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.