All posts

Product & Company Updates 5 min read

Clustering is the Key to AI for Tax

Neo.Tax
Clustering is the Key to AI for Tax
On this page
4
  1. Clustering is Key
  2. Why Clustering is the Neo.Tax Solve
  3. An Update that Lets You Have More Control
  4. Automation + Customization

By employing best-on-the-market LLM technology, Neo.Tax’s approach is much more rigorous and thorough than the traditional ASC 350-40 and R&D credit filing process. Rather than relying on dozens of engineer interviews, our automated system combs through a company’s existing project management and payroll systems to determine what projects our Qualified Expenses and how many hours each engineer spent on these tasks. 

These datasets are massive, which is why traditional tax teams and third-party firms rarely lean upon them, even though they contain the most detailed (and most IRS-preferred) record of R&D work at a company. Which is why AI is the solution for ASC 350-40 and R&D credit filing.

Clustering is Key

A typical software Fortune 500 company will generate about 2-5 million tickets in a given fiscal year. In order to understand that work well, you have to understand both what each ticket is about, and how it relates to each other piece of work. That translates to roughly 80 million tokens worth of context.

Somewhere in the 80M tokens is the context that tax teams need in order to identify R&D projects to submit to the IRS. The question is, who is going to parse and understand that context for you?

It’s virtually impossible for a human to absorb that much information. They’re going to focus on high level aggregates. As a result, they may miss out on many pockets of work that deserve to be highlighted that would otherwise get lost.

An off-the-shelf AI model is able to process large volumes of data, but it is well known that model performance degrades severely in long contexts, especially in settings where most tokens are irrelevant. For tax team looking to boil down 4M tickets into about 75 R&D projects, fewer than 0.01% of tokens are usually relevant to identifying and describing those projects. This puts the underlying transformer architecture that drives these models under severe strain.

The solution to the problem is to get smart about identifying just the right set of tokens that matter for the problem. But tax and engineering teams do not usually have the time to sort through and identify where the relevant context is together, and in large organizations, there is usually not just one place where this information will be found. As a company built on Engineering and Tax working together, Neo.Tax has figured out a generalizable solution that is able to identify when a token matters, and brings the best of Engineering (understanding the work context) and Tax (understanding of the IRS’s requirements) together in one place.

Why Clustering is the Neo.Tax Solve

Here’s how our AI-powered system works:

  1. Our LLM reads through every single ticket and task within your company’s project management software.
  2. It evaluates each project’s eligibility step by step in the 4-part test using a tailored model, trained on IRS codes, regulations, court cases, memorandums, letters, and guidance.
  3. Finally, it creates a detailed narrative of each project, explaining why or why not it qualifies for the R&D credit.
  4. You can then inspect each step, and determine whether you agree or disagree with Neo.Tax’s determination.
  5. By delivering the detailed reasoning for each determination, Neo.Tax creates an audit-ready study automatically.

An important step here (and one often overlooked in the tax automation process) is “Clustering” — the step in which our AI system finds patterns within tickets and automatically groups them into Projects. 

As our LLM reads through each ticket, before determining the eligibility, it moves that ticket into a specific project. In a perfect world, where every engineer and executive tagged and named projects identically, this would be a pretty straightforward step. But in reality, teams often have different naming conventions and, of course, some activities are mis-tagged, misspelled, or not tagged at all. 

That is why we trained our LLM to be able to pattern-match work to specific projects, based on the people involved, certain terms or figures within the work, or other aspects. Then, that work is grouped in a way that the Neo.Tax user at the company can review each grouping to check the work and understand the rationale behind the grouping.

An Update that Lets You Have More Control

What’s changed in our recent update, which just went live last week, is that we are allowing customers to be more involved in the Clustering process. Our previous approach implemented a particular point of view on how to group tickets together. Because it was trained to work across a wide range of customers, it didn’t allow for as specific an approach as some companies needed. 

Since then we’ve realized that some companies may have existing groupings that make more sense for their internal processes. Now they can “bring their own” and we can use that grouping information to better match their tickets. 

For example, the project list usually is developed by a separate team than the engineers. That means that they often use different language to describe the same work. The project list will have a public facing name like “Improve XYZ” but internally the engineers might call it “Operation Snowball”. This means we need to learn how to build context within a company’s data over time, so that we can learn those associations. We’ve developed a system that adaptively learns from user feedback, so that the matches become more accurate with increased usage.

This is especially valuable in ASC 350-40, which is largely an internal process and varies from company to company. Now, companies can bring their own sensibility about how granular they want their projects to be and how aggressively they want to capitalize and we can accommodate quite flexibly by updating the algorithm on a case-by-case basis. 

But there is value in R&D as well. Often, a customer will “bring their own projects” to a credit — perhaps utilizing a credit put together a year earlier by an internal or external tax team — only for our LLM to unearth new projects that the interview process may have missed. Because Neo.Tax reads every ticket, we can often find the corners of work that were overlooked.

And don’t worry: if you prefer to not bring any projects to the algorithm, our automated system can still ingest raw Jira data and create an accurate and comprehensive filing. 

Automation + Customization

The big idea here is that automation can be a tool in a company’s toolbox. Certainly, if the priority is effort-reduction and speed, we have a system that can file faster and more accurately than the one you’re using. But we’ve seen that companies want more than that — they want the efficiency and peace of mind of a filing built off of their data, but also the flexibility to correct and adjust the output. Strategic finance is an important part of being an innovative company, and Neo.Tax can be a tool for the most strategic companies. 

To be strategic with R&D and ASC 350-40, you need to be able to ingest and utilize your contemporaneous data. Neo.Tax allows you to do just that. Now, our system also allows you to adapt the algorithm to better fit your needs.

AI for Enterprise R&D. See how it works for your teams.