PRAJYOT MAINKAR
At a taluka office in most of India, the real gatekeeper is not the clerk. It is the register. There’s this fat old ledger going soft at the edges, the kind where someone once scribbled your name – maybe they even knew how to spell it, maybe not. One tiny letter decides everything: Sanjay or Sanjai, Lakshmi or Laxmi. That’s all it takes for the government to decide whether you exist. Anyone who’s stood at that counter watching the clerk squint from your Aadhaar card to some yellowed entry from years ago knows that quiet panic when it’s almost a match. Seventy percent similar. Close enough, maybe. The guy just shrugs, picks the one that looks right, and your whole afternoon gets saved by a guess.
I keep thinking about that guess as the government prepares to hand the whole apparatus a brain.
The helpful machine
Earlier this year, the Ministry of Electronics and Information Technology empanelled six companies to bring artificial intelligence into ministries, states and public sector units. The list was chosen from a field of around eighty bidders that reportedly included Deloitte, EY, PwC and KPMG, and the winning bids were, tellingly, the cheaper ones. One of the six is CoRover, the firm behind the government’s own flagship chatbots, and is now a pre-approved partner for teaching the government how to think.
The pitch is that six firms will finally teach the ministries how to use AI. The quieter worry runs the other way. A chatbot built on generations of mismatched records does not become wise. It becomes wrong faster.
Ask ‘AskSarkar,’ the citizen-information bot, a plain question about starting a business, and it is said to answer with a bureaucrat’s reflex: a bouquet of schemes, most of them about housing and cleanliness, almost none about what you asked. Ask ‘AskDISHA’ to book a ticket and, on a bad day, it deposits you back at the ordinary IRCTC page to scroll and select by hand, like a waiter who walks you to the kitchen and points. These tools were meant to spare us the labyrinth. Often they simply narrate it.
It is tempting to blame the machines. That would be a mistake.
The bottleneck was
never the interface
Here’s the problem. An AI assistant is only ever as good as the records beneath it, and India’s public records are a stack. The land documents at the bottom descend from colonial surveys; the welfare rolls above them are younger than the Republic. Digitising all of it didn’t clean it; it photographed the old paper, errors and all. Different ministries, different eras, different spellings, different ideas of what a citizen’s name even is. The government has admitted as much itself: its own draft data governance policy conceded, in the flat language of officialdom, that government data is managed and accessed in inconsistent ways across departments. That single sentence explains more about the chatbot’s confusion than any critique of its software could.
Consider the ordinary miracles that already fail. Millions of ration cards have been cancelled over the years for want of a clean Aadhaar link, some genuinely fake, many belonging to people very much alive and hungry. Studies of biometric authentication in Jharkhand and Rajasthan found large shares of beneficiaries unable to collect their grain because a fingerprint would not read or a record would not match. In one Odisha case, an eleven-year-old boy’s family lost its food entitlement because the name on the ration card and the name on Aadhaar did not agree. A spelling error is a small thing until it is the reason a child goes without rice.
Now imagine layering a confident, fluent language model on top of that. It will not repair the mismatch. It will explain it to you politely.
What procurement
teaches the machine
Oddly, the deeper flaw is not technical at all. It is cultural, and it lives in how the government buys things. Public procurement in India runs largely on the logic of the lowest bid. Qualify technically, then win on price: a rupee cheaper and the contract is yours. It is a system designed to prevent theft, and it does. What it also does is reward the smallest viable thing that can be delivered on paper. Build the portal, tick the box, move on. Nobody’s incentive is aligned with the citizen who will actually stand at the counter.
This is why India has a thousand portals that do not speak to one another. Each was procured separately, built separately, and declared a success separately. The health mission has digitised hundreds of millions of records that still struggle to travel across state lines. The vaccine platform, once celebrated, spent stretches throwing errors no ordinary user could decode. Every one of these was, in its own procurement file, a triumph.
An AI copilot dropped into this culture does not transcend it. It inherits it. Trained on siloed, duplicated, contradictory data, and procured under the same haste, it learns the institution’s confusions and repeats them with a straight face. Garbage in, as the old line goes, except now the garbage arrives beautifully phrased.
The thing India actually got right
And yet it would be dishonest to be only gloomy, because India has built something the world envies. The payments system works. You can send ten rupees to a vegetable vendor with a phone and a QR code, and it settles in seconds, and it rarely fails. Digital identity reaches nearly every adult. Direct benefit transfers have moved staggering sums with real savings.
Why did those succeed where the chatbots stumble? Not because the technology was smarter, but because someone did the boring, unglamorous work first: they agreed on standards. They built rails that everyone could run on. Interoperability was the whole point, not an afterthought. Payments worked because the plumbing was fixed before the taps were installed.
That’s the awkward truth sitting under the AI rush. Nandan Nilekani, who helped create much of that basic plumbing, has been saying lately that the future won’t go to whoever trains the biggest model. It’ll go to whoever cleans and sorts the data underneath, small systems fixing real problems for real people. The analysts at Carnegie and the consultants at EY, who rarely agree on anything, agree on this:
India’s missing layer is not compute or clever models. It is governed, trusted, connected data.
What would actually work
So what would a useful government AI require? It would require someone, somewhere, deciding that Sanjay and Sanjai are the same man, and making that decision hold across every
database that touches his life. It would require ministries to surrender their private spreadsheets to a common standard. It would require
procurement that pays for outcomes a
citizen can feel, not features a file
can list. None of this photographs well at a launch event. All of it matters more than the model.
Perhaps that is the quiet irony of the whole enterprise. We are trying to buy intelligence for a system whose real problem is memory, a state that
cannot reliably recall who its people are. A chatbot cannot fix that. It can only stand at the same softening
register, squinting at the same
near-match, and, like the tired operator, make its best guess. The difference is that it will sound certain
while it does.
