The AI Podcast: Using AI to Make Tax Day Easier

Engineering, Tech Leadership, Technology Studio microphone for recording podcasts over blue background

The AI Podcast connects with some of the world’s leading experts in artificial intelligence, deep learning, and machine learning to explain how it all works, how it’s evolving, and how it intersects with every facet of human endeavor, from art to science. Recently, host Noah Kravitz sat down with Intuit Senior Vice President and Chief Data Officer Ashok Srivastava in this AI Podcast episode to discuss how the company is using AI to make tax day easier, and to support small businesses and self-employed individuals during the COVID-19 global economic crisis. Following are highlights from their conversation. Click here to listen to the full podcast.

The AI Podcast: A lot has happened since we last spoke, to say the least. How has your team at Intuit been using AI to help the small business community with the challenges they’ve been facing?

Ashok Srivastava (AS): The COVID-19 pandemic has had an incredible impact on small businesses worldwide—it’s really alarming. With our Intuit Aid Assist program, we’ve used artificial intelligence and knowledge engineering to make it simple for small business owners in the United States to determine whether they’re eligible for government relief, how much of a loan they can get, and what the terms would be. This is an exciting program because it helps us help our customers get the money they need to make it through these difficult and unprecedented times.

We’ve defined artificial intelligence to have three major components: machine learning, natural language processing, and knowledge engineering. Knowledge engineering is the part of artificial intelligence which takes rules that people have written, such as government regulations, and converts them into code automatically so that computers can take action on them. This becomes the brain behind a lot of the questions that are asked in Intuit Aid Assist, TurboTax, and many other places within Intuit’s product lines.

This is especially powerful given the size and complexity of these rules and regulations. Consider the U.S. Tax Code itself, which is 80,000 pages long. No matter how smart a person is, it’s very difficult to say, “Oh yes, I’ve read the U.S. Tax Code and I understand it.” Now we’ve created systems that do that. It’s a perfect task for artificial intelligence.

The AI Podcast: What are some of the other ways you’re using AI and deep learning to help small and medium sized businesses?

AS: I’ll give you a couple of examples. One that’s very relevant to COVID-19 is cash flow forecasting. Small businesses need to make sure they have enough money in the bank to meet all of their obligations to their employees and their customers. This depends on a number of things: What are the outlays that are coming up? What are the major invoices they’ll need to pay? What are the major invoices that their customers will need to pay so that they have money coming in? To help them manage their cash flow more accurately, we’ve built deep learning algorithms and other kinds of algorithms that help us forecast, for example, how much money they’ll have in the bank two weeks from now. This is one of the most important applications that we’re focusing on for small businesses.

We’re using learning in very novel ways to address particularly challenging nuances in this area as well. It’s important for a small business to know how much money they’re going to have in the future, but it’s very hard to say something as precise and definite as, “You’re going to have $30,000 in the bank next Tuesday.” Ideally, you’d like to be able to say, “We think you’re going to have $30,000 in the bank next Tuesday, plus or minus 10 percent.”

Figuring out that confidence interval turns out to be a hard problem that statisticians have worked on for a long time. We’re building deep learning neural networks that can allow us to identify and predict not only the $30,000, but the confidence intervals as well. Because 10 percent might be okay, but if I say, “Next Tuesday you’ll have $30,000 plus or minus 100 percent,” that’s not very good. You can’t take action on it. We want to be able to give people that level of information. Doing so takes not just deep learning, but also areas like quantile regression and new algorithms for bringing together nonlinear predictions. It’s a very complicated and very exciting area of machine learning and deep learning that we’re exploring.

[For a deep dive into how AI can help build resiliency in a global economic crisis, see this Intuit Blog.]

The AI Podcast: We’ve talked in the past about chatbots, natural language processing, and voice recognition. What have you been working on lately to make it easier for customers to interact with your products?

AS: One of the most important things that we need to do in that area is to build systems that can understand what people are saying in common conversation, because when we have those kinds of capabilities, we can serve our customers better. For instance, when a person is using TurboTax or QuickBooks, and they have a question, they often don’t know what to search for or where to look for that information. When they go into the application and start looking around, the AI systems in the background are taking that information and making predictions about what their real question might be.

When the person asks a question in our help box, we can look at the question plus all of the feedback they’ve been providing us implicitly, find a match in the background, and say, “We think these are the most important articles for you to look at,” or, “We think this is where you need to get help.”

The AI Podcast: How do you see the work that you’re doing progressing over the next five years? For example, what’s ahead for computer vision?

AS: Think about the sheer number of documents small businesses and consumers have to deal with to file their taxes—receipts, W-2 forms, 1099s, and so forth. Now imagine the ability to just snap a picture of that pile or upload it in some other way, have the machine parse all of that information, extract what’s needed, put it into a database, and then your taxes are almost done. That’s the world that we’re trying to drive to with deep learning and a lot of other technologies.

Of course, this is easier said than done. There are all kinds of W-2s out there, some handwritten, some in different forms, not to mention the different ways a receipt can look. Making a general purpose computer vision system that can look at all of that and get the right information with extremely high reliability is a great challenge. We have to be very, very accurate—even a small error can have a costly impact. Building this system requires deep learning, machine learning, as well as knowledge engineering in the background. If you think about it, all of the data fields on a W-2 form have relationships to each other, so our knowledge engineering system can make determinations about whether or not we’re reading the form correctly.

The AI Podcast: Are there other roles that deep learning can play, particularly during tough times when small businesses are struggling to make rent, to meet payroll and to otherwise manage cashflow?

AS: If you think about that small business owner, the consumer, every day they’re making decisions that affect their livelihood. Those decisions need to be informed by data, and by recommendations based on the judicious use of that data. For the foreseeable future, these two things are going to come into play again and again. Data and machine learning will help people get the best information at the right time so that they can make the best decisions about their financial future.

To listen to the full podcast, go here.

Tags: