From Digital India to Intelligent India: Three tests for the next decade

By March, India had empanelled more than 38,000 GPUs under its subsidised national AI-compute programme. Another statistic may be more revealing. Karnataka says ₹115 crore was credited to the accounts of nearly 1.1 lakh deceased beneficiaries of its Gruha Lakshmi welfare scheme, largely because information about their deaths reached the payment system too late.

It is not only Karnataka’s problem. India’s Comptroller and Auditor General has found pensions flowing to the dead and the ineligible in other schemes. The next decade of Indian technology will be judged less by how much AI the country acquires than by whether the state can move a basic fact, a birth or a death, from one register to another, act on it, and reverse course when the record is wrong.

Few people drafting Digital India in 2015 could have expected UPI to process 24.5 billion transactions in a single month, as it did this August. The useful question, then, is not what technologies India will have in 2035, but what would make the state meaningfully more intelligent.

Digital India built rails for identity, payments, documents and public services at a scale few countries have matched. But access does not guarantee outcomes. Citizens still have to discover the scheme, establish their eligibility, resubmit information another department already holds and chase the benefits.

An Intelligent India would use what the state already lawfully knows to anticipate routine needs, notice when a programme is failing in practice, and use AI where fixed rules are insufficient for messy information, language and patterns that cannot easily be specified in advance. Consequential judgment would remain accountable to people.

Three tests would show whether that transformation is actually happening.

Does the citizen still have to ask?

A birth, a 60th birthday, a death in the family: the state often records these events, then waits to be told about them again.

In Austria, registering a child’s birth can trigger the family allowance without an application from the parents. Haryana has experimented with a similar principle, using its family database to determine eligibility for some pensions without requiring citizens to apply.

None of that requires artificial intelligence. That is precisely the point.

Rules should execute rules. A government that reaches for a language model before connecting its registers is confusing sophistication with intelligence. AI becomes useful where the problem is harder to specify in advance: interpreting an unstructured document, communicating across languages, helping an official navigate an unusual case or finding patterns hidden across thousands of interactions.

The first test of an intelligent state is therefore not how much AI it uses. It is how much unnecessary work it stops asking citizens to do.

Does the system notice what isn’t working?

A conventional digital system can reject the same eligible people for the same obscure reason thousands of times without anyone asking why. It can administer a scheme with unusually low take-up in one district and treat each missing beneficiary as an isolated absence.

This is where AI can add something that ordinary digitisation cannot, and India has begun to use it. The central grievance portal, CPGRAMS, now uses AI-enabled analytics to categorise grievances, identify trends and help officials look for root causes.

The more important question is how often those insights change a rule, a form or a process.

Suppose thousands of complaints in several Indian languages turn out to involve the same document. Or pension enrolment in one district falls far below what its demographics imply. Finding the pattern is useful. An intelligent system will be able to also flag the issue and recommend a course of action.

Can it explain itself, and be overruled?

Karnataka’s overpayments illustrate one side of automation: a system that fails to stop paying when it should. Connecting death records more quickly to welfare databases could prevent much of that waste.

Haryana shows what happens when the record is wrong. In 2023, the state government told the assembly that over 63,000 old-age allowances had been stopped on the basis of its family-ID data. After verification, about 44,000 were restored.

The asymmetry is simple. A wrong yes costs the state money that it may be able to recover. A wrong no can cost a citizen her income, and she must do the recovering herself.

That means correction cannot be an afterthought to automation. Where an automated system stops or reduces a social benefit, the citizen should receive a reason she can understand and have access to a person with the authority to reverse the decision within a defined, short period.

The standard will vary with the stakes. AI translating a government webpage does not require the controls appropriate to a system flagging welfare fraud. A model helping an officer summarise a file is different again from one whose output determines whether a household gets paid.

The hardest constraints here are institutional rather than computational. Building an intelligent state requires ministries and states to share data they may treat as power. It requires officials to accept that making correction easy also makes errors more visible. And it requires governments to fund reviewers, appeals and data quality as readily as they fund computing.

There are signs that the prerequisite is understood. The Ministry of Statistics and Programme Implementation has been pushing states to harmonise administrative data, move beyond departmental silos and make datasets interoperable and linkable by design. That is less glamorous than buying GPUs. It may matter more.

The three tests may therefore be more useful than any forecast of what AI in government will look like in 2035.

Ask again, what happens when a woman becomes eligible for a widow’s pension. Does the state notice without waiting for her application? If women like her are systematically falling through the cracks, does the system detect the pattern and force someone to examine why? If she is wrongly refused, can she find out what went wrong and reach a person who can put it right quickly?

She may never know whether any of that involved artificial intelligence. She will know whether the state became more intelligent, more responsive, more concerned.

(The author is Country Director India at the Tony Blair Institute for Global Change. Views are personal.)