Election Duty in Tamil Nadu

The alarm goes off at 4:30 AM in Girnar Guest House in Kaushambi. It is the 18th of March 2026. I have packed the night before. Three copies of my appointment letter. Six passport-size photographs. A hardbound diary. Blue pen, black pen, red pen. The red one, I suspect, will see the most use.

On my laptop, my phone, and a pen drive, there is a PDF I have been reading for two days. The Compendium of Instructions on Election Expenditure Monitoring, December 2024 edition. Five hundred and some pages. It is going to be my reference manual for the next forty-five days.

I have been following money trail in Income Tax. Today, I am going to follow money in a different place.

The Election Commission of India has appointed me as Expenditure Observer for the Tamil Nadu Legislative Assembly Elections, 2026. My constituency is 86-Edappadi, Salem District. 

My job is straightforward. Every candidate contesting this seat can spend a maximum of Rs. 40 lakh under Rule 90 of the Conduct of Election Rules, 1961. I am here to make sure that limit is respected. And to catch the spending that candidates would rather not declare: the cash, the liquor, the freebies, the vehicles that never make it into any register.

Different domain. Same discipline. Follow the money.

The flight leaves Delhi at 7:05 AM. I land in Bangalore by mid-morning and make my way to Salem by a connecting flight. The landscape changes as you cross into Tamil Nadu. The soil turns red. Granite quarries appear on hillsides. Steel plants. Sago factories. Roadside stalls selling tender coconut. The signboards switch to Tamil, a script I cannot read.

Salem is bigger and more industrial than I expect. Thirty lakh people live in this district. Eleven Assembly constituencies. One Lok Sabha seat. A steel plant, a river, a busy corporation. This is not a sleepy town. This is a serious place.

A blue-beacon Innova is waiting for me. My Liaison Officer, Thiru. Duraisamy, and my PSO, Thiru. Samuel, are standing beside it. In the grammar of Indian government, a blue-beacon vehicle is a statement. It says: the Election Commission is here. Word travels in a constituency. By evening, people in Edappadi will know that the Expenditure Observer has arrived.

That is by design.

Most people think of elections as rallies and speeches and ballot boxes. They do not think about what runs underneath.

India's election expenditure monitoring system is, quietly, one of the most ambitious real-time financial surveillance operations in any democracy. Flying Squads patrol highways at 3 AM, stopping vehicles and checking for cash. Static Surveillance Teams sit at junctions with cameras and logbooks. Video teams record every rally and public meeting so that someone can later count the chairs, measure the pandal, estimate the cost, and ask: did the candidate report this?

Check posts on every major road. An Election Seizure Management System that logs every seizure of cash, liquor, drugs, or freebies in real time. Shadow registers that track observed expenditure independently of what candidates declare.

And at the centre of this system, an Expenditure Observer in every constituency. 

I meet the police first. The Superintendent of Police, Salem.

We sit across a table and talk about money. Not money in the abstract. Cash movement corridors through the district. Drug transit routes that get repurposed during elections for distributing inducements. Liquor supply chains. The geography of how money and material flow into a constituency when an election is on.

We exchange phone numbers. That is the real deliverable of the meeting.

Then the Collector. District Collector and District Election Officer, Salem. 

She manages eleven constituencies simultaneously. She knows the machinery.

I also meet the Commissioner of Police. That is three institutional meetings in a single afternoon. Three sets of phone numbers saved. Three relationships established.

The foundation is being laid.

Then I leave Salem for Edappadi.

Forty-two kilometres west. The road passes through Tharamangalam. The landscape opens into the kind of small-town Tamil Nadu that runs on agriculture, granite, and local commerce. The Mettur Dam, one of the largest in India, sits at the edge of this constituency.

I arrive in the late evening. The temperature is 36 degrees. My accommodation is an Inspection Bungalow. 

I meet the Assistant Expenditure Observer and the Returning Officer's team. These are the people who will actually run the expenditure monitoring on the ground every day. Block-level officials. Taluk-level officers. People who know every lane in this constituency.

I explain the shadow register. This is the most important tool in my kit. A parallel ledger maintained independently by my team, recording every observed campaign activity and its estimated cost based on the DEO's rate charts. When a candidate declares they spent fifteen lakh and my shadow register shows thirty-five lakh of observed activity, that gap is where the questions begin.

I set up the daily reporting protocol with the AEO. Summary every evening by 8 PM. Seizures reported immediately. No delays. No filtering.

They save my number. I save theirs. And a whatsapp group is created.

Here is what I have learned on Day 1.

The machinery exists. The teams are being assembled. The infrastructure is taking shape. Salem district has a competent Collector, a sharp SP, and a police commissionerate that is engaged. At the constituency level, the AEO and RO team are in place.

Whether the machinery works under pressure is a different question. That is what the next four days are for. I will inspect check posts. I will visit Flying Squads in the field. I will meet State Excise officials. I will walk through the Expenditure Sensitive Pockets identified in this constituency. I will look at registers and formats and verify that the teams are not just formed on paper but actually functioning on the ground.

There is something I keep thinking about.

In tax administration, the stakes are revenue. Important, certainly. A nation runs on what it collects.

In election monitoring, the stakes are the process itself. Every rupee of unaccounted expenditure is a small crack in the integrity of a democratic exercise. Every cash envelope distributed to a voter is an attempt to convert a citizen's choice into a transaction. The expenditure monitoring system exists to hold that line.

I am not being dramatic about this. It is just a fact. The line holds because someone is watching. The observer is not only a checker of registers. The observer is a signal that someone is counting.

A note on language.

I do not speak Tamil. This is, to put it mildly, a limitation when you are posted in Tamil Nadu.

But I have discovered that vanakkam, said with folded hands and a genuine smile, opens every door. That nandri, said often, builds goodwill faster than any official letter. And that sari, meaning OK, meaning understood, meaning let us proceed, is the single most useful word in the working vocabulary of Indian governance.

My Liaison Officer, Thiru. Duraisamy, handles the rest with a patience I find remarkable. The block-level officials speak enough English that we get by. In government, the real language is not Hindi or Tamil or English. It is the language of registers, formats, reports, and deadlines. That language, I speak fluently.

It is late now. The ceiling fan turns. The night outside is warm and full of the sounds of a Tamil Nadu small town: insects, a distant temple speaker, the occasional auto-rickshaw.

I open my diary and write the day's observations. Four meetings. Two cities. One overnight transformation from a tax officer in Delhi to an election observer in a constituency I had never visited before this morning.

I have four more days in this visit. Then back to Delhi for a week. Then back here on the 30th of March when the gazette notification drops. Then continuously from the 5th of April through polling day on the 23rd.

The work has begun.

Vanakkam, Tamil Nadu. 

India's First Global AI Summit

The first-ever global AI summit hosted in the Global South just kicked off in Delhi. And I was there on the expo floor yesterday.

Bharat Mandapam was buzzing — 300+ exhibitors from over 110 countries, 20+ Heads of State converging later this week, and Sundar Pichai landing in Delhi for a keynote. The India AI Impact Summit 2026 is massive by any measure.

Which of these AI tools can actually survive contact with Indian government reality?

Not the pitch deck reality. The ground reality. The 10,000-user, legacy-system, compliance-heavy, can't-afford-downtime reality.

With that lens, I walked the expo floor for hours. And seven solutions stood out — not because they had the flashiest booths, but because they addressed problems I deal with every single day.

Here's what I found.

1. Deloitte PRAGYA — When Consulting Meets AI at Scale

Deloitte's PRAGYA platform wasn't just another enterprise AI dashboard. What caught my attention was how it bridges the gap between strategic advisory and operational execution.

In government, we've seen no shortage of consulting reports that gather dust on shelves. What we need are tools that take a recommendation and help you implement it — across hundreds of offices, with varying levels of digital maturity, in real time.

PRAGYA seems to be built for exactly that kind of complexity. For large-scale government transformation programs — the kind where you're implementing a new Act across an entire department — this is the type of AI-assisted project intelligence that could change how we manage reform.

2. LKS In-House Litigation Management — This One Hit Home

If you've ever managed litigation portfolios in government, you know the pain. Thousands of cases. Multiple courts. Overlapping deadlines. Paper trails that would fill a warehouse.

I've spent years in ITAT litigation and transfer pricing disputes both in Mumbai and Amritsar, including cases worth hundreds of crores. The single biggest bottleneck isn't legal strategy — it's tracking, coordination, and institutional memory.

LKS has built an AI-powered litigation management system designed for in-house legal teams (as of now). Automated case tracking, deadline management, outcome analysis, and pattern recognition across your case portfolio.

For the Income Tax Department — which handles lakhs of cases across the country at any given time — a system like this isn't a luxury. It's an operational necessity. The question isn't whether we need it. It's how fast we can adapt it to our scale.

3. Smoothtalk AI — Virtual Calling That Could Transform Citizen Services

Picture this: a taxpayer in a Tier 3 city has a query about their assessment. Today, they call a helpline, wait, get transferred, explain their problem three times, and maybe — maybe — get a resolution.

Smoothtalk AI is building virtual calling agents that can handle these interactions with natural, human-like conversation. Not the robotic IVR menus we've all grown to hate. Actual contextual dialogue.

For any government department that handles millions of citizen queries — and the Income Tax Department certainly does — this technology could fundamentally reshape the service delivery experience. Imagine every taxpayer getting an intelligent, patient, context-aware agent on the other end of the line, available 24/7, in multiple languages.

We're not there yet. But the demo I saw suggests we're closer than most people think.

4. Government AI — A UK Perspective on Sovereign Deployment

This one was fascinating for a different reason. Government AI is a UK-based platform built specifically for public sector use cases.

What made me stop and engage wasn't just the product. It was the philosophy. They've clearly thought through the unique constraints of government — compliance requirements, audit trails, data sensitivity, and the fact that "move fast and break things" is not an acceptable operating principle when you're dealing with citizens' data and rights.

Seeing how another country approaches sovereign AI deployment gives useful comparative perspective. At CBDT, as we think about integrating AI into tax administration, understanding global best practices — not just Silicon Valley practices — is essential. The UK's approach to government-specific AI platforms is worth studying closely. They infact have their own courses on iGOT platform.

5. ACTUALITY — On-Premise Deployment for Data That Can't Leave the Building

This is the unsexy but critical conversation that most AI summits skip.

Every AI vendor will tell you their cloud solution is secure. And maybe it is — for a private company. But when you're dealing with taxpayer data, national security information, or sensitive government records, "trust our cloud" is not sufficient.

ACTUALITY offers on-premise AI deployment. Your data stays on your servers. Your models run on your infrastructure. Full control, full compliance.

For Indian government agencies bound by data localization requirements and handling some of the most sensitive personal data in the country, on-premise deployment isn't optional. It's the baseline requirement for any serious AI adoption. ACTUALITY understands this, and that alone puts them ahead of a lot of flashier competitors who haven't thought through the government procurement and compliance lens.

6. Sovereign AI — Local LLMs for Government Solutions

If ACTUALITY addresses where the data lives, Sovereign AI addresses something equally important: where the intelligence comes from.

Large Language Models trained on Western internet data don't inherently understand Indian tax law, Hindi administrative procedures, or the nuances of how a CBDT circular differs from a notification. They can approximate. They can't natively operate in our context.

Sovereign AI is building local LLMs designed for government use — models trained on local data, in local languages, for local administrative contexts. Data stays within borders. Models understand the operating environment they're deployed in.

This is the future of public sector AI, and India — with its scale, linguistic diversity, and digital infrastructure ambitions — should be leading this charge. The IndiaAI Mission's focus on building indigenous AI capacity aligns perfectly with what Sovereign AI is demonstrating.

I can tell you: the gap between a generic LLM and one that understands Section 148A of the Income Tax Act 1961 is not a nice-to-have. It's the difference between a tool that helps and one that creates more problems than it solves.

7. FUSKI.ai — Training Your Workforce Before You Deploy Your AI

Here's a truth that doesn't get enough airtime at AI summits: the biggest bottleneck to AI adoption in government isn't technology. It's people.

You can have the most sophisticated AI system in the world, but if the 50,000 officers who are supposed to use it don't understand it, don't trust it, and haven't been trained on it — you've just bought a very expensive piece of software that nobody opens.

FUSKI.ai builds AI-powered training modules for workforce upskilling. Custom learning paths. Adaptive difficulty. Progress tracking. The kind of structured capability building that large organizations need before they can meaningfully adopt AI tools.

When I think about our challenge at CBDT — implementing a new Act across 700+ offices with staff at wildly different levels of digital comfort — this is exactly the gap that needs filling. You can't just send a circular saying "use AI now." You need a systematic, scalable training infrastructure.

FUSKI.ai could be that infrastructure.

The Bigger Picture

Walking the expo floor, what struck me most wasn't any individual product. It was the shift in the overall conversation.

Three years ago, the question at these events was: "Should government use AI?"

Two years ago, it became: "Can government use AI responsibly?"

Yesterday at Bharat Mandapam, the question had evolved to: "How fast can we deploy AI, and what's stopping us?"

That's a seismic shift. And it tells you something about where India — and the Global South more broadly — stands in the AI landscape. We're not spectators in this revolution. We're not waiting for Silicon Valley to build solutions and then adapting them for our context.

We're building. We're deploying. We're setting the terms.

The India AI Impact Summit 2026 — anchored in the three Sutras of People, Planet, and Progress — isn't just a diplomatic gathering. It's a statement of intent. India is positioning itself as a global convenor for responsible, inclusive AI. And based on what I saw on the expo floor, the ecosystem is rising to match that ambition.

The summit continues through February 20th (in fact extended to 21st Feb for general public), with PM Modi's inaugural address tomorrow setting the tone for the main event. I'll be watching closely.

But if the expo is any indication, the future of AI in governance isn't coming.

It's already here. Walking the floor at Bharat Mandapam.

Teaching AI to Global Tax Officials

NADT had invited me to deliver two sessions as part of their ITEC programme on "Innovations in Tax Administration: Building Capacity for the Future." The morning session was about opportunity — how generative AI is reshaping everything from taxpayer services to complex cross-border investigations. The afternoon was about the guardrails — data privacy, algorithmic bias, ethical governance, and the cautionary tales of what happens when governments deploy AI without asking the right questions first.

On paper, these are two separate topics. In practice, they're inseparable. You can't talk about the transformative potential of AI in tax administration without immediately confronting the question: at what cost, and to whom?

That tension — between innovation and protection, between speed and fairness — became the throughline of my entire day in Nagpur.

The Room That Changed My Perspective

Here's what I wasn't fully prepared for: the quality of the questions.

I walked in expecting to explain foundational concepts. What is a large language model? What does RAG architecture mean? How does prompt engineering work? And yes, some of that was necessary. But the participants — officers from developing economies across Asia, Africa, and beyond — weren't just absorbing information passively. They were interrogating it.

One officer asked how you prevent an AI-assisted audit system from encoding historical biases in taxpayer selection. Another wanted to know how her country, with limited digital infrastructure, could adopt AI without deepening the digital divide. A third pushed back on the idea that AI could meaningfully assist in transfer pricing without understanding the specific regulatory nuances of his jurisdiction.

These weren't theoretical objections. They were the questions of people who would actually have to implement these systems back home, with real constraints and real consequences.

That shift — from "should we adopt AI?" to "how do we adopt it responsibly?" — told me something important. The global conversation about AI in government has moved past the hype phase. What people want now is practical, honest guidance.

What I Showed Them (And What I Learned Showing It)

In the morning session, I walked through real examples rather than hypotheticals. India's NUDGE campaign, where AI-powered behavioural nudges led to 24,678 taxpayers voluntarily revising their returns and disclosing ₹29,208 crore in foreign assets — with zero litigation. Singapore's IRAS chatbot, which saved over 11,000 officer-hours. The CBDT's ₹3,000 crore data analytics project with LTIMindtree that's building predictive capabilities at a national scale.

I also got personal. I showed them the time-savings analysis from my own transfer pricing work — how document review that used to take eight weeks can now be compressed to two, how comparability analysis that consumed a month now takes three days. The room went quiet when I put up the numbers. Not because the technology was surprising, but because the implications were immediate. Every officer in that room was mentally mapping those efficiency gains onto their own caseload.

But the moment that stayed with me came in the afternoon.

The Cautionary Tales That Matter

When I brought up Australia's Robodebt scandal — where an automated debt recovery system wrongly targeted hundreds of thousands of welfare recipients — the energy in the room shifted. This wasn't abstract anymore. These were real governments, using real algorithms, causing real harm to real people.

The Netherlands childcare benefits crisis, where an AI system flagged families — disproportionately those with dual nationalities — for fraud based on flawed data and biased algorithms, hit even harder. Some participants came from countries with similar demographic complexities. The lesson wasn't subtle: if wealthy, technologically advanced nations can get this catastrophically wrong, what does that mean for countries with fewer resources to build safeguards?

I made a point that I believe deeply: these failures weren't caused by bad people. They were caused by good people who asked "Can we automate this?" before asking "Should we automate this?" Who asked "How much will this save?" before asking "Who might this harm?"

The room didn't just nod along. They debated. They shared examples from their own contexts. They pushed me to be more specific about what "human oversight" actually looks like in practice when you're understaffed and under-resourced.

What I'm Taking Away

Three things crystallised for me during that day in Nagpur.

First, the demand for practical AI literacy in government is enormous and largely unmet. Officers don't need more keynote speeches about the "fourth industrial revolution." They need to sit down with a tool, try a prompt, see what works, understand what fails, and build intuition through practice. The most engaged moments in my sessions weren't during the slides — they were during the live demonstrations, when participants could see AI drafting a tax notice or analysing a scenario in real time. I gave a live demonstration of how we give a context in a prompt and how does it lead to a customised notice generation.

Second, developing economies have a genuine opportunity to leapfrog. They're not burdened by legacy systems the way some advanced administrations are. If they build thoughtfully — with ethical frameworks baked in from day one rather than bolted on after a scandal — they can set a global standard. That's not wishful thinking. It's a strategic possibility, and it requires exactly the kind of cross-country learning that programmes like ITEC enable.

The Question I Keep Coming Back To

Tax administration isn't glamorous. It doesn't make headlines unless something goes wrong. But it's the infrastructure that funds schools, hospitals, roads, and defence. When we get it right, societies function. When we get it wrong — through bias, opacity, or carelessness — the most vulnerable bear the cost.

So the question isn't whether AI will transform tax administration. It already is. The question is whether we — the people in the room that day, and the thousands like us around the world — will shape that transformation with the care it demands.

I left Nagpur cautiously optimistic that we will.

When AI Builds AI

There's something uniquely humbling about watching an AI agent build another AI agent in less than two weeks - and then realizing that the thing it built might fundamentally change how millions of people work every day.

Anthropic just launched Cowork, a desktop AI agent that manages files on your Mac without you having to babysit it through every step. But here's what stopped me cold: they built this entire feature using Claude Code. In approximately ten days. The AI helped build the AI that now helps us build... well, whatever we need to build.

If you're in tax administration, public finance, or really any corner of government that drowns in paperwork, you should be paying very close attention to what just happened.

The Expense Report That Changed My Mind

Let me paint you a picture that probably sounds familiar. A financial operations team receives hundreds of receipt images every month - crumpled photos from field visits, scanned hotel bills, blurry restaurant checks. Someone (usually several someones) manually extracts vendor names, amounts, dates. They sort by category. They build spreadsheets. They reconcile. Four hours of mind-numbing work that, let's be honest, nobody entered public service to do.

Now imagine this: you point Cowork at a folder of those receipts. You go get coffee. You come back to a reconciled monthly expense spreadsheet, categorized and formatted, with near-zero data entry errors. The four-hour process took fifteen minutes, and the humans involved spent that time thinking about patterns in spending rather than typing numbers into cells.

That's not a future scenario. That's available today for Claude Max subscribers on macOS.

And working at DOMS, thinking constantly about how we transform tax administration while respecting taxpayer dignity, I find myself asking: if Anthropic can build something this sophisticated in ten days using their own AI tools, what's our excuse for still requiring taxpayers to manually re-enter information the government already has? (Though we have succeeded to a greater extent with pre-filled forms)

The Recursive Loop We're Living In

Here's what fascinates me about this moment. Anthropic used Claude Code - an AI coding agent - to build Cowork - an AI file management agent. We're watching AI accelerate AI development, which will accelerate how we deploy AI, which will accelerate... you see where this goes.

In transfer pricing work in Mumbai, I've spent years analyzing patterns across thousands of transactions, looking for anomalies that might indicate profit shifting. It's intellectually demanding work that requires both pattern recognition and contextual judgment. The pattern recognition part? That's increasingly AI territory. The contextual judgment - understanding what unusual circumstances might legitimately explain an outlier, recognizing when similar-looking cases require different treatment - that's profoundly human.

But here's the thing: I couldn't do the contextual judgment part nearly as well if I was drowning in the pattern recognition grunt work. The AI doesn't replace my expertise; it creates the conditions where my expertise can actually matter.

Cowork represents a particular philosophy about this division of labor. It's not a chatbot that requires constant prompting. You specify a folder, define a task, and it works autonomously within those boundaries. It reads, creates, edits - without asking permission at every step. That's a fundamentally different relationship between human and AI than most of us are used to.

What Government Gets Wrong About AI (And Why Cowork Matters)

In my role working on India's new Income Tax Act 2025 and developing Guidance Notes at DOMS, I see two competing visions of AI in government service constantly colliding.

Vision One: AI as Institutional Efficiency Engine. Automate processing. Speed up compliance checks. Reduce headcount needs. Make government run faster, cheaper.

Vision Two: AI as Citizen Empowerment Tool. Reduce coordination costs for taxpayers. Make complexity navigable. Shift the relationship from adversarial compliance to collaborative partnership.

Most government AI initiatives, if we're being honest, default to Vision One. It's easier to measure. It fits existing budget frameworks.

But Vision Two is where the transformation actually happens.

What Anthropic did with Cowork - built primarily by AI, for everyday human tasks, designed to work autonomously once properly directed - points toward Vision Two. It doesn't make Anthropic's team smaller; it makes them more capable of building ambitious things quickly. The constraint shifted from "how many hours can our engineers spend on this?" to "what's actually worth building?"

Now translate that to tax administration. The constraint shouldn't be "how many officers can we hire to process returns?" It should be "how do we help taxpayers understand and meet their obligations with minimum friction?"

If an AI agent can take a folder of messy receipts and produce a reconciled expense report in minutes, could a similar agent help a small business owner take a folder of invoices and produce an accurate GST return? Not just fill in the forms - actually understand what qualifies, what doesn't, flag potential issues, suggest legitimate deductions they might have missed?

The technology is clearly there. The question is whether we have the imagination and will to deploy it this way.

The Ten-Day Test

Here's a thought experiment I keep coming back to: If Anthropic can build a sophisticated desktop agent in ten days using AI-assisted development, what could a well-resourced government innovation team build in ten weeks?

A taxpayer assistance chatbot that actually understands the Income Tax Act 2025's streamlined 536 sections? An agent that automatically identifies eligible deductions from uploaded financial documents? A tool that helps taxpayers model different scenarios - should I claim this under section X or Y? - with clear explanations in plain language?

Working closely with CBDT Board Members and the Chairman on policy implementation, I've seen how much brilliant thinking goes into tax reform. The new Act itself is remarkable - reducing 800+ sections to around 536, written in genuinely plain language, designed for accessibility.

Cowork suggests a different approach: build the tools as fast as you rebuild the rules. Use AI to create AI-powered assistance that evolves alongside the policy. Don't wait for the perfect centralized solution; empower teams to build, test, learn, iterate.

The Privacy Architecture We're Not Talking About

There's a critical detail buried in Anthropic's announcement that everyone in government should internalize: Cowork operates within user-specified folders. It doesn't roam freely across your system. You define the boundaries; it works within them.

This matters enormously for government AI deployment. The resistance to AI in tax administration often centers on privacy concerns, and rightfully so. Citizens worry about algorithmic surveillance, about AIs that know too much, about data being used in ways they didn't consent to.

But the Cowork model offers a different architecture: bounded AI assistance. The taxpayer uploads their documents to a specific, secure environment. The AI works within that environment to help them complete their obligations accurately. The AI doesn't have access to their entire financial life - only what they explicitly provide for the specific purpose of tax compliance.

This isn't just technically feasible; it's a fundamentally different social contract. Instead of "trust the government with AI access to all your data," it's "use this AI tool, bounded by your choices, to interact with government more effectively."

That shift from institutional efficiency to citizen empowerment I mentioned earlier? It requires this kind of privacy-preserving architecture. You can't empower citizens if they fundamentally don't trust the tools you're asking them to use.

What Gives Me Hope

I'll be honest about something that troubles me. The gap between what's technically possible and what government actually deploys is widening dangerously fast. Private sector companies are building sophisticated AI agents in weeks. Government procurement cycles measure timelines in years.

If a taxpayer can use a commercial AI tool to manage their finances more easily than they can use government-provided tax compliance tools, we've failed. And right now, in 2026, that's increasingly the reality.

But here's what gives me hope: The India we're building through initiatives like the new Income Tax Act isn't trying to compete with the private sector on technological sophistication. We're trying to create the legal and policy frameworks that make sophisticated technology serve public purpose.

The Act's move to plain language, the Taxpayers' Charter revision we're working on, the focus on collaboration over confrontation - these create the conditions where tools like Cowork-for-tax-compliance could actually flourish.

I've been struck by how hungry students and young professionals are for this vision. They don't want to choose between public service and technological sophistication. They want to bring AI's potential into government, to build tools that genuinely help people navigate complexity.

That energy matters. If we can channel it - if we can create environments where talented people can build meaningful solutions quickly, learn from real users, iterate rapidly - the ten-day timeline that seems remarkable today might just become normal.

The Question That Matters

So here's what I keep coming back to: Are we building AI for government, or are we building AI for citizens that happens to interact with government?

Cowork is definitely the latter. It doesn't exist to make Anthropic's operations more efficient (though it probably does). It exists to make Anthropic's users' lives easier. The benefit to Anthropic is indirect - happier, more productive users who see more value in their subscription.

Most government AI is the former. Built to make processing faster, compliance checking more automated, administration more efficient. The benefit to citizens is supposed to be indirect - cheaper government, faster processing, fewer errors.

I think we have this backwards.

What if we started with: How do we help this specific taxpayer understand what they owe and why? How do we make it genuinely simple for this small business to comply accurately? How do we reduce the cognitive load on this individual trying to claim legitimate deductions?

And then worked backwards to: What AI capabilities would we need to build to achieve that? What data infrastructure? What privacy protections? What training for our officers who work alongside these tools?

That's a fundamentally different procurement process. A different innovation culture. A different success metric. Not "how many returns processed by CPC” but "how many taxpayers report feeling confident they complied correctly?"

An Honest Admission

I don't have all the answers here. Working at DOMS, engaging with policy implementation at the highest levels, I'm acutely aware of constraints I couldn't have imagined before serving in this role. Government AI deployment isn't slow because bureaucrats are lazy or unimaginative. It's slow because the consequences of getting it wrong affect millions of lives, because privacy architecture for government AI is genuinely harder than for consumer applications, because equity considerations require us to ensure AI benefits don't accrue only to the tech-savvy.

These aren't excuses; they're real challenges that deserve serious thought.

But watching Anthropic build an AI agent using an AI agent in ten days, seeing what's now possible for ordinary users, I also know this: The complexity argument only holds if we're trying to build centralized, one-size-fits-all solutions. If we're trying to create bounded, privacy-preserving tools that help individuals navigate their specific situations, the path forward is clearer than we often admit.

The Income Tax Act 2025 gives us a once-in-a-generation opportunity to rethink not just the legal framework but the entire compliance experience. We're writing Guidance Notes to help people understand the new Act. What if those Guidance Notes were interactive? What if they adapted to your specific situation?

Building Differently

If I had to distill what Cowork represents into one sentence, it would be this: AI building AI to help humans focus on what actually matters to them.

For Anthropic's users, that's managing files and tasks efficiently so they can do their real work.

For taxpayers, it could be navigating tax obligations confidently so they can focus on their businesses, their families, their lives.

For us in tax administration, it should be about creating the conditions where that kind of empowerment becomes normal, expected, achievable - not exceptional.

The technology is here. The legal framework is evolving.

And I think the next ten days, ten weeks, ten months of government AI deployment will tell us whether we meant it.

 

When Coffee Machines Know Calendars

There's something quietly revolutionary happening in the space between our kitchen counters and our calendars. Last week, Amazon launched Alexa.com - bringing their AI assistant to web browsers - and while the tech press buzzed about "cross-device integration" and "unified interfaces," I found myself thinking about something else entirely: What happens when AI stops being a novelty and starts being a utility?

I've been spending considerable time at DOMS thinking about AI in taxation - pattern recognition in transfer pricing, compliance automation, the architecture of systems that could shift us from adversarial enforcement to collaborative partnership. But this Amazon announcement landed differently for me. Not because it's revolutionary technology (it isn't), but because it reveals something about where we are in the maturation of AI as infrastructure.

And that matters for anyone working in public service, policy, or governance.

The Invisible Infrastructure Question

Here's what caught my attention: Amazon's new platform doesn't just connect your voice commands across devices. It coordinates your Bosch coffee machine with your calendar, manages grocery shopping across Amazon Fresh and Whole Foods, and adjusts your smart home settings - all without you actively managing the connections. The company claims busy professionals are reclaiming 3-5 hours per week previously lost to managing household operations.

Three to five hours. Per week.

Now, set aside whether you trust Amazon with that level of household visibility (a legitimate concern). The underlying pattern is what interests me: AI becoming genuinely useful when it reduces coordination costs rather than just automating individual tasks.

This isn't about making one thing faster. It's about eliminating the cognitive overhead of connecting multiple things.

And that's precisely the conversation we're not having enough in public finance and governance.

What Taxpayers Actually Need (Hint: It's Not Just Faster Processing)

Working on the Taxpayers' Charter revision and the implementation guidance for India's new Income Tax Act 2025, I keep returning to a fundamental question: What if we're optimizing for the wrong thing?

Most AI applications in taxation focus on institutional efficiency - faster processing, better fraud detection, automated compliance checks. All important. All necessary. But they're solving the government's problem, not necessarily the taxpayer's problem.

The taxpayer's problem isn't usually that their return takes three weeks instead of two to process. Their problem is understanding which of seven different deduction categories applies to their situation, remembering whether they need Form 10E or Form 10BA, coordinating information across multiple financial institutions, and doing all of this while holding down a job and managing a household.

The taxpayer's problem, in other words, is coordination cost.

What would it look like if we designed AI systems in taxation the way Amazon designed this household coordination platform? Not to make our processes faster, but to eliminate the cognitive overhead citizens face in navigating our systems?

The Power Dynamic Embedded in Design

Consider the difference between these two approaches:

Approach A: AI-powered system that automatically flags discrepancies in your return and sends you a notice demanding clarification within 30 days.

Approach B: AI-powered system that, while you're preparing your return, proactively identifies potential issues, explains why they might be flagged, suggests documentation you should gather, and walks you through the reasoning - before you even file.

Both use the same underlying technology. Both might even result in the same compliance outcome. But only one treats the taxpayer as a partner in the process rather than a subject of it.

This is what I mean when I talk about shifting from adversarial models to collaborative ones. It's not about being "nice" to taxpayers. It's about recognizing that better compliance outcomes emerge when citizens understand and trust the system - and when the system genuinely serves their need to comply, not just the government's need to enforce.

When Privacy Architecture Becomes Democratic Architecture

The Amazon launch also surfaced something uncomfortable: the company stores conversation history and personalization settings across all your devices. Convenient? Absolutely. Concerning? Also absolutely.

But here's what struck me: In consumer AI, we've largely accepted this trade-off. We surrender privacy for convenience, and we do it consciously (if not always thoughtfully). We know Google reads our email to make search better. We know Amazon tracks our purchases to refine recommendations. We've normalized surveillance capitalism as the price of utility.

In public service AI, we cannot make that trade.

The privacy architecture of government AI systems isn't just a technical consideration - it's a democratic one. When the Income Tax Department deploys AI for transfer pricing analysis or compliance monitoring, the question isn't just "Does this protect data?" but "Does this preserve the proper relationship between citizen and state?"

This is why I'm increasingly convinced that equity considerations and privacy protections in AI aren't add-ons to be addressed after we build the systems. They're foundational design constraints that should shape what we build in the first place.

And honestly? I think we're still figuring this out. The new Income Tax Act 2025 reduces complexity from 800+ sections to around 500 and emphasizes plain language compliance. That's movement in the right direction. But compliance simplification and AI-enabled assistance need to evolve together, not sequentially.

The Coordination Challenge That Actually Matters

Here's where this connects to the broader work we're doing at DOMS with the CBDT leadership: The real coordination challenge in taxation isn't technical. It's institutional.

We have multiple departments, multiple systems, multiple data sources, multiple compliance touchpoints. From a taxpayer's perspective, this creates exactly the kind of coordination overhead that Amazon's new platform claims to solve for household management - except with much higher stakes and far less user-friendly interfaces.

The question isn't whether AI can help with this coordination. It obviously can. The question is whether we're willing to redesign our institutional architecture to let it.

Because here's the uncomfortable truth: Truly effective AI in public service requires dismantling some of the silos and turf protections that currently define how government works. It requires data sharing across departments. It requires common standards and interoperable systems. It requires trusting that better citizen outcomes serve everyone's institutional interests.

That's not a technology problem. That's a governance problem.

And it's one that decades of e-governance initiatives have repeatedly confronted - often unsuccessfully - because we've treated it as a technology problem.

What I'm Carrying Forward

As I work on implementation guidance for the new Act and continue the Charter revision, I find myself returning to a simple test: Does this make it easier for a taxpayer to understand what they need to do and why?

Not "Does this make our process more efficient?"

Not "Does this reduce our processing time?"

But: Does this reduce the coordination cost for the citizen trying to comply?

If we're honest, most of our current systems - even the digitized ones - fail that test. We've automated complexity, not eliminated it. We've made our processes faster without making them more navigable.

The Amazon announcement is a reminder that the technology exists to do better. What we need now is the institutional courage to design differently.

A Final Thought

I don't know if Amazon's vision of AI-coordinated household management will actually deliver on its promise of reclaiming hours per week. The history of productivity technology is littered with overpromises.

But I do know this: The future of AI in public service won't be determined by what the technology can do. It'll be determined by who we design it to serve - and whether we have the imagination to prioritize citizen empowerment over institutional efficiency.

That's the conversation I want to be having. Not just within DOMS or CBDT, but across government, across policy communities, across anyone grappling with how we make public institutions work for the people they're meant to serve.

What AI applications in governance have actually made your life easier as a citizen - not just as a policy professional or administrator, but as someone navigating systems from the outside? I'm genuinely curious what's working out there.

Because the best ideas for collaborative systems rarely come from inside the institutions alone.

 

When AI Needs Nuclear Power

There's something deeply paradoxical about our digital age. We've built technologies that can recognize faces across billions of images, translate languages in real-time, and generate human-like text - all running on invisible infrastructure we rarely think about. Until, that is, the power goes out.

Last week, Meta announced deals to secure up to 6.6 gigawatts of nuclear power by 2035. To put that in perspective, that's roughly equivalent to the entire electricity generation capacity of a country like Austria. This isn't just about keeping the lights on at Facebook. It's about ensuring that the next generation of AI models - the ones that might transform everything from drug discovery to climate modeling - can train without interruption.

And it got me thinking about a question we're wrestling with at DOMS as we explore AI applications in tax administration: What happens when the infrastructure requirements for transformative technology become so massive that they fundamentally reshape industries we thought were settled?

The Hidden Cost of Intelligence

Working on AI implementation in taxation, I've become acutely aware of something most people outside the tech world don't fully appreciate: artificial intelligence is hungry. Not metaphorically hungry for data - though that's true too - but literally hungry for electricity.

Training a single large language model can consume as much energy as hundreds of homes use in a year. When you're Meta, running the Prometheus AI supercluster, you're not talking about hundreds of homes. You're talking about powering a small city, continuously, for months at a time.

Here's what makes this particularly challenging: these training runs can't be interrupted. Imagine you're ninety days into a hundred-day training cycle for a frontier AI model - an investment potentially worth hundreds of millions of dollars in compute time and researcher expertise. A power fluctuation doesn't just pause your work. It can corrupt the entire run, forcing you to start over.

This is why Meta's nuclear bet matters. It's not about being green (though that's a welcome benefit). It's about reliability. Nuclear power plants run at 90%+ capacity factors, compared to 35-40% for solar and 25-35% for wind. When you're making a hundred-million-dollar bet on uninterrupted computation, that difference between 90% and 40% isn't academic - it's existential.

What This Means Beyond Silicon Valley

Now, you might be wondering: what does Meta's energy strategy have to do with public service or tax administration?

More than you'd think.

At DOMS, as we explore AI applications - from pattern recognition in transfer pricing to compliance automation - we're confronting a scaled-down version of the same question: What infrastructure do transformative applications actually require?

It's not just about computing power (though that matters). It's about the entire ecosystem that makes sustained innovation possible. Reliable data pipelines. Uninterrupted processing capacity. The ability to run complex analyses without worrying about system failures mid-stream.

But here's where it gets interesting for public service: while Meta can sign multi-gigawatt nuclear deals, government agencies need to think more creatively. We can't just throw money at the problem. We need to be smarter about architecture, partnerships, and what we're actually trying to achieve.

This brings me back to something I emphasize when speaking to students about AI in finance: The constraint isn't the limitation - it's the clarifying force. Meta's constraint is power availability. Ours in government might be budget or legacy systems. Both constraints force us to think more carefully about what problems we're actually solving and whether AI is genuinely the right tool.

The Deeper Question About Sustainability

There's an elephant in the room that Meta's announcement highlights: if artificial intelligence is going to transform healthcare, accelerate scientific discovery, and help solve climate change, it's going to need a lot of power. The estimates vary, but data centers could consume 3-4% of global electricity by 2030, up from about 1% today.

This raises an uncomfortable question: Are we willing to make the infrastructure investments required for the future we say we want?

I find myself thinking about this in the context of India's development trajectory. We're simultaneously trying to:

• Expand electricity access to everyone
• Reduce carbon emissions
• Build digital infrastructure
• Deploy AI for public benefit

These aren't contradictory goals, but they're certainly in tension. Meta's solution - nuclear power - might work for a company with virtually unlimited capital. But what's the pathway for developing economies? For government agencies? For small startups with transformative ideas but limited resources?

What I'm Taking Away

Meta's nuclear deals won't be the last of their kind. We're going to see more announcements like this - tech giants securing dedicated energy sources, building their own infrastructure, effectively becoming their own utilities.

But what's calling to me isn't just the scale of these investments. It's the reminder that transformative technology requires transformed infrastructure. You can't bolt revolutionary capabilities onto legacy systems and expect them to just work.

This applies whether you're training AI models or modernizing tax administration. The question isn't "Can we use AI?" but "Have we built the foundation that makes AI sustainable, reliable, and equitable?"

As we work on the new Income Tax Act 2025 and revise the Taxpayers' Charter, I keep coming back to this: the most important innovations aren't always the flashiest ones. Sometimes they're the unglamorous infrastructure decisions - the data architecture, the processing pipelines, the reliability standards - that make everything else possible.

Meta is betting billions on nuclear power because they understand something fundamental: the future they're trying to build requires infrastructure decisions made today.

In public service, we're making similar bets, even if they're denominated in different currencies. The question is whether we're being as intentional about our infrastructure choices as Meta is being about theirs.

What infrastructure investments - technical, institutional, or human - do you think are missing from conversations about AI in government? I'm genuinely curious, especially as we navigate these questions in real-time at DOMS.

 

ChatGPT Health - When Health Meets Intelligence

There's something uniquely humbling about realizing that the frontier of transformation isn't always where you're looking.

For the past few months at DOMS, I've been immersed in how artificial intelligence might reshape taxation - from pattern recognition in transfer pricing to automating compliance checks. I've spoken to students at DTU and LBSIM about AI in finance, always circling back to efficiency, accuracy, and scale. But when OpenAI announced ChatGPT Health last week, I found myself thinking less about algorithms and more about my father.

He passed away in 2011. In those final months, I watched him navigate a maze of medical appointments, test results scattered across multiple hospitals, medication schedules that changed with bewildering frequency. My mother would carry a worn folder stuffed with reports, trying to piece together narratives for each new specialist. "What did the cardiologist say about the kidney function tests?" Simple questions that demanded complex archaeology through fragmented records.

What struck me about the ChatGPT Health launch wasn't the technology itself - we've known AI could process medical data for years. It was the fundamental reorientation of the question it answers.

The Question We're Actually Asking

Most health technology asks: "How do we make healthcare systems more efficient?"

ChatGPT Health asks something more intimate: "How do we help people understand their own bodies?"

The distinction matters immensely. With 230 million health questions being asked on ChatGPT weekly, OpenAI identified something profound: people aren't just looking for medical expertise - they're looking for translation, synthesis, and partnership in making sense of their health journey.

This is where my work in tax policy and health technology unexpectedly converge. Both deal with systems that have grown so complex that the gap between expert and citizen has become a chasm. Both struggle with fragmentation - multiple touchpoints, different data formats, institutional silos. And both are being transformed not primarily by making the system work better, but by empowering individuals to navigate the system more effectively.

Consider the practical reality. A person managing diabetes doesn't just need their glucose levels measured - they need those levels contextualized against their medication timing, exercise patterns, sleep quality, and stress levels captured across different apps and devices. The old approach: spend 15-30 minutes daily manually logging information, then several hours before each appointment trying to compile three months of scattered notes. The new possibility: upload your data, receive synthesized insights, generate a comprehensive health summary in two minutes. Arrive at your appointment with specific, data-informed questions.

That's not just efficiency - that's a fundamental shift in agency.

What Makes This Different

Working this closely with policy implementation at CBDT, I've learned to distinguish between technological novelty and genuine transformation. ChatGPT Health demonstrates several design choices that signal the latter.

First, the separation of health data from the general ChatGPT environment. All health conversations exist in a protected space, encrypted by default, with 30-day deletion options. This data won't train their foundation models. In an era where data privacy concerns often derail promising innovations, OpenAI chose to build walls between their business model and your medical information.

That matters because trust is the currency of health technology.

Second, the explicit framing: "designed to support, not replace, medical care." The technology positions itself as infrastructure, not authority. This reminds me of how we've been thinking about AI in tax administration - the goal isn't to replace tax professionals with algorithms, but to free them for complex judgment calls while AI handles what's systematic.

The Questions That Surface

But every powerful tool creates new responsibilities. Several questions keep surfacing for me:

The equity dimension: Connecting medical records and wellness apps assumes you have both. In India, where healthcare records are increasingly digital but far from universal, where Apple Health penetration remains limited to a small urban demographic, who benefits from this technology? How do we prevent health AI from becoming another layer of advantage for the already advantaged?

The interpretation gap: AI can identify patterns in your glucose levels, but can it distinguish between correlation and causation in complex biological systems? The AI might flag the pattern, but who owns the interpretation?

The data dependency: What happens when people begin outsourcing their health literacy to AI? There's profound value in learning to read your own body's signals. Does AI synthesis enhance that literacy or erode it? (or may increase anxiety!!)

These aren't questions with clear answers. They're tensions to be navigated.

A Personal Thought

I keep thinking about those folders my mother carried, stuffed with medical reports. In some future I can now imagine, those reports would flow seamlessly into a secure space where patterns emerge, where questions form themselves, where the doctor's appointment becomes a genuine conversation.

My father won't benefit from that future. But millions of others might. And if we get this right - if we build these tools with intention, with equity, with clear boundaries - they'll benefit not just from more efficient healthcare, but from deeper understanding of their own wellbeing.

That's the possibility that keeps me engaged with AI. Not because technology solves everything, but because thoughtfully deployed, it can shift the balance of agency back toward individuals navigating complex systems.