Taxes are mostly thought to be about numbers (dollars and cents), but foundationally, it’s about the literature. There are over 80K pages of statutory laws passed by Congress, over 180K pages of rules from federal agencies, and at least 100K pages published by the IRS alone. When you add state and local law and code, the numbers explode into the millions. These numbers on their own are astounding. They’re further complicated by the fact that they change yearly: estimates suggest five to ten percent of federal regulations are edited per year, and as much as twenty percent of IRS tax guidance is edited annually. Executive orders can have an even higher propensity for change (and we’re having a record year).
This body of data is incredibly unwieldy. To comply with the tax code, the IRS estimates that Americans spend roughly $133 billion annually in out-of-pocket costs, including preparers and software. This is because mistakes are both stressful and expensive. Every dollar of tax error costs, on average, $33 to resolve.
This spending has created some incredible vertical giants, such as Intuit’s TurboTax, Thomson Reuter’s Westlaw, and the International Code Council. They all employ the same business model: a historical data moat built by an army of humans interpreting the relevant code, maintaining that on an ongoing basis, and expressing that through the industry-specific software experience on the front-end. Our belief is that these businesses are in for a rude wake up call: LLMs just ate their moat.
We know it can work because it’s already happened
Tax codes are not the only market with a large, continually evolving dataset. Pretty much any area where industry meets the law needs a software solution that, to this point, has required a fleet of subject matter experts to keep it current. To show us what’s coming, we can look to two verticals a few steps ahead in their AI disruption journey: case law and building codes.
Casetext was the first major outcome that showed what can happen when you blow up a data moat. Casetext was founded out of YC in 2013 with the goal of creating a free legal crowdsourcing tool (“Wikipedia for the law”). They were early to AI, using it in 2014 in their flagship product, but struggled to gain traction against the data moat that Westlaw and LexisNexis had built around historical case law. The team was brilliant, they had a distinct point of view on the market, but, nevertheless, no amount of venture capital gave them the ability to replicate the data set of the incumbents. That changed overnight with access to OpenAI. Casetext developed a custom model using GPT4 that let them more rapidly build out their data assets and bundle in new research tools and capabilities that let them leapfrog ahead of incumbents. In March 2023, they launched CoCounsel, an AI legal assistant for research, drafting, and analysis. In June 2023, Thomson Reuters bought Casetext for $650M.
We saw something similar in building codes. In 2022, ICC, a hundred-year-old organization (by merger), sued UpCodes, a startup founded in 2016 that used AI to parse building codes. Why would they do that? Because they realized that LLMs had completely collapsed decades of data moat that they had built. UpCodes was able to take all the available public building codes and via LLMs produce a slick, modern interface that vastly outmatched the physical books and reference sheets they sold. They also claimed a win for public data.
Both of these categories point to specific characteristics we know make a market vulnerable to AI disruption: an abundance of data that is complex and dynamic, in highly regulated, high stakes fields where mistakes are costly. Taxes are an obvious market on that list.
We also know that, unlike some verticals, the tax industry has been a ready adopter of new tech. Your standard, off-the-shelf tax software is supported by what’s essentially a calculator and an if-this-than-that logic system that flags common errors and surfaces relevant information. Filers get a lot of value out of that, with the error rate of e-filing at 1%, compared with paper filing at 21%. And the market says they’re ready. According to Thomson Reuters, 88% of tax professionals believe AI will be central to their work within the next five years, and 42% of tax departments are actively exploring AI solutions.
Business sales tax is suddenly sexy: Anrok, Kintsugi, Numeral, and Zamp
Sales tax law is its own special kind of headache. It’s hyper local, layering state and jurisdiction taxes. California alone has over 400 sales tax jurisdictions.
There’s an interesting duopoly in the sales tax software market between TaxJar and Avalara. Both companies unfortunately are less competitive today than they were five years ago, TaxJar having been acquired by Stripe and Avalara being taken private by a PE buyer. Both also collectively do over a billion in revenue, based predominantly on having deep expertise on very local sales tax codes. So, why are there so many new startups chasing this market? In part, because of the South Dakota v Wayfair (2018) case that allowed states to start collecting sales tax for out-of-state sellers, adding a significant layer of complexity with origin-based and destination-based tax rates. But that demand is only part of the puzzle. The more fascinating thing is the speed at which these companies have come to and are maintaining product parity with entrenched competitors. This is only possible because they can create incredibly efficient data workflows to interpret and monitor any changes to tax code which then is immediately pushed into product code.
We wouldn’t be surprised if these startups are also being pushed ahead by regulatory tailwinds. Cross-boarder trade is in flux, intensifying the demand for global solutions. Anrok raised on being global first, as did Kintsugi, which doubled its valuation in six months. Meanwhile, Numeral has entrenched itself with leading D2C brands, attracting Benchmark as its lead investor, and Zamp has diversified itself across multiple segments aiming for end to end automation.
Can someone slay the behemoth: April, Column Tax, and Keeper Tax
If there has ever been an obvious giant to slay, it’s Intuit. Intuit has a ~$5B business in TurboTax supported by almost $5M of annual lobbying spend to retain the status quo. There are fewer than ten vendors with the full data set for federal and statewide filings, with none founded after 1990. Intuit has not been competing with other startups, but with the actual federal government to keep itself relevant.
And suddenly, there is a whole wave of startup activity in replicating the underlying product of TurboTax and its peers. These new startups replace the functionality of the legacy software and then take it a step further by offering a more embedded experience, using AI to let your financial datasets “talk” to each other: finding deductions directly from your bank account, real-time forecasting, financial planning, etc., bringing more expensive financial services downmarket.
This dynamic is one we’re seeing in a few markets. There’s value in AI for B2C (individual filers) and B2B (accounts and CFAs), but the boundaries between the markets blur as AI eats the labor and shifts offerings from services to software.
This is a massive prize worth winning, and one we’re watching very closely. At the core, again, is the shift in how LLMs can suddenly power the data pipeline to interpret and understand the complexity of the IRS tax code and translate that into software code, with a supporting workflow to maintain tax code changes in a cost efficient way.
Where else?
In short, you tell us. Tax is not the only market in which there are deeply entrenched incumbents with a rapidly evaporating data moat. The next vertical giant could emerge here, or it could come for compliance automation, bill analysis, or legislative intelligence—or all five. What we know is that the pace of change is rapid, and that what was previously impossible is suddenly being reinvented by very smart teams. A lot of revenue could shift hands, which is exciting for the entrepreneur and venture communities.