Governments repeatedly inscribe memoranda of understanding with global technology titans, posing before logo walls for commemorative photographs. The fine print seldom articulates data residency protocols, model ownership provisions, or exit cost structures. Five years hence, when contracts renew at triple tariff rates, officials discover their departments cannot operate without foreign infrastructure. No conspiracy necessary—merely procurement shortcuts masquerading as progress. Sovereign artificial intelligence gets dismissed as expensive and slow, while vendor dependence becomes irreversible. The arithmetic is straightforward: renting feels economical today, extracts everything tomorrow. That logic is flawed, and the price of embracing it compounds daily.
India stands at a singular moment in the global artificial intelligence transition. We have demonstrated world-class capacity in digital public infrastructure—Aadhaar, UPI, the account aggregator framework—but these are fundamentally transactional systems. Artificial intelligence differs. It is interpretive, generative, and increasingly constitutive of state power. When an algorithm prioritizes which farmer receives drought relief, which urban settlement gains municipal services, or which tax return triggers investigation, it exercises sovereignty’s essence. The question emerges whether that sovereignty remains with the Indian state or is subcontracted to entities beyond its jurisdiction.
The sovereign artificial intelligence mandate therefore constitutes not merely a procurement preference but a constitutional necessity. It requires three foundations: models we own and can audit, data that persists within national boundaries under Indian legal frameworks, and computational infrastructure operating without foreign license or connectivity. This appears expensive. It need not be. Recent advances in model efficiency—demonstrated most dramatically by DeepSeek’s architectures—have collapsed the cost of capable artificial intelligence. What necessitated million-dollar GPU clusters two years ago now operates on standard servers. The barrier is no longer capital but coordination.
State-as-platform governance emerges as this coordination’s crucible. India’s administrative architecture was designed for paper files and vertical silos. Artificial intelligence agents traverse these structures, creating horizontal capabilities existing departments are ill-equipped to govern. The response cannot be another information technology department with expanded ambit. What is needed is a cross-cutting artificial intelligence council with genuine authority—the power to mandate standards, obstruct noncompliant procurements, and arbitrate disputes between departments. Without this, every ministry will commission its own conversational interface, its own data repository, its own vendor relationship, recreating fragmentation that has crippled previous digitization endeavors.
The technical specifications follow from governance clarity. Small, agentic architectures—task-specific programs chained into reliable workflows—offer decisive advantages over monolithic models for public administration. They are interpretable: when a Right to Information processing agent makes determination, its reasoning can be traced. They are isolable: a failure in the agriculture agent does not cascade into health system. They are efficient: operating on commodity hardware, they deploy at district level without perpetual cloud dependencies. Most significantly, they align with India’s constitutional values. A grievance redressal agent routing complaints to appropriate authorities, maintaining audit trails for Right to Information Act, and respecting Digital Personal Data Protection Act consent architecture is not merely efficient—it is legible to courts, legislators, and citizens.
The data transformation required appears less glamorous but equally urgent. Decades of administrative records remain trapped in proprietary formats, scanned images, or physical ledgers. The sovereign artificial intelligence stack demands machine-readable infrastructure—JSON-normalized records, open APIs, secure query layers with lineage tracking. This is not archival vanity. An artificial intelligence system is only as grounded as the data it accesses. Without verified, structured sources, government agents will hallucinate policy interpretations, invent precedents, and erode rule of law they are meant to serve. Investment in data cleansing and standardization constitutes the unglamorous prerequisite for everything that follows.
Language and access present both technical challenge and political opportunity. India’s linguistic diversity has historically been treated as friction to be managed—hence English’s dominance in digital governance interfaces. Sovereign artificial intelligence inverts this logic. Voice-first, multimodal systems in Marathi, Telugu, Bengali, and dialects become competitive advantages, enabling service delivery to populations excluded by literacy barriers. Technology for speech recognition in Indic languages has matured rapidly; what remains is policy commitment prioritizing these deployments over English-first demonstrations impressing visiting delegations but serving few citizens.
The Digital Personal Data Protection Act provides regulatory frame, but implementation requires institutional courage. Redzone data—biometric information, health records, financial data—must be categorically excluded from foreign processing. This constitutes not protectionism but risk management. Foreign intelligence laws’ extraterritorial reach, commercial artificial intelligence training practices’ opacity, and technology supply chains’ geopolitical volatility create unacceptable vulnerabilities for sensitive citizen information. Data Protection Impact Assessments should be mandatory for every artificial intelligence agent, not bureaucratic checkboxes but genuine evaluations reviewed by independent authorities. Privacy-preserving defaults—technical architectures minimizing data collection and retention—must be engineered in, not added later.
Vendor strategy in this environment requires sophistication procurement departments often lack. The market will offer subsidized pilots, “sovereign cloud” offerings merely foreign infrastructure with local branding, and long-term contracts with opaque pricing escalators. The response must be systematic triage: evaluating vendors against strict sovereignty criteria, preferring open-source solutions with portable code, and mandating contractual provisions for data escrow and orderly exit. The goal is not autarky but optionality—maintaining capacity to switch providers, modify systems, and sustain operations independently if commercial or political conditions change.
Delivery discipline separates serious reform from perpetual pilot projects. The temptation will be announcing comprehensive “artificial intelligence for governance” initiatives covering dozens of departments. This approach has failed before. Better identifying sharp, measurable use cases—Right to Information request processing, grievance redressal, multilingual helplines—and delivering them completely in controlled environments. Quarterly capability increments, rigorous performance metrics, and aggressive sunsetting of underperforming pilots create feedback loops necessary for genuine learning. Components proving themselves should be abstracted and reused; failure should be fast and documented, not hidden behind press releases.
Safety and transparency mechanisms must be institutionalized, not assumed. Artificial intelligence systems encode and amplify existing biases; they generate confident falsehoods; they create accountability gaps when decisions are attributed to “the algorithm.” Mitigation requires continuous bias testing, diverse training data, and mandatory human-in-the-loop protocols for consequential determinations. Model cards documenting capabilities and limitations should be public documents. Audit logs should be accessible to oversight bodies and, where appropriate, to affected citizens. Appeals mechanisms must be clear and functional—the right to human review when artificial intelligence systems make adverse determinations is essential for democratic legitimacy.
The international context adds urgency. Major economies race to establish sovereign artificial intelligence capabilities, not from hostility toward global cooperation but from recognition technological dependence creates strategic vulnerability. India’s position is unique: we possess talent, data scale, democratic legitimacy, and institutional experience with digital public infrastructure leading this movement. What we risk lacking is time. Every major foreign artificial intelligence system procurement for governance functions creates path dependencies—training data flows overseas, staff capacity atrophies, vendor relationships become entrenched. The window establishing sovereign alternatives narrows with each contract signed.
The choice is not between modernity and nostalgia, or efficiency and control. It is between a future where India’s governance is augmented by tools we understand, govern, and modify—and one where critical state functions operate on blackbox infrastructure we rent but do not own. The kiosk in the district office speaking local language, accessing verified records, rendering decisions subject to appeal constitutes not technological vanity. It tests whether world’s most complex democracy can maintain mastery of its administrative transformation. Hardware is affordable. Architectures are proven. What remains is institutional will building rather than buying, coordinating rather than fragmenting, treating technological sovereignty prerequisite for democratic continuity.
-
Brijesh Singh is a senior IPS officer and an author (@ brijeshbsingh on X). His latest book on ancient India, “The Cloud Chariot” (Penguin) is out on stands. Views are personal.
