Building the Next Wave of AI: Responsible Frameworks & Standards

Contents

Executive Summary

This panel discussion at a global AI summit addresses the critical imperative of building responsible, ethical, and inclusive AI systems through collaborative frameworks and practical governance mechanisms. The speakers emphasize that safety benchmarks must emerge from real-world deployment contexts—not isolation—and that responsible AI should be embedded as a core product value proposition, particularly for MSMEs and startups operating at scale.

Key Takeaways

Responsible AI is a competitive advantage, not a compliance burden—when embedded as core product design from inception, governance and safety become value propositions that attract enterprise customers and enable scale, especially for startups and MSMEs.
India is uniquely positioned to shape global AI standards because its operating context (scale, multilingual, infrastructure constraints, dual growth/inclusion imperatives) reflects the conditions most of the developing world faces—making Indian frameworks more globally applicable than Western-centric designs.
Compliance and governance must be productized as APIs, not static documents, to drive adoption among resource-constrained organizations. The shift from "here's a 200-page PDF" to "invoke these compliance APIs into your system" is essential infrastructure for democratizing responsible AI.
Technology selection should follow problem definition, not capability availability. Using LLMs everywhere because you have compute capacity, or deploying generative AI when classical NLP suffices, undermines both safety and efficiency. Composite AI approaches are more trustworthy and performant.
Trust is a relationship outcome of sovereignty and transparency—organizations gain confidence in AI systems when they have complete data control, observability into decision-making, human-in-the-loop override capability, and clear documentation of how systems are designed, tested, and governed.

Key Topics Covered

Safety benchmarks and responsible AI frameworks: Designing benchmarks that emerge from deployment reality rather than research labs
The RAISE Index: India's first quantified assessment tool for AI safety and responsibility, harmonizing global frameworks
Government innovation hubs as critical infrastructure: Bridging policy intent and operational reality
Trust and governance as product design elements: Embedding compliance and observability into core product architecture
Sovereign AI and data control: Meeting enterprise and defense sector demands for complete data sovereignty
Composite AI and model selection: Choosing appropriate technologies (NLP, generative AI) based on problem definition rather than capability availability
Global standards harmonization: Aligning diverse regulatory frameworks (EU AI Act, NIST, Singapore, UK) into portable assessments
Human-in-the-loop systems: Designing AI to transition appropriately between autonomous and human-assisted modes
Small Language Models (SLMs) vs. Large Language Models (LLMs): Strategic technology selection over the deployment lifecycle
Productization of governance: Making compliance and guardrails integral to product rather than post-hoc additions

Key Points & Insights

Benchmarks must validate in real deployment contexts: Safety benchmarks fail when developed in research isolation. The most effective benchmarks come from institutions actively building, deploying, and maintaining AI at scale, with government innovation hubs positioned at the critical intersection of policy and operational reality.
The RAISE Index as practical standardization infrastructure: Co-developed by ICOM and The Dialogue over 18 months, this is India's first quantified index for measuring AI safety and responsibility impact during development and deployment. It harmonizes requirements across EU AI Act, NIST AI Risk Management Framework, Singapore guidelines, and UK AI assurance frameworks into a single, portable assessment.
Governance and compliance should be APIs, not PDFs: For mass adoption among MSMEs and startups, governance frameworks cannot remain static checklists or 200-page documents. Compliance should be productized—embedded as core APIs that can be invoked throughout the AI lifecycle (input, reasoning, tool-calling, output stages).
India's competitive advantage in global AI standardization: Most global frameworks are designed for high-resource, homogeneous environments. India's context—multilingual populations, infrastructure constraints, massive scale, simultaneous economic growth and social inclusion imperatives—is actually a strategic asset for developing validated, globally applicable frameworks.
Trust as organizational culture, not just technology layer: Salesforce's 10-year history with an "Office for Humane and Ethical Use of Technology" demonstrates that responsible AI requires institutional commitment. Trust must be prioritized alongside innovation, customer success, equality, and sustainability.
Composite AI over reductive generative AI selection: Not every problem requires LLMs. At IRCTC, 80-90% of interactions are successfully handled by classical NLP-based intent classification and entity extraction, with generative AI reserved for genuinely novel or complex queries. Technology selection should follow problem definition, not available compute capacity.
Probabilistic vs. deterministic expectations create compliance friction: AI systems are fundamentally probabilistic, but compliance requirements demand deterministic correctness. This tension can be managed by converting compliance requirements into versioned, open-source APIs that all ecosystem participants can reliably invoke.
Default vs. optional design shapes adoption at scale: Privacy and governance outcomes depend on what is default vs. optional. Customer data should not train models by default (it should be opt-in). Guardrails and observability should be defaults, not afterthoughts. Default settings drive 80-90% of actual behavior.
Sovereign AI and edge deployment as trust mechanism: Defense and government sectors increasingly demand data residency, on-premises solutions, and even edge AI (processing on local devices/appliances like the V-GPT desk box). Complete sovereignty enables higher trust relationships, even if it constrains some deployment flexibility.
Benchmarks must be living infrastructure, not static checklists: AI capabilities evolve faster than regulatory cycles. The RAISE Index incorporates phase-based assessment ensuring benchmarks remain relevant to company maturity stages. Frameworks must institutionalize continuous evolution through pilot phases, stakeholder consultation, and iterative refinement.

Notable Quotes or Statements

"Safety benchmarks fail when developed in isolation. The most effective ones come from institutions building, deploying and maintaining AI at scale." — Stefani Naguran (ICOM)

"The question comes down to how is India leveraging its innovation hubs and its leadership position in shaping the global dialogue on inclusive and responsible AI... India is uniquely positioned because most global frameworks are designed for high resource homogeneous environments. India operates in the context that most of the developing world shares—multilingual populations, infrastructure constraints, massive scale." — Stefani Naguran

"Unless and until we have sufficient transparent information exchange, unless and until we all say together that this is not something we will allow, it would be very difficult for us to stop the bad actors." — Arunati Batara (Salesforce)

"Technology this powerful should not and cannot be stopped because bad actors are misusing it... It's up to all of us to come up with a framework, a global compact, to ensure we are all trying to stop bad actors and using this for the good of humanity." — Arunati Batara

"If governance looks like a 200-page PDF for all companies, MSMEs will struggle. Governance should be the core product... Productize it, productize it, productize it, and that allows mass adoption." — Karna (Blue Machines AI)

"Human in the loop is a first class feature, not a failure point. Design the system so it knows when to transition from fully autonomous to assisted agent to human." — Karna

"Trust is important for any relationship, but with AI, trust is more important because they are trusting us with their data to create models. Enterprises want more of trust, scale, security than innovation." — Ankush (Sovereign AI provider)

"Safety should be the core of design... Don't do just GenAI because it's easily available. Begin with and end in the mind: first see what problem you're solving, then which solution, then which model." — Ankush

"Converting compliance into APIs... Every company can pick what compliances they need. These APIs should ideally get open-sourced in the market so there's enough validation across all players." — Karna

"Default vs. optional: whatever is the default selection has 80-90% adoption. Customer data should not be used by default to train LLMs—it should be optional. That's how you enable scale." — Karna

"Responsible AI practices from the beginning by design... The RAISE Index will encourage that approach." — Kazim Risby (The Dialogue)

Speakers & Organizations Mentioned

Speaker	Title/Role	Organization
Stefani Naguran	(Speaker, Keynote)	ICOM (Telangana AI innovation hub)
Kamesh Shaker	Associate Director, Moderator	The Dialogue
Arunati Batara	(Speaker, Global Enterprise Leader)	Salesforce
Karna	(Speaker, Founder/CEO)	Blue Machines AI; Upna (hiring platform)
Ankush	(Speaker, Technical Lead)	Sovereign AI/Defense AI provider (unnamed)
Kazim Risby	Founding Director	The Dialogue

Institutions & Organizations Referenced:

ICOM (India's first AI innovation hub, Telangana)
The Dialogue (AI policy think tank/convener)
Salesforce
Blue Machines AI (voice agents)
Upna (hiring/recruitment platform)
NVIDIA (edge computing collaborator)
IRCTC (Indian Railways Catering and Tourism Corporation) — customer case study
LIC, NPCI, Indian Army/Defense — customers/stakeholders
Global regulatory bodies: EU, NIST, Singapore, UK

Technical Concepts & Resources

Tools & Frameworks

RAISE Index: India's first quantified assessment index for AI safety and responsibility. Harmonizes EU AI Act, NIST AI Risk Management Framework, Singapore guidelines, UK AI assurance. Available via QR code; includes phase-based assessment and live iteration model.
Telangana Data Exchange: First digital public infrastructure for AI. Provides sandboxed government datasets to startups for testing AI systems against real use cases, actual constraints, and actual data before production deployment.
V-GPT Desk Appliance: NVIDIA-powered edge AI supercomputer with 1 petaflop floating-point capacity, 4TB storage. Runs models up to 1 trillion parameters. Enables sovereign, on-premises, room-level AI processing (designed for defense/sensitive government use).

AI/ML Methodologies

Composite AI: Strategic combination of classical NLP (intent classification, entity extraction) with generative AI, deployed where each is most appropriate. Reduces hallucination risk and unnecessary compute/token burn.
Human-in-the-Loop Design: Systems architectured to:
- Detect when they should transition from fully autonomous → assisted agent → human handoff
- Include guardrails at input, reasoning, tool-calling, and output stages
- Not treat human intervention as a failure mode but as a designed feature
Phase-Based Assessment: Benchmarking methodology that evolves with company maturity stages rather than static snapshots.
Compliance as APIs: Converting regulatory requirements into versioned, open-source APIs that can be invoked throughout the AI lifecycle, enabling consistent enforcement across organizations.

Model Sizes & Types

Large Language Models (LLMs): Used when novel/complex queries require generative reasoning. Recommended early in deployment journey for speed of innovation.
Small Language Models (SLMs): Transition target after understanding what is actually needed. Potential advantages: lower latency, lower cost, reduced hallucination risk. Strategic evolution path rather than initial choice.
Classical NLP: Intent classification, entity extraction, rule-based logic. Handles 80-90% of use cases in high-scale deployments (e.g., railway booking systems).

Governance & Compliance Concepts

Observability: Visibility into AI decision-making and system behavior. Critical for enterprise trust and debug/audit trails.
Explainability: Core API output requirement. Systems must articulate why decisions were made, not just what was decided. Enables human partners to be decision-makers alongside AI.
Data Sovereignty: Complete organizational control over data—storage location, access, audit permissions. Increasingly demanded by government, defense, and regulated sector clients.
Bias, Toxicity, Hallucination Detection: Trust layer components that validate data quality and model outputs before deployment.

Additional Notes

Closing Context: This panel was the concluding session of a multi-day global AI summit, positioning responsible AI as the capstone framework for the week's discussions.
Scope of RAISE Index: Currently in first edition; roadmap includes ongoing pilot phases, stakeholder consultation, and iterative refinement to keep pace with AI capability evolution.
Regional Positioning: Emphasis on India's role in global AI standardization reflects both practical advantages (operating in heterogeneous, resource-constrained, multilingual, high-scale environments) and diplomatic/geopolitical positioning in AI governance discussions.