Safe and Responsible AI at Scale: Practical Pathways

Contents

Executive Summary

This panel discussion addresses the critical challenge of making enterprise and government data "AI-ready"—structured, trustworthy, and interoperable so that large language models can deliver accurate, contextually relevant answers. The speakers emphasize that AI readiness is fundamentally a governance and data infrastructure problem, not primarily a technical one, and that successful implementation requires federated data ecosystems, clear frameworks, human oversight, and transparent accountability mechanisms. They argue that LLMs are tools to supplement human judgment (10–15% of a solution), not autonomous solutions, and that public data infrastructure, when properly structured, can democratize access to critical information for MSMEs, policymakers, and researchers across the Global South.

Key Takeaways

AI readiness is primarily a governance and data infrastructure challenge, not a technology problem. Success requires agreed frameworks, federated ownership, standardized metadata, and policy enforcement at the API level.
LLMs are tools (10–15% of a solution), not solutions. They require human oversight, guardrails, grounding in verified facts (via knowledge graphs), and clear communication about their probabilistic nature and limitations.
Open-source, federated data platforms (like Data Commons and MCP-enabled NSO data) democratize access to information for small businesses, policymakers, and researchers—reducing information asymmetry and unlocking economic opportunity at scale.
Metadata, glossaries, and context files are as important as data itself. Without machine-readable definitions, data cannot be trustworthy or interoperable; data silos persist even when data is digitized.
Data orchestration across disconnected sources is a practical necessity for real-world problems. Solutions must integrate administrative, alternate, and verifiable data sources while embedding accountability and auditability into the system architecture.

Key Topics Covered

Data silos and fragmentation — Enterprise and government data trapped in PDFs, disconnected systems, and inaccessible formats
AI readiness frameworks — What it means to prepare data for LLM consumption; absence of agreed definitions
Data governance and interoperability — Federated models, data stewardship, standardized metadata, and machine-readable catalogs
Knowledge graphs and contextual grounding — Using structured knowledge to anchor LLM outputs in verifiable facts
Domain-specific challenges — Dialect support, specialized vocabulary, translation gaps, and localization needs
Data quality and verification — Distinguishing declared vs. verifiable data; survey quality and metadata rigor
Public data commons — Open-source platforms (Data Commons) making statistical and administrative data discoverable and queryable
Policy enforcement and accountability — Embedding governance into APIs and policy engines rather than relying on manual compliance
Business models for data ecosystems — Incentive structures, pricing discovery mechanisms, and monetization frameworks
Use cases and practical applications — MSMEs location decisions, policy analysis, health data integration, price indices

Key Points & Insights

The Information Divide Exists: Critical government information (e.g., subsidies for women entrepreneurs in biotechnology) is stuck in notifications and PDFs, unreachable to both LLMs and the public. Making this data AI-ready could unlock enormous value—one organization manages 5 million compliance queries annually.
AI Readiness Requires a Shared Framework: There is currently no agreed-upon definition of what "AI readiness" means. Speakers called for foundational (core) and aspirational tiers of readiness standards, established collaboratively rather than imposed by single actors.
Data Ready for AI ≠ Data Ready for Humans: LLMs scan entire data domains at once (unlike humans focusing on specific questions). A folder with 10 budget versions will yield different answers depending on which version is queried—a fundamental difference requiring explicit attention to structure and context.
Governance > Technology: Whether to centralize or federate data, standardize vs. allow flexibility, or mandate vs. incentivize—these are governance decisions, not pure tech problems. The audience poll showed split opinion, but panelists emphasized governance as primary.
LLMs Are Probabilistic Tools with Inherent Limitations: Same prompts across different models (and same model across different runs) yield different answers. A mathematician from MIT explained mathematically why perfect consistency is impossible—it would remove the source of the model's generative capability. Guardrails, human-in-the-loop processes, and risk assessment are essential.
Metadata and Context Are Non-Negotiable: Without machine-readable metadata, context files, and glossaries/knowledge graphs defining terms, LLMs "drive blind." Data must declare meaning (e.g., "frequency = quarterly"), not assume it will be understood.
Data Orchestration Across Silos Is Critical: Real-world problems (e.g., a child's health) require integrating data from multiple disconnected sources (nutrition data from Women & Child Development; immunization from Health Ministry). Solutions must orchestrate these flows; data alone won't do it.
Verification and Auditability Are Practical Imperatives: Much public data is "declared" (self-reported) rather than verified (doctor-confirmed). For sensitive decision-making, provenance, auditability, and enforcement mechanisms must be embedded at the API/policy engine level, not left to human oversight.
Democratization via Access (with Guardrails) Beats Restriction: Panelists rejected the argument that average people shouldn't access health/financial data. The response to imperfect information should be education and better tools, not denial of access. Google Search faced similar objections; the alternative (suffer or lack information) is worse.
Federal, Open-Source Models Reduce Lock-In and Enable Choice: Centralizing data with one provider creates risk and limits LLM choice. Open-sourcing stacks (e.g., Data Commons) and exposing APIs (MCP servers) allow organizations to maintain sovereignty, choose which LLMs to use, and avoid vendor lock-in.

Notable Quotes or Statements

Rohit G (NSO/Mosby): "People are not aware what it takes to make data AI-ready... The first idea is to create a framework agreed framework — not my way or highway, but all of us work together create that framework."
Rohit G: "Unlike human... where I am focused on this question AI is designed to take scan the entire thing available. So it's a big difference between human and AI."
Prem Shankar Garg (Google Data Commons): "The majority of the world is flying blind... The majority of problems are 50–60 dimensional problems. Machines are really good at this... We have to approach AI as a tool we can use, not as the answer."
Prem: "Is the answer to that question [imperfect AI] suffer? Or is the answer to that question do less harm and give people a pathway that they can learn from?"
Rohit G: "Open data is not free data. Somebody has paid for it... Depending on the use we provide the data."
Ashish Choudhary (Microsoft/TripleIT A4i Lab): "LLM will give different answer, how you are compensating with guardrails, human in the loop, risk assessment... It is a probabilistic model. It cannot ever become as perfect... every time consistent... because then you are taking the main source of its creativity away."
Prem: "Can we reduce that risk for that individual [MSME]? Can we help them model, understand, derisk the decision they're making based on the audience they want?"
Rohit G: "The data should be federated. It should be located at every organization and governed locally by the organizations that are using it."

Speakers & Organizations Mentioned

Speaker/Organization	Role/Focus
Rohit G	Ministry of Statistics & Programme Implementation (MOSBY/NSO, India)
Prem Shankar Garg	Google, Data Commons project
Ashish Choudhary	Microsoft & TripleIT Bangalore, A4i Lab (AI innovation for Inclusion)
Sali (host/moderator)	AI summit moderator
Mosby/NSO	India's National Statistical Organization; manages GDP calculation, village/taluka-level data
United Nations Statistical Department	User of Data Commons backend
WHO, ILO	Data providers integrated into Data Commons
MSME sector (India)	74 million micro, small, medium enterprises; mentioned as key beneficiary group
Google Search	Historical reference for debates over public access to health information

Technical Concepts & Resources

Frameworks & Standards

AI Readiness Framework — Foundational vs. aspirational tiers for data readiness (proposed by NSO/Rohit G)
Data Boarding Pass — Concept for onboarding B2B users (policy makers, researchers, market players) onto knowledge graphs; checklist-based certification model
MCP (Model Context Protocol) servers — API standard allowing any LLM to plug into federated data sources; Mosby/NSO exposed 10 datasets via MCP server at mosby.gov.in
Knowledge Graph — Structured, machine-readable representation of facts and relationships; alternative to feeding unstructured data directly to LLMs
RAG (Retrieval-Augmented Generation) architecture — Combining domain-specific knowledge (retrieved facts) with LLM intelligence; debated as alternative to fine-tuning entire models

Data Preparation Components

Machine-readable metadata (JSON, XML) — Replacing PDF-based documentation
Business glossary — Standardized definitions of terms (e.g., "frequency = quarterly")
Context file — Points LLM to source definitions and meaning
Data catalog — Inventory of datasets, indicators, and definitions
Structured database definitions — Specifying dimensions, attributes, temporal roles

Platforms & Resources

Data Commons (datavommons.org) — Open-source platform aggregating 50,000+ datasets globally; overlays new datasets with existing ones to create network effects; 20-minute setup guide provided
Data Boarding Pass (databodingpass.org) — Tool for certifying and onboarding data as "AI-ready"
NSO/Mosby MCP integration — Accessible via mosby.gov.in; allows querying wholesale price indices, commodity indices, etc. via Claude, ChatGPT, etc.
A4i Lab — Collaboration between Microsoft and TripleIT Bangalore; working on education, accessibility (stem teaching for blind children), and last-mile health worker solutions

Models & Approaches

Glossary + LLM combination — Domain-specific vocabulary (e.g., for class 6 physics translation) combined with LLM to compensate for model gaps
Give Model for Data (G.I.V.E.) — Guaranteed trust, Incentive, Value, Exchangeability; framework for incentivizing data contribution and reuse
Benchmark for LLM consistency — Work underway to measure whether same question yields same answer across multiple LLM calls (stability measure)

Challenges & Cited Research

Paper (Canadian undergraduates) — Demonstrated same prompt + dataset yields different analysis types across LLMs
Paper by Rohit G (just accepted at major conference) — "AI Readiness of Data"
Book reference — "Raw data is an oxymoron" (cited by panelists to emphasize data always carries context)

Use Cases Demonstrated

Wholesale price index queries (Mosby/NSO) — "What's been the price of moong dal over quarters/months?"
Tamil song CPI index — User-created Consumer Price Index for grains mentioned in a Tamil song (shows creative reuse potential)
MSME location optimization — Using demographics, traffic, affordability data to help small business owners choose store locations
Women entrepreneur subsidy discovery — Finding applicable government schemes (e.g., biotechnology subsidies) via AI-ready notification databases
Child health integration — Combining anthropometric, nutrition, immunization, and birth data from fragmented ministries for holistic decision-making
Climate/poverty analysis — Multi-dimensional analysis combining education, health, income, and economic data

End of Summary