSecOps – Security-first ML operations
  • Integrity Before Intelligence – Principle: ensure data/model integrity at each phase before intelligence outputs
  • Governance as Engineered, Not Retrofitted – Continuous testing and governance built into architecture
  • Hybrid Compliance Architecture – Balancing cloud-scale workloads with on-premises/sovereign requirements
  • Data Modernization Strategy – Cleaning, structuring, and preparing fragmented data for AI

Optimization Concepts

  • Frugal AI – Designing AI systems for cost-efficiency and resource scarcity (not feature bloat)
  • Task-Specific Model Pruning – Removing unnecessary neural network capabilities for specific domains
  • Prompting vs. Programming – Shift from natural language prompting to structured programming (agents, tool-calling, planning) for efficiency and optimization

Multilingual & Voice AI

  • Telugu Model (Chanda Mama Kadagalu) – Trained using CPUs by free software developers in Hyderabad; language-specific training example
  • Voice-First Interface – Friction-free access for non-literate and non-English-speaking populations
  • Linguistic Diversity Encoding – Building language support into foundational models, not as post-hoc layers

Sector-Specific Applications

  • Financial Services – Fraud detection (real-time, sub-millisecond UPI transaction screening); micro-credit scoring for informal workers
  • Healthcare – Claims processing; diagnostic co-pilots; telehealth accessibility
  • Agriculture – Pest management recommendations; irrigation planning; supply chain optimization; precision farming
  • Education – Teacher co-pilots; student learning personalization; accessibility across languages
  • Government – Unorganized worker skilling (CVs, interviews); policy implementation at scale

Learning & Skill Development

  • Google Career Certifications – Months-long, bite-sized reskilling programs (analytics, AI foundations)
  • Microcertifications – Short, on-the-job upskilling modules
  • Learning-by-Doing – Productivity gains + comfort with tools (Google Workspace, YouTube, Play Store with embedded AI) building confidence and native AI literacy
  • Tech MBA + BS AI Programs – 4-year undergraduate + co-terminal graduate