India AI Impact Summit Insights
All sessions

Launch of the AI Evidence Playbook: From Policy to Practice | India AI Impact Summit 2026

Contents

Executive Summary

JPAL (Abdul Latif Jameel Poverty Action Lab) launched the AI Evidence Playbook, a practical resource designed to help policymakers, practitioners, and funders make informed decisions about AI-enabled programs. The session featured panelists from JPAL, Google.org, and the Indian government discussing how to identify where AI can create impact, design effective AI programs for real-world contexts, and evaluate outcomes rigorously—with particular emphasis on India's digital public infrastructure model as a template for global learning.

Key Takeaways

  1. The AI Evidence Playbook provides a three-part framework: Where should AI help? How should AI-enabled programs be designed? How should outcomes be evaluated?—applicable across sectors and geographies.

  2. India's DPI model is replicable globally: Layering AI on top of foundational digital infrastructure (digital identity, payments, data lockers) demonstrates how to achieve scale, access, and equity simultaneously.

  3. Successful AI adoption requires problem diagnosis first, then technology selection—not the reverse. Community engagement, regulatory alignment, and real-world constraints must shape design from the start.

  4. Jobs will transform, not disappear: Entry-level coders should upskill into AI transformation engineering and robotics; countries that train their workforces for these roles will capture disproportionate value in the AI economy.

  5. Evidence-driven evaluation is critical to avoid waste and inequity: Rigorous randomized evaluations, equity analysis, cost accounting, and sustainability planning separate impactful pilots from failed deployments at scale.

Key Topics Covered

  • AI Evidence Playbook: A new practical resource synthesizing randomized evaluations and behavioral insights for AI program design
  • Digital Public Infrastructure (DPI) + AI: How AI amplifies the reach of existing DPI systems (Aadhaar, UPI, DigiLocker) through voice and local language interfaces
  • Criteria for successful AI applications: Problem-first thinking, user-centered design, real-world constraints (low connectivity, low-cost devices), regulatory alignment
  • Skills transition in the AI economy: Moving entry-level coders from traditional coding into AI transformation and deployment engineering roles
  • Scaling challenges and evaluation: Cost efficiency, equity considerations (benefits shouldn't accrue only to high performers), hardware durability, regulatory compliance
  • Google.org accelerator approach: Selection criteria including scale potential, open-source commitment, and long-term sustainability vision
  • India as a model for global AI development: Lessons on community-driven design, DPI architecture, and data sovereignty (e.g., Karda's data creation initiative)
  • Sectoral deployment: Healthcare, agriculture, education, and urban management (traffic optimization) as primary application domains
  • Labor market transformation: Concern about entry-level job displacement offset by opportunities in AI transformation engineering and robotics
  • Regulatory coordination: Role of central agencies (ICMR, drug control authority) in certifying AI solutions; state-level scaling responsibility

Key Points & Insights

  1. DPI + AI as exponential multiplier: AI layers on top of India's existing digital infrastructure (Aadhaar, UPI, DigiLocker) make services accessible via voice and local language—enabling access for farmers with basic phones to access agricultural advisories, healthcare workers to diagnose via AI medical assistants, etc.

  2. User-centered design is non-negotiable: Solutions fail when designers assume users have smartphones or reliable connectivity. Successful deployments account for 5,000 rupee phones, intermittent/offline functionality, low literacy, and language barriers.

  3. Problem-first, not solution-first: Technology should address real problems identified through community engagement, not retrofit existing solutions to invented problems.

  4. Equity within average impact matters: A technology can show positive average impact but still fail if benefits accrue disproportionately to high-performing users (e.g., best students) while harming lower-performing users.

  5. Regulatory approval ≠ deployment readiness: Solutions like AI tuberculosis diagnostics must pass regulatory certification (ICMR, drug authorities) but then require training, change management, and operational scaling at state/local levels.

  6. Cost and sustainability are evaluation metrics: Scaling an overly expensive solution creates bottlenecks; JPAL and Google.org both prioritize cost efficiency and demonstrated pathways to long-term sustainability.

  7. Open-source and data transparency matter: Google.org prioritizes applicants committed to open-source code, data, or model weights to prevent vendor lock-in and enable broader ecosystem learning.

  8. Job displacement will be selective: Entry-level traditional coding jobs will decline with AI agents, but demand will surge for AI transformation engineers (integrating agents with legacy systems) and physical AI/robotics specialists—India is positioned to become a hub for these higher-value roles.

  9. Hardware is not an afterthought: Evaluation must account for tablets dropped 11 times daily, shared devices in classrooms, and environmental durability—not idealized lab conditions.

  10. Community design for scale: Solutions should be co-designed with target communities (parents, teachers, farmers, health workers) to understand real problems, then built with local design principles in mind (offline-first, low bandwidth), but architected to identify and reach similar communities globally.

Notable Quotes or Statements

  • Abhishek Singh (Ministry of Electronics & IT, India AI Mission CEO): "DPI raised to the power of AI"—capturing how voice-enabled AI services expand reach exponentially beyond browser/app-based digital infrastructure.

  • Abhishek Singh: "People with AI skills will take away the jobs of people without AI skills"—reframing job displacement as a skills gap rather than net job loss, with entry-level coders becoming AI transformation engineers.

  • Abhishek Singh: "It's too early to call what succeeds and what fails because AI is just coming in... solutions which are deployed many of them are in pilot stage... we'll have to see after giving them a finite time."

  • Maggie Johnson (Google.org): On Karda's data creation model—"the actual pay is way more than minimum wage in rural India... and they're actually scaling it to Ethiopia and Kenya"—highlighting social enterprise models that benefit data creators directly.

  • Abhishek Singh on design principles: "You have to think it from the point of view of the person who is accessing the solution... Does the solution work on a low cost device in a low connectivity area?"

  • Sam Asher (JPAL) on equity: "if the benefit accrued disproportionately to the best performing students in the class and the lower performing students in the class actually went down then that particular technology is not what you probably had in mind."

  • Maggie Johnson on design for scale: "Design with the community to build for the community and then design for scale... you want to be able to design for scale and that really means understanding the attributes of this community and identify where else in the world it might make sense."

Speakers & Organizations Mentioned

Panelists:

  • Abhishek Singh – Additional Secretary, Ministry of Electronics & Information Technology; CEO, India AI Mission; Director General, National Informatics Center (NIC); Named in Time 100 Most Influential People in AI 2025
  • Maggie Johnson – Global Head & Vice President, Google.org (philanthropic arm of Google); formerly a professional musician
  • Sam Asher – Co-author, AI Evidence Playbook; JPAL researcher
  • Bal Daliwal – Global Executive Director, JPAL
  • Audrey Loro – AI Policy Associate, JPAL
  • Kaneka (moderator) – Former Indian Administrative Service officer; IAS topper 1996 batch

Key Organizations:

  • JPAL (Abdul Latif Jameel Poverty Action Lab) – Research institution behind the AI Evidence Playbook; partnered with Google.org on Project AI Evidence Initiative
  • Google.org – Anchor funder for JPAL's Project AI Evidence; runs GenAI accelerator programs
  • India AI Mission – National government initiative led by Abhishek Singh; architected along layered principles (compute, datasets, foundation models as common layers; applications at state/private sector level)
  • Ministry of Electronics & Information Technology (MeitY) – Government body overseeing India's AI and emerging tech agenda

Featured Example Organizations/Projects:

  • Karda – Data annotation social enterprise in rural India; labels datasets for AI models; creators receive royalties on reuse
  • Tarjimi – Translation service platform for refugees/asylum seekers; integrated GenAI via Google.org accelerator; expanded languages and reduced response time
  • Green Light – Google Research traffic optimization system using real-time pedestrian/traffic data; deployed in multiple Indian cities; reduces emissions 10%, full stops 30%
  • CURE AI (tuberculosis diagnosis) – Validated by Central Drug Control Authority and ICMR; example of regulated medical AI solution

Research & Certification Bodies:

  • ICMR (Indian Council of Medical Research) – Certifies medical AI solutions
  • Central Drug Control Authority – Regulates medical interventions including AI diagnostics

Technical Concepts & Resources

AI/ML Tools & Frameworks

  • GenAI/Large Language Models: Foundation for voice-enabled services, translation, diagnostic decision support
  • Coding Agents: GitHub Copilot, Claude, Cursor, Code Execution engines—tools that augment entry-level developers
  • Digital Twins: Virtual replicas of physical entities (cities, infrastructure) used for simulation and intervention planning before real-world deployment
  • Agentic AI: AI systems that autonomously integrate with legacy IT systems; referenced as a growth area
  • Physical AI & Robotics: Emerging domain expected to create 3–5 year growth in deployment and design roles

Infrastructure & Data

  • India Stack Components:
    • Aadhaar (digital identity)
    • UPI (unified payments interface)
    • DigiLocker (digital documents)
    • Associated GVD projects
  • Foundation Models: Government of India providing as common layer in India AI Mission architecture
  • Datasets: Local-language datasets in short supply; Karda model addresses via community-created, creator-compensated annotation
  • Open Source Commitments: Code, data, and model weights prioritized by funders for ecosystem health and vendor independence

Evaluation Methodologies

  • Randomized Evaluations: Core to JPAL's evidence approach; controls for bias and impact attribution
  • Equity Analysis: Disaggregating impacts by user subgroup (performance level, geography, income) to detect regressive effects
  • Cost-Effectiveness Analysis: Sustainability and scalability metrics alongside impact
  • Regulatory Certification Pathways: Solutions must meet regulatory standards (ICMR for medical, drug authority for interventions) before population-scale deployment

Design Principles (Recurring)

  • Offline-first / Asynchronous connectivity: Data syncs when connectivity available
  • Low-device hardware: Function on ₹5,000 phones; work with intermittent power
  • Multilingual/Voice interfaces: Overcome literacy and language barriers
  • Hardware durability: Account for tablets dropped repeatedly, shared devices, outdoor use

Sectoral Applications Referenced

  • Healthcare: AI diagnostic tools (ECG, tuberculosis), telemedicine companions, maternal/child health
  • Agriculture: Plant disease diagnosis, agricultural advisories, seed/fertilizer access
  • Education: AI tutors aligned to syllabi and learning levels; asynchronous learning tools
  • Urban Management: Traffic optimization, emissions reduction, city digital twins

Project/Funding Mechanisms

  • Project AI Evidence Initiative: JPAL + Google.org funded call for proposals; received 96 well-curated proposals (up from 65 in first round)
  • Google.org GenAI Accelerator: Cohort-based program for nonprofits already using technology; selection criteria: scale, open-source commitment, long-term vision

Policy/Governance Frameworks

  • India AI Mission layered architecture:
    • Common layer: Compute, datasets, foundation models (provided by government)
    • Application layer: Built by state governments and private sector
  • Regulatory coordination: Central agencies (ICMR, drug control) certify; states handle training, rollout, scaling
  • Seven-pillar whole-of-government strategy: Mentioned but not detailed in transcript

Context & Significance

This session represents a pivotal moment in global AI governance:

  1. Scaling evidence-based AI evaluation: JPAL's playbook formalizes lessons from randomized trials and digital transformation into actionable guidance, filling a gap between "let's experiment with AI" and "how do we know if it worked?"

  2. India as a tested model: With 4+ years of AI governance experience post-ChatGPT and deep digital infrastructure (Aadhaar, UPI), India demonstrates that layered, federated approaches—combining national standards with state autonomy—work at scale.

  3. Inclusive AI development: The summit's reach (11,000+ online viewers, students and high school attendees from across India) signals intent to democratize AI literacy and entrepreneurship beyond traditional tech hubs.

  4. Addressing the equity challenge: Repeated emphasis on disaggregated impact, serving low-resource users, and regressive effects suggests sophistication about AI's risk of amplifying existing inequalities.

  5. Labor transition planning: Candid acknowledgment of job displacement paired with proactive skilling strategies (AI transformation engineering, robotics) positions this discussion ahead of many countries still in denial about automation.