AI for the Last Mile: Human-Centred Design for Bharat

Key Topics Covered

• AI Implementation Challenges in Government & Public Systems

  • Data quality, availability, and integration across departments and states
  • Scaling from pilot to production in resource-constrained environments
  • Regulatory and institutional resistance to adoption

• Compute & Infrastructure Barriers

  • High upfront costs and ongoing compute requirements for AI systems
  • Edge computing and sovereign AI infrastructure as alternatives to cloud APIs
  • Cost-sharing models between government and startups

• Human-Centered Design & Localization

  • Language diversity (multilingual, regional language support)
  • User behavior and context differ significantly from Western markets
  • Blended digital-physical solutions required for last-mile users

• Business Model Innovation

  • Shifting from direct-to-consumer to B2B2C and government partnerships
  • Reframing mental health interventions as education to unlock government budgets
  • Identifying "painkillers" (acute problems) over "multivitamins" (nice-to-haves)

• Trust, Safety & Explanability

  • Data security and privacy in sensitive domains (courts, health, finance)
  • Regulatory compliance and professional liability
  • Bias and accuracy in non-English, domain-specific language processing

• Behavioral Change & Implementation

  • Training, change management, and cultural adoption among end-users
  • Building relationships with government stakeholders
  • Securing stakeholder buy-in (teachers, judges, parents, community workers)

• Government's Role & AI Governance

  • Data democratization and public data platforms
  • Subsidizing compute and infrastructure for impact startups
  • Building government AI capability for long-term ownership
  • Coordination between fragmented AI initiatives

• Skills & Education

  • Curriculum and training needs for AI-first workforce
  • Bridging technology gaps in MSMEs (micro, small, medium enterprises)
  • Police and cyber-security upskilling

---

Key Points & Insights

1. Data is the largest bottleneck, not compute. Across healthcare, agriculture, and justice sectors, accessing quality, clean, integrated data takes 6+ months. Government data platforms (e.g., Telangana's agriculture data exchange, now expanded to 1,100+ datasets across sectors) are essential enablers for startups and pilots.

2. Scaling pilot to production requires re-engineering for efficiency. Compute costs increase disproportionately at scale in AI systems—unlike previous software eras. Teams must plan for edge computing, small language models (SLMs), and proprietary optimization from day one, not as afterthoughts.

3. Business models must align with how government and end-users actually pay. Wize (mental health chatbot) succeeded in India only after reframing from "mental health intervention" (no budget) to "social-emotional learning via workbooks" (education budgets exist). Government subsidizing compute, not direct consumer payment, is realistic for impact sectors.

4. Behavioral change and implementation costs rival or exceed technology costs. Training teachers, judges, community workers; securing parent consent; building trust; and redesigning workflows are non-trivial, expensive, and often underestimated. Two-pronged approach: tech + programming (workshops, training, change management).

5. Localization goes far beyond translation. User behavior, device access, literacy, and trust differ fundamentally in Bharat. Wize's shift from chatbot (high data/phone access) to physical-digital hybrid (workbooks + QR codes) was necessary. Regional languages require domain-specific speech-to-text and language models, not generic NLP.

6. Sovereign AI infrastructure is critical for sensitive sectors. Adalat AI (justice sector) does not use third-party APIs (no OpenAI, Google); builds proprietary models; uses decentralized encryption; and manages legal Tamil and technical language—because court transcripts are confidential and highly sensitive. This adds compute and engineering complexity.

7. Trust is earned through solving acute pain points first, not big-picture multivitamins. Adalat AI's first product was speech-to-text transcription for judges (immediate, painful problem). Only after establishing credibility did courts open doors to paperless filings and chatbots. This "painkiller first" approach builds momentum and government pull.

8. Edge AI and low-cost compute devices are emerging solutions. A₹15,000 AI computer running models on Chrome (launched in Karnataka for teacher personalization) is expected to drop to ₹5,000–6,000. These enable deployment without cloud API dependency and make solutions cost-effective at scale (e.g., courtrooms, rural schools).

9. Government IT teams must evolve to AI teams. Current IT teams in ministries are built for transactional systems. Future solutions must be designed from the outset to be transferable to government ownership and operation—not permanently dependent on external nonprofits or startups. This changes engineering and architecture decisions.

10. Philanthropy is shifting from traditional impact to cost-effectiveness and systemic scale. Rising AI capability + shrinking public funds (e.g., US foreign aid pullback) + focus on state capability = funding models rewarding technology-driven, cost-effective, government-sustainable solutions. Randomized controlled trials (RCTs) and rigorous research embedded in product development are now expected.

---

Notable Quotes or Statements

• "We don't look at profits, but we look at impact or solving a certain problem." — Government representative (Telangana), on what drives government procurement and priorities.

• "You cannot build a business model around here [mental health in India], so this should not be a digital intervention. It has to be a physical intervention." — Ramakant (Wize co-founder), reflecting the necessity to redesign products for Indian ground realities.

• "We are the proud plumbers of the courts, trying to unclog the pipes of all these processes and paperwork." — Utkash (Adalat AI CEO), on identifying and solving systemic pain points.

• "The painkiller before the multivitamin." — Utkash, articulating the strategy of solving acute problems first to build credibility and open doors for systemic change.

• "If the funder is the payer, the pain points that are going to be solved are the funder's pain points. But really what you want is for the organization to solve the pain points of the government." — James Walsh (Agency Fund), on alignment of incentives for sustainable impact.

• "AI is not just tech, it is tech plus people." — Ramakant (Wize), on the dual requirements for implementation success in impact sectors.

• "Judges are developing spondilitis. They will say, 'We are not judges, we are scribes in the system.'" — Utkash, describing the physical and cognitive burden that motivated Adalat AI's speech-to-text solution.

• "The hairy problems that we were never as philanthropists comfortable taking on, I think we'll take on bigger challenges." — Prashant Prakash (Excel, ACT), on how AI and cost-effectiveness enable philanthropists to tackle harder systemic problems.

---

Speakers & Organizations Mentioned

Panel Members:

• Government representative — Electronics & Communications, Government of Telangana (data platforms, government AI policy)
• Prashant Prakash — Founding partner, Excel; Board member, ACT (philanthropy); awarded Praamshri 2025
• James Walsh — Senior behavioral scientist, Agency Fund; focused on social innovation and scale
• Ramakant — Co-founder, Wize (mental health AI chatbot; multilingual; billion conversations)
• Utkash Saka — Co-founder & CEO, Adalat AI (legal tech for courts, justice reform); LLM Harvard Law, MPA Kennedy School

Organizations & Initiatives:

• ACT Foundation — Knowledge partner; supports tech-driven innovation in education, health, environment
• Agency Fund — Funder; supports early-stage nonprofits in development and social impact
• Rocket Learning — Nonprofit; WhatsApp-based education for Anganwadi workers and parents; randomized trials embedded
• Wize — For-profit mental health chatbot; launched 2017; now 100+ countries, billion conversations, multilingual
• Adalat AI — Nonprofit legal tech; works with Indian courts (district to Supreme Court); speech-to-text, case management
• Telangana Government — Launched agriculture data exchange platform (2+ years ago); expanded to state-wide data exchange (1,100+ datasets)
• Government of Maharashtra — Piloting Wize in education (social-emotional learning, school-based)
• Government of Odisha — Health sector data lifecycle project
• Government of Kerala — First state to mandate Adalat AI speech-to-text in all courtrooms
• Welcome Trust — Funding Wize's clinical trial (₹7M, 5-year grant) in Uttar Pradesh
• NITI Aayog — Government AI policy, Frontier Tech Hub, AI for Jobs roadmap
• Ministry of Education (GoI) — Setting up AI skill development centers
• RBI, MHA, Banks — Involved in cyber-security and financial crime prevention
• Saram — Mentioned for voice AI capabilities

Other Referenced Entities:

• IITs — Building duplicate chatbots; coordination gap on compute
• E6 Data — Compute optimization for large companies; potential model for government AI stack
• Microsoft — Edge AI computer initiative (Karnataka, ₹15,000 device running on Chrome)
• Flip Cart, Swiggy — Referenced as recipients of compute optimization

---

Technical Concepts & Resources

AI/ML Approaches:

• Speech-to-text (transcription) — Domain-specific models for legal Tamil, technical language, regional languages; challenges in accuracy without third-party APIs
• Small Language Models (SLMs) — Mentioned as more efficient alternative to large LLMs; necessary for cost-effective edge deployment
• Multilingual NLP — Wize shifted from English-only (2017) to Hindi, then multilingual; business scaled significantly post-localization
• Randomized Controlled Trials (RCTs) — Embedded in product iteration (Rocket Learning); standard for demonstrating cost-effectiveness to government funders
• A/B Testing — Rocket Learning iterates on solution variants mid-trial based on RCT feedback
• Decentralized/Blockchain Encryption — Adalat AI uses judge-specific encryption keys requiring all keys to decrypt court database; security approach for sensitive data
• Edge Computing — Low-cost AI devices (₹15,000 → ₹5,000–6,000) running models locally without cloud API dependency; deployed for teacher personalization in Karnataka

Data & Infrastructure:

• Government Data Exchange Platforms — Telangana agriculture data exchange; expanded state-wide Telangana data exchange (1,100+ datasets); key enabler for startup access
• Data Quality — Core bottleneck: 6+ months to acquire clean, integrated government data; 30% of time spent on re-processing; cross-department integration slow
• Sovereign AI Infrastructure — Building proprietary models instead of relying on OpenAI/Google APIs; necessary for confidential sectors (courts, health, finance)
• Compute Cost Management — APIs to proprietary models changes unit economics; government subsidy of compute for impact startups; differential pricing for government stack
• AI Talent Mission — Recommended by NITI Aayog; government standing committee on AI; coordination of fragmented skilling initiatives

Domain-Specific Challenges:

• Legal Language (Law) — Pedantic, technical syntax ("as per doctrine of rest ipssa loquery per order 39 rule 1 and 2"); "Law is its own language"; requires custom speech-to-text and legal NLP models
• Clinical Safety & Professional Liability — Wize forced to prioritize clinical safety after NHS inquiry (2019); liability risks in financial, health, legal advice from AI
• Regulatory Compliance — Bar council licensing, professional malpractice in legal; RBI/MHA oversight in finance; slow regulatory frameworks lagging technology
• Product-Researcher Integration — Rocket Learning embeds rigorous research; allows real-time iteration and impact measurement

Business & Delivery Models:

• Blended Digital-Physical Solutions — Wize's workbook + QR code model; reduces phone dependency, leverages teacher trust, enables school budget alignment
• Painkiller vs. Multivitamin Framework — Adalat AI prioritizes solving acute, high-impact problems (stenography delays) before systemic improvements (paperless filing)
• Government as First Customer — Government provides infrastructure, customer base, distribution, ground-level nuance (credit relationships, cultural context); critical for market entry and scale
• Subscription vs. One-time Purchase — Schools buy books (scalable budget); apps purchased per-student less viable
• Cost-Effectiveness Framing — Solutions must demonstrate cost-per-outcome (e.g., cost per child impacted) to secure government/philanthropic funding in post-pandemic aid environment

Measurement & Validation:

• Impact Evaluation — RCTs, cost-effectiveness analysis; prerequisite for government adoption
• Continuous Monitoring & Evaluation — AI products require ongoing model evaluation, not just launch; differentiates good teams from poor ones
• User Engagement Tracking — WhatsApp open rates, adoption curves; product-led iteration informed by real usage
• Behavioral Metrics — Judge adoption of transcription; Anganwadi worker engagement; parent consent rates

---

Document Notes:

• Transcript has repetition and audio artifacts (typical of automated transcription). Interpretation is based on context and speaker intent.
• Some organizations (e.g., Saram, E6 Data) mentioned briefly; limited detail available in transcript.
• Government representatives did not provide on-camera identification but spoke as Telangana official and NITI Aayog representatives.