Rethinking AI Governance for India and Indian Enterprises
AI governance has moved very quickly. A few years ago, the discussion was largely about ethical AI: fairness, transparency, accountability and explainability. Those principles still matter. But the debate has moved from principles to risk management, from risk management to assurance, and now from assurance to sovereignty.
In recent months, model access decisions, cloud infrastructure incidents and regulatory enforcement milestones have shown that AI dependence can become an operating risk, a legal risk and a sovereign risk at the same time.
For India, the question is therefore not whether we copy the EU, align with the US, or simply adopt ISO standards. The more practical question is whether we can build an AI governance architecture that is innovation-friendly, citizen-protective, enterprise-adoptable, globally interoperable and sovereign in its control over critical digital capabilities. Sovereignty in this context is not isolationism; it is the ability to remain globally interoperable while retaining choices, safeguards and recovery options.
Why governance has changed
AI governance has evolved in layers rather than through a clean replacement. The ethics layer gave us the language of human dignity, fairness and accountability. The risk-management layer, reflected in NIST AI RMF and ISO/IEC 23894, focuses on what can go wrong, who may be affected and how risk should be mapped, measured and managed. The assurance layer, reflected in ISO/IEC 42001, focused on whether organizations have repeatable controls, ownership, documentation, monitoring and review. The sovereignty layer considers who controls the compute, data, models, platforms, standards, audit capacity and recovery options behind the AI systems on which society and enterprise increasingly depend.
That last point is especially important for India. A country is not sovereign in AI merely because it trains a domestic model or gets access to GPUs. It is sovereign when it can decide which AI systems should be trusted, which should be constrained, which should be localized, which should be audited, and which dependencies should be avoided or diversified.
India must draw selectively from the global AI governance stack
The global AI governance landscape should be read as a stack. ISO/IEC 22989 gives vocabulary. UNESCO, OECD and G20 principles provide a human-centered direction. NIST AI RMF and ISO/IEC 23894 provide operational risk methods. The EU AI Act and the Council of Europe AI Convention show how rights, safety, democracy and rule of law enter the regulatory frame. ISO/IEC 42001 gives an enterprise management-system baseline. The G7 Hiroshima Process covers foundation-model risk. Platform governance shows the practical power of cloud, model and API providers. IndiaAI Mission, AIKosh and DPI-aligned AI offer India a path to build its own assurance capacity.
Comparative view of the AI governance stack
| Governance layer | Representative frameworks | Core contribution | Recent signal | Implication for India |
| Conceptual foundation | ISO/IEC 22989 | Creates a shared vocabulary for AI systems, actors, concepts and lifecycle terms. It reduces ambiguity before policy, procurement, audit or enterprise governance begins. | Fragmented definitions can turn governance into inconsistent paperwork across sectors. | Use common AI language for regulators, procurement bodies, enterprises, auditors and courts. Without shared definitions, governance will fragment. |
| Ethical and public legitimacy | UNESCO Recommendation on AI Ethics; OECD AI Principles; G20 AI Principles | Defines values AI should protect: human dignity, inclusion, fairness, transparency, accountability, privacy, safety, sustainability and human-centred innovation. | Public trust is becoming as important as model performance, especially where AI touches citizens directly. | Translate principles into measurable obligations for inclusion, language diversity and developmental outcomes rather than leaving them as abstract intent. |
| Operational risk management | NIST AI RMF; ISO/IEC 23894 | Moves governance from principles to practice: context of use, foreseeable harms, risk mapping, measurement, mitigation, monitoring and lifecycle controls. | Model and agentic AI risks are increasingly contextual, dynamic and operational rather than purely technical. | Create sectoral risk profiles for finance, healthcare, education, welfare, manufacturing and public administration. Voluntary adoption may not be enough in rights-sensitive domains. |
| Rights and legal guardrails | EU AI Act; Council of Europe AI Convention | Places limits around AI where it can affect rights, safety, democracy and rule of law. Supports prohibited practices, transparency and high-risk controls. | The EU AI Act moves from policy to enforcement in stages: Article 50 transparency and GPAI enforcement from 2 Aug 2026; high-risk rules follow later. | Absorb risk proportionality and stronger safeguards for sensitive domains while avoiding a compliance-heavy model that disadvantages startups, MSMEs or public innovation. |
| Enterprise assurance | ISO/IEC 42001 | Turns AI governance into an auditable management system covering policies, roles, responsibilities, risk assessment, lifecycle controls, monitoring and continual improvement. | Regulatory deadlines make assurance an enterprise design issue, not just a legal review at deployment. | Use as a practical baseline for Indian firms building or deploying AI at scale, especially those serving global markets. Certification should support accountability, not replace it. |
| Advanced AI and foundation-model governance | G7 Hiroshima AI Process and code of conduct | Recognizes that foundation models, generative AI and advanced AI need safety evaluation, misuse prevention, cyber-risk controls, data provenance, transparency and post-deployment monitoring. | The Anthropic Fable/Mythos episode (June 2026) showed how frontier-model access can change abruptly when national-security concerns intervene. | India’s foundation-model and agentic-AI ambitions need model evaluation, red-teaming, misuse monitoring and substitution options. Risks include frontier capability and Indian realities such as language failure and exclusion. |
| Platform governance | Cloud, model, API, app-store and enterprise SaaS provider policies | Embeds governance into infrastructure: acceptable-use policies, content filters, safety layers, access restrictions, logging, model availability and data-use terms. | Foreign platform rules or provider constraints can interrupt enterprise AI workflows with little notice. | Do not reject global platforms, but avoid hidden dependence on opaque foreign rules. Sovereignty should mean the ability to choose, negotiate, audit and substitute. |
| Sovereign public infrastructure | IndiaAI Mission; AIKosh; DPI-aligned AI; sectoral sandboxes | Connects AI governance to compute access, datasets, public digital rails, language resources, open evaluation, startup ecosystems and sovereign assurance capacity. | The July 2026 CloudFront VPC Origins outage underlined that AI systems also depend on resilient cloud and network paths. | Build AI on resilient, interoperable public and private rails. Global interoperability should be paired with domestic redundancy, audit capacity and recoverability. |
Recent incidents make the issue practical.
• Model access can become a geopolitical dependency. On 12 June 2026, Anthropic said a US government export-control directive required it to restrict Claude Fable 5 and Claude Mythos 5 from access by any foreign national. Because the order took effect immediately and real-time nationality verification was not feasible, Anthropic suspended access for all users until the controls were lifted on 30 June. The lesson for enterprises is clear: if a critical workflow depends on a foreign frontier model, model availability is not only a vendor issue; it can become a sovereign and geopolitical dependency.
• Cloud architecture is part of AI sovereignty. On 16 July 2026, AWS reported increased 5xx errors for CloudFront customers using VPC Origins connectivity; customers using other origin types were not affected. The practical lesson is that AI operating resilience requires multi-region, multi-origin, fallback and recovery design. Sovereign AI must include resilient architecture, not just sovereign models.
• Regulation is moving from debate to enforceable operations. The EU AI Act is now applying in stages. Article 50 transparency obligations apply from 2 Aug 2026, while GPAI obligations have already begun applying from Aug 2025 and high-risk rules to follow later according to the official timeline. This is a useful precaution for Indian enterprises: AI governance must be built into design, documentation, monitoring and incident response rather than treated as a later compliance add-on.
The enterprise agenda
Many organizations may still not have a full inventory of AI use: formal pilots, embedded AI in SaaS platforms, vendor models, analytics models, generative AI assistants, automation agents and informal employee usage. Without an inventory, risk classification is guesswork. Without ownership, monitoring is fragmented. Without fallback options, resilience is accidental.
Indian enterprises need to act on three priorities:-
First, create visibility before scaling AI. Build an inventory of AI use cases, embedded SaaS AI, vendor models, internal models, generative AI tools and automation agents. Classify them by business criticality, data sensitivity, regulatory exposure and impact on customers, employees or citizens.
Second, build AI assurance as an operating capability. Use NIST AI RMF, ISO/IEC 23894 and ISO/IEC 42001 to define ownership, testing, human oversight, red-teaming, monitoring, audit trails, incident response and evidence packs. AI governance should not sit only with legal or technology teams; it must connect business, risk, data, cyber, compliance and procurement.
Third, reduce hidden dependency risk. Critical AI workflows should not depend blindly on one model provider, cloud route, API policy or foreign platform rule. Enterprises need vendor due diligence, contractual safeguards, data-boundary controls, fallback models, multi-region resilience and clear recovery plans.
A Sovereign AI assurance stack for India
India needs a sovereign AI assurance stack rather than a single AI law or a purely voluntary code. This stack should include common AI vocabulary, ethical principles linked to inclusion and dignity, risk-based legal guardrails for sensitive domains, operational risk management that enterprises and public institutions can implement, an ISO/IEC 42001-style enterprise governance baseline, specific governance for foundation models and agentic AI, sectoral sandboxes, domestic capacity for audits, red-teaming, model evaluation, incident review and certification.
For policymakers, this balance is also important. Under-regulation leaves citizens and institutions exposed. Over-regulation favours large firms that can absorb compliance cost. India’s path should be graduated: clear prohibitions for unacceptable uses, stronger controls for high-risk domains, standards-based enterprise assurance, sandboxes for experimentation, public procurement rules for responsible AI and domestic audit capacity.
India’s choice
In the next phase of AI, nations will not be secured only by the models built internally. Nations will be secured by the trust architectures they create around those models. For India, the strategic choice is not EU-style firm guardrails, US-style market-led flexibility, or certification as a substitute for accountability. The better path is selective absorption: learn from global frameworks, adapt them to Indian realities and build a sovereign assurance architecture that is trusted, inclusive, resilient and enterprise-ready.
The essence of Sovereign AI is not just national compute, not just Indian datasets and not just indigenous models. Sovereign AI must also mean sovereign governance capacity – the ability to build, evaluate, assure and recover AI systems in a way that protects citizens, supports innovation and gives India strategic room to act on its own terms.
References: Anthropic statements on Fable/Mythos access suspension and redeployment (June 2026); Reuters report on lifting of export controls (30 June 2026); AWS Health Dashboard and The Register report on CloudFront VPC Origins 5xx errors (16 July 2026); European Commission AI Act Service Desk implementation timeline and FAQ.
