Thank you dear subscribers, we are overwhelmed with your response.Your Turn is a unique section from ThePrint featuring points of view from its subscribers. If you are a subscriber, have a point of view, please send it to us. If not, do subscribe here: https://theprint.in/subscribe/
In February 2026, at the India AI Impact Summit, Union Minister Ashwini Vaishnaw laid
out a five-layer stack for India’s AI ambitions: energy, infrastructure, compute, models, and, at the top, application. That top layer is where AI actually reaches a citizen, and Vaishnaw said India could lead the world in deploying it at scale.
The budget disagrees with him. Of the India AI Mission’s ₹10,372 crore, compute takes ₹4,563 crore and the Innovation Centre that funds indigenous models takes ₹1,971 crore, together about 63 percent. The Application Development Initiative, the pillar meant to build that top layer, gets ₹689 crore: about 6.6 percent, a tenth of what compute and models draw. At the same summit, panellists from the Gates Foundation, the Boston Consulting Group and the UNDP made a point the budget does not reflect: the binding constraint on AI’s public value is adoption and institutional capacity, not the model.
None of this argues for abandoning compute or models. India simply cannot win this race
by spending: OpenAI and its partners have committed roughly US$500 billion to
compute through Stargate alone, against the entire India AI Mission’s US$1.2 billion.
India does not need to close that gap, because frontier capability is commoditising fast,
and what was proprietary a year or two ago now ships in open weights.
What stays scarce is the harness around the model, the orchestration layer that turns raw intelligence into something that finishes a task on a citizen’s behalf. That layer is where India already has an asset few countries can match: live rails at population scale. Aadhaar processed 231 crore authentications in November 2025 alone. UPI handled 2,272 crore transactions in June 2026. DigiLocker’s own counter shows more than 900 crore issued documents.
AnAI agent needs no new infrastructure to act for a citizen, only permission to ride the
infrastructure that already exists. Consider Direct Benefit Transfer, India’s flagship welfare system, which routes funds across more than 300 schemes run by 56 ministries. DBT digitised the payment. It never digitised the journey to get there. A citizen still has to discover which scheme she qualifies for, work out the eligibility rules, assemble documents, and fill in the forms. The myScheme portal lists what she may be owed, but it only redirects her; it cannot apply on her behalf.
Parliament’s own Public Accounts Committee has flagged high rates of Aadhaar biometric-verification failure that quietly exclude the elderly and manual labourers from benefits they are entitled to. Most public-sector AI deployments so far have been chatbots, not agents. Jugalbandi, the multilingual scheme-information bot built with Microsoft Research and AI4Bharat, could tell a villager which scheme exists. It could not act on it. An agentic layer flips that: instead of answering a question, it does the legwork a citizen cannot do herself, pausing
to ask her authority at each step the law reserves to her. A citizen's own agent could scan eligibility across hundreds of schemes, pull the right documents from DigiLocker, and
prepare the application, then submit it once she authenticates through Aadhaar and see
any grievance through to resolution. Where biometrics fail, as the PAC found they do at
high rates, the agent should route her to a human path rather than silently drop her from
the system.
The Digital Personal Data Protection Rules, notified in November 2025, already give this
a fixed date: their consent-manager provisions come into force on 13 November 2026,
giving a citizen a registered, auditable way to grant and withdraw permission over her
own data. That is the missing half of the trust layer an agent would need, and it arrives in
fourteen months.
The fix does not require new ambition, only different sequencing. Fund the Application
Development Initiative as infrastructure, not as a scatter of pilots, the way UPI itself was
built as shared public rails rather than left to any one bank. A common, model-agnostic
agentic layer, open to any provider, could carry every scheme built on top of it. Vaishnaw
put application at the top of the stack. Only the budget still needs to follow him there.
Ayan Pahwa builds and ships production agentic AI systems. He writes about his work at
codensolder.com.
These pieces are being published as they have been received – they have not been edited/fact-checked by ThePrint.