How I Used AI In An Indian Election?

What do you see here?

A stack of papers.

Nothing particularly exciting about that. We have all seen stacks of paper. Government offices are full of them. So are lawyers' chambers, railway booking counters and the desks of men who have promised to read something but never quite got around to it.

But this particular stack is different.

It is a 638-page document.

And if you ask me what it is, I would call it something far more useful than a report.

It is the Kundli of a constituency.

And before you accuse me of being dramatic, let me explain.

Recently, I met a very prominent Member of Parliament in Guwahati. He wanted me to take him up as a client. During our conversation, the subject of artificial intelligence came up.

He told me, rather confidently, that he had access to the best Large Language Model in the world.

That was when I realised that there was probably no stopping his party from losing.

Not because the man was stupid. Quite the contrary. But because he had misunderstood the nature of the beast.

AI is a tool. It is not a political strategy.

Give the world's most powerful LLM a lazy prompt asking it for "10 campaign ideas to defeat anti-incumbency" and you have achieved roughly the same thing as using a chainsaw to trim your beard.

Technically possible. Strategically idiotic.

The real question is not which AI model do you use?

The real question is:

What do you make the machine do?

That is where my story begins.

The Problem

I was working in Assam when I first began seriously experimenting with AI for political intelligence.

The problem was obvious.

AI hallucinates.

It hallucinates even in English. Feed it data in Assamese, particularly messy, field-collected, bilingual Assamese-English data, and its imagination can become positively literary.

That is rather inconvenient when you are dealing with elections. Because a fabricated sentence in a blog post is embarrassing. A fabricated voter statistic can cost you an election. So before I tell you what I did with AI, let me tell you what I was trying to solve.

My client was a very well-known political leader. He was secular. But, as has happened to many secular politicians in India, his secularism had been painted by his opponents as "minority appeasement." Mahatma Gandhi, incidentally, had endured a version of the same accusation. The difficulty was that my client was contesting from an Assamese Hindu-majority constituency. The narrative was beginning to stick. There was discomfort among sections of voters. There was also the familiar disease of electoral politics: anti-incumbency.

My job was simple to describe and extremely difficult to execute:

Understand what people were actually thinking—and prevent a political narrative from becoming an electoral reality.

46,000 Households

There were approximately 46,000 households in the constituency. We assembled a team of more than 100 fresh graduates. More than 90 per cent of them were women. Within a month, we surveyed more than 28,000 households.

Twenty-eight thousand is not a focus group. It is not a sample of fifty "influencers" sitting in an air-conditioned hotel room. It is a sizeable chunk of an actual constituency.

We collected responses across 22 data points.

There was nothing revolutionary about conducting a household survey. Political consulting firms have been doing surveys for years. The difference was what happened after the survey. Most political agencies would eventually produce a neat 20- or 30-page constituency report. There would be graphs. There would be pie charts. There would be a section called Key Findings. Everyone would nod gravely. And then someone would put the report in a drawer.

I wanted to go considerably deeper.

The Constituency Was Not One Constituency

A constituency is not a homogeneous geographical unit. It is an accident of political cartography. Inside those boundaries live thousands of different Indias.

A Hindu household is not necessarily politically identical to another Hindu household. A farmer and a government employee do not experience the same state. Two families belonging to the same caste may have completely different political loyalties because one received a government scheme and the other did not.

Income matters. Occupation matters. Religion matters. Caste matters. Gender matters. Age matters. Political history matters. And, very often, the combination of two apparently unrelated variables matters even more.

So we began creating intersectional segments. Religion × caste × income × occupation × employment × welfare benefits × political leaning—and much more.

We were no longer asking: "What does the constituency think?"

We were asking: "What does this particular kind of household think, and where exactly are these households located?"

Every household was geo-tagged. That changed everything. Suddenly, the constituency stopped looking like a coloured map in an election office. It began looking like thousands of individual political stories stitched together. And patterns started emerging.

Then Came the AI Problem

There was, however, one small difficulty. I was doing all this largely on my own. The data was bilingual. English and Assamese.

The field data was not produced by Oxford scholars. It was produced by human beings in the real world—which meant inconsistent spellings, variations in terminology, duplicated entries, different ways of recording the same response and all the other small disasters that make data scientists reach for aspirin. So the first job was not to ask AI for insights.

The first job was to clean the data. We standardised the variables. We created uniform categories. We removed inconsistencies. We structured the bilingual dataset so that machines could understand what human beings had actually recorded.

Only then did I bring in AI.

But I did not trust one model. That would have been rather like asking one astrologer to predict the future and then blaming Mercury when he got it wrong. Instead, I created what I jokingly called my Council of LLMs. It consisted of the major models available to me at the time — ChatGPT, Gemini, Claude, Grok and their various versions.

Each model was given the same structured data. Each was asked to independently analyse the constituency. They produced their findings separately.

Then came the interesting part.

I compared the outputs. Where they agreed, confidence increased. Where they disagreed, I investigated. Where one model produced an extraordinary insight, I checked whether it was actually supported by the underlying data or whether the machine had simply invented a plausible-sounding story.

In other words, I did not ask AI to tell me the truth. I made AI argue with AI.

And then I checked the argument myself. The machine was not the strategist.

It was the research assistant. And a rather argumentative one at that.

From 28,000 Households to 638 Pages

The field team completed the data collection in the first month. The second month was spent doing something far less glamorous: reading the constituency at microscopic scale.

We went down to the Panchayat and Municipality level. We analysed the intersections.

We examined the geographical distribution. We identified patterns. And eventually, we produced a 638-page constituency report.

Why 638 pages?

Because the constituency had not asked us to produce a 30-page report. It had given us 28,000 households. The report was simply the consequence of taking those households seriously. And because each household had been geo-tagged, the data did not become useless when polling-station boundaries changed. A polling station could be shifted. Another could be added. The administrative map could change. The underlying household-level intelligence remained. That, to me, was the real value of the exercise.

We had not created a survey.

We had created an electoral intelligence database.

The Data Eventually Had to Become Politics

Of course, nobody votes for a spreadsheet. People vote for people. Data can tell you where the problem is. It cannot shake hands with a voter. It cannot sit in a Namghar. It cannot listen to an old man complain about his road. It cannot make a family believe that its MLA is one of their own. That is the job of politics.

Our research showed us that the opposition narrative had created genuine discomfort among sections of the electorate. So we did not try to fight every accusation with another accusation.

We changed the emotional frame.

The campaign gradually moved towards a simple proposition: "No matter what, he is our son."

The politician stopped being an abstract political figure. He became local. Familiar. Accessible. Part of the community. We made him visit every Namghar and Satra in the constituency. His relationship with the community became visible. People began talking not merely about his politics, but about his devotion and his contribution to the community.

The narrative changed. And eventually, so did the electoral arithmetic.

My client won his Assembly seat.

The Lesson Is Not About AI

That is the part people will probably misunderstand.

They will read this and say: "Ah, so AI helped him win an election."

No.

That is not the lesson.

The lesson is that AI is only as intelligent as the system in which you place it. Give it an unclean dataset and it will give you beautifully written nonsense. Give it vague instructions and it will give you generic campaign ideas. Give it the question "How do I win this constituency?" and it will probably produce the same list of ten ideas it has produced for a thousand other consultants. But give it structured, granular, geo-tagged field data. Give multiple models the same problem. Make them challenge one another. Cross-check their conclusions against the underlying evidence. Then combine machine intelligence with human political judgment. Now you have something interesting.

You have a machine that can help one political consultant do the work of a much larger research operation. That is the real disruption.

The Big Agencies Are Not Magicians

There is an unfortunate tendency in politics to look at organisations such as I-PAC and assume that they possess some secret technology unavailable to ordinary political consultants. They don't. They have people. They have systems. They have money. They have data. They have processes. They have experience. And, most importantly, they know how to turn information into political action.

The interesting question is not whether a small political consultant can suddenly become I-PAC. He cannot. The interesting question is whether technology allows a small team—or even a determined individual—to compress the gap. I believe it does.

The 638-page Kundli sitting in front of me is proof of that possibility. It represents 100-plus enumerators. Twenty-eight thousand households. Twenty-two variables. Thousands of geographical coordinates. Multiple AI models. Hundreds of hours of analysis. And an enormous amount of human judgement. But there is one thing it does not represent. It does not represent the victory of AI over human intelligence. Quite the opposite.

It represents what happens when human intelligence learns how to command machines properly.

So, the next time a politician tells me that he has the best LLM in the world, I shall resist the temptation to ask him which one. I shall ask him a much simpler question: "Fine. What have you made it do?"