How Does Neo.Tax’s AI Work?

By now, everybody and their mother has heard of AI. But that doesn’t mean most people understand it. Apparently, it can drive cars, write blog posts, create fine art, and do most other tasks if you believe the hype. But how does it work? And can you really trust it?

When it comes to taxes, trust is essential. To build that trust, we believe in being transparent about how our AI works. Which is why we’re pulling back the curtain on how our AI works at Neo.Tax.

So, here’s what you need to know.

The Three Issues of R&D Taxes

There are three hard problems to solve when it comes to filing an R&D credit. 

1. Which of your projects count as “qualified” for the credit under the 4-Part Test?
2. How much time did each employee spend performing qualified activities on those R&D projects?
3. Can you create R&D project narratives justifying the credit to the IRS? 

When we set out to automate the R&D filing, we knew our AI solution would need to successfully master those three problems. Here’s how we did it.

Identify Qualified Projects

Project management systems like Jira, Linear, or Azure DevOps contain most of the information needed to identify R&D work. But no human has the time to sort through every ticket in those systems — large teams have hundreds of thousands to millions of tickets per year. Even if they did, the data is messy, disorganized, with different engineers and teams tracking their work in different ways. That’s why Neo.Tax’s AI has been trained on messy data, which allows it to identify where certain tickets should live based on context clues like textual and structural signals. 

The first task for Neo.Tax was creating a tool that could search through a company’s existing project management system data to identify qualified R&D projects. Neo.Tax’s AI begins by scanning through the tens to hundreds of thousands of tickets and grouping them into separate projects based on both content and hierarchy information with each ticket.

At this point, the AI has instantaneously sorted a year’s worth of development work into a dozen (or in some cases, many dozens of) buckets, each labeled as a separate project. 

Now, AI can do what it does best: digest a mountain of data and sort it based on a defined ruleset, the IRS’s 4-Part Test. Is the work Technical in Nature? Does it have a Qualified Purpose? Does it solve for Technical Uncertainties? Does it include a Process of Experimentation?

Neo.Tax’s AI system determines if each of the Project groupings is Qualified, Partially Qualified, or Unqualified for an R&D credit. Just like that, twelve months’ work has been grouped, tested, and labeled.

Create R&D Narratives

From there, based on the project management data, Neo.Tax’s LLM creates a rigorous narrative explaining every project. This is where the LLM technology (think: ChatGPT) becomes such a gamechanger. 

The IRS prefers contemporaneous data—something Neo.Tax leverages in a way no human accountant could. Better still, it creates narratives to explain exactly how and why each Project qualifies under IRS rules. 

This step has two advantages for filers. First, the company is audit ready without the pain of doing interviews and manual data gathering. This saves the company tens to hundreds of hours and gives them confidence knowing that they have an unprecedentedly detailed study that they can immediately present in the event of an audit.

Second, by taking a holistic approach and looking at all projects, Neo.Tax can help identify qualified work that traditional methods may have missed. A company to dig into the automated filing and understand how the AI is getting to its outputs. By showing its work, a company’s Head of Tax or CFO can catch places where Qualified work is not being credited in the way it should. 

Calculates R&D Percentages for Employees

Once the Projects are sorted and put through the 4-Part Test, the final step is calculating the percentage of each employee’s time that can be counted as R&D work.

Because Neo.Tax’s AI has already ID’d the tickets that relate to Qualified R&D Projects, tying the employee to the project is relatively simple (at least, simple if you are AI). 

The payroll, vendor, and project management data required to calculate the credit in Neo.Tax is readily available at many companies. That means that rather than relying on Quarterly or Yearly estimates filled out by engineers, you can have exact percentages per employee without any additional work for your engineers.

But Neo.Tax also understands that our solution only works if it’s built for the messy reality of how people actually track their work. So, even if your data is inconsistent or incomplete, it’s not a problem for our AI. Rather than relying on tagging, Neo.Tax’s AI has been trained to scour content and find patterns to sort. Nonsensical entries that would befuddle any human attempt to use keyword searches to sort these massive datasets are non-issues for our AI. We trained our AI on messy/incomplete datasets, because we know that engineers are not experts at tagging and labeling their work.

The last hurdle for Neo.Tax’s AI is one you’d expect: missing data. But even that challenge is not insurmountable. We know, especially for senior employees, not every activity is tracked in your system. So, the AI identifies unrecorded time and flags it for the filer.

Not a Black Box; A Better Way

The fact is, the old way of filing R&D credits is an inexact (and expensive) science. Most filers rely on interviews with engineers months after work was completed, relying on fuzzy memory and guestimates. 

Now that the data exists within your system, the IRS has been signaling that the burden of proof will be rising for R&D filings. But, as you can see, sorting and digesting the existing data is a yeoman’s task for any human (or even team of humans). 

This is a problem that AI was created to solve. And luckily, Neo.Tax has built the AI to solve it.