I’ve decided to build an AI agent to simplify the U.S. tax code. I’ll dig into the technical details in a later post, but for now, I want to share some staggering stats that helped convince me this work is necessary.

These numbers come from a deep analysis of the source documents: the full U.S. tax code plus every IRS bulletin issued since 1995.
TL;DR: it’s a mess. A really big, expensive mess.

📚 The Core Problem

Let’s start with the basics. The U.S. tax code, as written into law, spans 7,061 pages. That alone would be intimidating for anyone trying to understand how taxes actually work.

But that’s just the beginning.

Every year, the IRS publishes guidance bulletins — documents that interpret the tax code, clarify its meaning, and offer legal determinations. These aren’t optional reading. If you want to truly understand the tax code, you have to understand the guidance too.

Between 1995 and 2024, these bulletins have averaged 2,518 pages per year. When you add it all up, that’s over 80,000 pages of guidance layered on top of the already sprawling tax code.

That’s 1.8 GB of raw PDF files — and growing. If current trends hold, we’ll hit 140,000 pages by 2050.

Projected growth of U.S. tax documentation

🧠 The Human Cost

No human being can reasonably be expected to digest all of this. Reading 140,000 pages of dense tax material, at an average reading speed for legal text, would take more than two years of full-time effort — 8 hours a day, no weekends, no holidays.

And what a two years it would be.

💸 The Financial Cost

This complexity isn’t just abstract — it has very real downstream effects.

Each year, Americans spend over $14 billion just to file their taxes. That includes the cost of tax professionals, accountants, and software. It’s essentially a tax on the tax system itself — a hidden surcharge we pay to navigate the bureaucracy.

🤖 A Different Approach

We’ve built incredible tools in the last decade — machine learning, large language models, generative AI. But most tax software today still focuses on helping people survive the complexity, not eliminate it.

That feels backward.

So for my capstone project in Google’s Gen AI course, I’m building an AI agent that takes a different approach: simplify the tax code itself. Not just interpret it — distill it. Reduce it. Make it human-readable again.

As Oliver Wendell Holmes once said, “I wouldn’t give a fig for simplicity this side of complexity, but I would give my life for simplicity on the far side of complexity.”

More soon.