Equitable agency: The case for an AI agent for every Indian

Her name is Lakshmi. She lives in a small village in the Krishna delta, Andhra Pradesh, where the fields flood in September and the nearest bank branch is a one-hour bus ride away. She wants to buy a buffalo, not as an aspiration, but as a business plan. A government scheme could fund it. But she cannot read or fill a form and has no one to help her navigate the paperwork.

The bank exists. The scheme exists. The money exists. The gap is not financial. It’s procedural. And procedural complexity, invisible to those of us who navigate it daily, is one of the most efficient destroyers of wealth in India.

This is the problem at the heart of a white paper I co-authored with colleagues from MIT, IIT Kanpur, IISc and other institutions for the India AI Impact Summit, titled ‘Doot: The AI Agent for Every Indian Citizen.’ It outlines the architecture for an AI agent designed to work for every Indian citizen.

Lakshmi talks to the Doot app on her smartphone about her need for a loan. Her personal AI agent finds the right scheme, verifies her identity through Aadhaar KYC, retrieves her income and eligibility certificate documents from DigiLocker and initiates the loan through the bank’s systems automatically, without requiring her to understand and fill out forms by hand, and without a middleman.

India’s digital public infrastructure (DPI) stack is formidable: Aadhaar authenticates a billion identities and UPI moves ₹250 trillion annually in payments. What is missing is the agentic layer that converts citizen intent into government process and banking workflow.

With sovereign large language models like Sarvam now capable of nuanced, culturally grounded conversation across India’s major languages, the linguistic barriers that kept Lakshmi outside the system can be bridged through a conversational AI app.

India’s digital plumbing is remarkable. The next step is making it truly intelligent and agentic: understanding the person it represents, their context and history, and then navigating the full complexity of what lies beneath: KYC norms, bank regulations, government eligibility rules, scheme documentation requirements.

Agentic AI doesn’t just converse. It can act, the way a knowledgeable and tireless community banker would. And it can be instructed in every Indian language via a low-cost smartphone.

But digital capability alone is not enough, it needs safeguards. At a billion-user scale, the architecture must be safe, sovereign and accountable. Every Doot agent is cryptographically bound to the citizen’s identity, with all authority flowing from the citizen, not the platform.

Consent is granular, time-limited and revocable. For high-stake actions like financial transfers, the agent requires explicit human permission to proceed. A policy layer encodes law, financial limits and data protection rules, ensuring that the agent can only act within legal and regulatory bounds. All actions must be logged in tamper-evident audit trails visible to citizens.

Even though all of this sophisticated digital orchestration takes place behind the scenes, Lakshmi is completely shielded from its complexity.

She controls her transaction through a simple, intuitive conversational interface in her language. Her agent tells her in plain language what it is about to do and waits for her voice confirmation before acting. It proves her eligibility without exposing her Aadhaar number. It accesses her income and eligibility certificates from DigiLocker only after getting her consent. When the disbursement arrives via UPI, she is immediately alerted, with an audit trail.

Lakshmi’s story is not exceptional. Across India, millions of capable, resourceful people are locked out of systems that could genuinely help them—not because they lack cognitive capabilities, but because the digital interface was never built for them.

If unchecked, AI can deepen divides in society. If we allow AI to develop only for those with education, English fluency and high-end smartphones, we risk widening historical divides that we are trying to close. The solution is AI designed as DPI—open, sovereign, identity-anchored and privacy-first—that uses conversational AI to bypass literacy and complexity barriers, thus enabling marginalized society to participate in mainstream economic activity.

Loan applications are just the beginning. Imagine the same Doot agent helping Lakshmi book a doctor for her ailing mother at a government health centre, navigating ABHA records and available appointments through a voice request. Or guiding her teenage daughter through a scholarship application, finding the right scheme, checking eligibility and submitting the forms, all in the local language. Healthcare, education, livelihoods can all be handled by the same agent, with the same simple conversation—the full weight of India’s DPI working quietly on her behalf.

Lakshmi does not need charity. She needs a tool that speaks her language, understands her world and navigates procedural complexity and completes tasks. The difference between those two things is the difference between a handout and a future.

The author is former founder CTO of Aadhaar, CEO of Khosla Labs and a research affiliate, Massachusetts Institute of Technology.