Intuit builds "done-for-you" tax model based on AI

Accounting Software

chatbox of intuit assist's chatbot

What happens when you take the power of generative AI and plug it into tax software? Intuit has fully embraced this new wave of technology; even using it as the foundation for an “operating system” for its whole business

Intuit has also built generative AI into TurboTax, its consumer tax software for US and Canada, in the form of its AI assistant, Intuit Assist. Unlike CoPilot for Microsoft, Intuit Assist is included for free in TurboTax, accounting software QuickBooks, marketing platform MailChimp and credit monitoring service Credit Karma.

Last September Intuit released a video showing how Intuit Assist works in TurboTax. Standout moments include:

  • Creating a personalised checklist for documents that automatically updates itself as you upload data
  • Reading and analysing documents and extracting data to fill out the respective fields on tax forms
  • Correcting mistakes on the fly in your form if the data looks incorrect or outside usual parameters 

TurboTax also searches an online database of tax concessions to calculate the largest possible tax break, which is presented as a gamified tally that rises as it finds more rebates.

From a strategic perspective, it shows the likely direction that all tax returns – including for businesses – will eventually travel. Towards a highly automated service where the AI will collect and organise information and answer most tax questions, and humans will come in to verify information and provide a sense of comfort that the numbers are correct.

It raises important questions for accountants about how much AI will disrupt their current processes for completing tax returns.

  • How has generative AI already changed the process and how much further will it go?
  • What skills will accountants need to complete the work?
  • Will accounting firms own the tax return process or will the software company?

Digital First recently interviewed Jamie Belsky, the lead for TurboTax and TurboTax Live who oversees Intuit Assist and all other AI work in the product. 

The interview has been edited for length and clarity.

DigitalFirst: Does Intuit Assist do everything in the video that is released on the Intuit Assist site?

Belsky: It does, capability wise. We are learning and iterating with our customers in real time. Tax season in October is really the final filing deadline here in the States. So that's when we get this last influx of customers where we can learn quite a bit for the next season. 

And so we launched that experience in the product, we learned from customers, and we've learned some new things and have made small pivots. But I think the benefits that are in the video and what Intuit Assist does, that is true to the announcement. As we're learning from customers, we're tweaking the experiences to make sure we're delivering the maximum benefit.

DigitalFirst: It does feel like quite a radical improvement on how people would normally do their tax return. When the customer decides to get in touch with a human agent from within TurboTax, does the agent get a Gen AI summary of the file? Because it does mention AI-enhanced support a couple of times. 

Belsky: We've always had some sort of document understanding capability that was usually leveraging OCR, optical character recognition. What we found with some of these large language models is their document understanding was just better and so it's able to accelerate it. 

What we're looking for is document identification. This is a W-2 (form), this is a 1099 (form), as well as the data extraction. So it's able to recognise all the data in every box for every field. And then application – it puts it into the right place in the tax returns. 

And so it really just sped up our coverage of documents that we can identify and extract data for, which was awesome. We use them in tandem; where our existing system was already working really well, we left it. We don't need to recreate things. But we use it as a hybrid, as a fallback when our own system fails. We're seeing some really great results. Then it applies (the information) to the return. 

In broad strokes, there's two ways you can get your taxes done with TurboTax. A tax expert can do them all for you. You're just uploading your documents and answering some personal questions. 

Or you're going to do the questions yourself – with or without some experts jumping in to help you along the way. 

So if you're in that former camp where an expert is doing it for you, then that expert can see all the data that comes across in those documents. What we do for everyone is mask personal information. So an expert will not be able to see your social security number, for example. Before it gets to the expert, it's going to go through that scrubbing where they'll see the data they need to do their jobs and be effective, but they're not going to see anything that might put you at risk. 

DigitalFirst: Do people care more about that? You've got to trust your accountant. People tend to be more worried about sharing how much they earn than some ID number.

Belsky: “It's a “say, do” difference. All customers will say, “I want my information secure. I don't want anyone to see my social security number.” And then they will go and email a document that has their social security number on it. So you're totally right. 

That's awesome as a one-on-one relationship. But we're scaling a network of over 14,000 experts. And so customer privacy is just really important. 

We trust them (the experts). They're all credentialled, they're all screened. And we still want to be careful as that network begins to grow, that you can feel confident that you're getting the best tax expert and you don't have to worry about anything else like sharing your information. Because what's new, in a lot of cases, is that you're not meeting these people. You don't meet them face to face. They might not even live in your city or your state. 

This is a faster relationship, a faster matching. You're basically saying, “I'm going to go to TurboTax, someone's going to get picked to do my taxes for me, and we're going to start this relationship”. Versus, maybe you're going to go to Google and you'll read some reviews on Yelp or you're going to get a recommendation from a friend. The models are just different. So we just take that extra step to make sure everything is secure.

DigitalFirst: Going back to the personalised checklist, the form is either a yes or no as to whether there's information in a particular field. And if there is information in a field, then the checklist requests other forms. So you’re not using generative AI at this point to interpret what's being put into the fields. It's more just field matching?

Belsky: It does field matching. The checklist is generated through some simple questions that you might ask and the documents that you've uploaded to create lookalike models that say, you like you also have these documents, too. Once you start uploading those documents, then we pull the data out and apply it to the return into the specific fields where it needs to go and do all the calculations.

DigitalFirst: But there's no interpretation? It's taking data and putting it into the field that it needs to go into in the tax return, as opposed to, say, an accountant saying, “It's interesting that you've had these kinds of incomes, we should talk about this kind of tax treatment”.

Belsky: Right, but I think that once folks get to us, it's a lot of looking back, right? Your financial year is over and you have already made your decisions. Our full service experts are able to say, “These are the things we might want to do differently next year.” 

And that’s what we're excited about with some of these large language models (the technology behind ChatGPT) and the new models that we're building around tax. We could give that (tax advice) to everybody through Intuit Assist.

DigitalFirst: Watching that video made me think, why is this software for filing tax returns? I mean, logically, these questions should be answered at the beginning of the year, and then you can make your buying decisions or whatever else. 

Then the return part is just filling out the decisions you'd already made at the beginning of the tax year. It’s more relevant for business, than individuals. Is that something that you're also thinking about? Releasing a version of the software for tax planning?

Belsky: It absolutely is. I think that the question that we are trying to figure out is where's the right place within Intuit for those sorts of things. Because we have Credit Karma, for example, which is a year-round engagement app that pretty much everyone has an account for. (Credit Karma is a platform for US consumers that shows their credit score and helps them apply for credit cards.) 

That's where folks are. What we found is that most people don't want to think about taxes when it's not time to do their taxes. And most people don't have a lot of options when it comes to tax planning. 

Where we see people who do need it are investors and self employed. And so we have started to roll out some tools. We have one that's called the Investor Center. Right now it's just for crypto. You can connect your crypto accounts and we will monitor your gains and losses as well as tax implications of any sales for short or long term capital gains. 

The same is true for self employed. What are the expenses that you should be tracking throughout the year in order to deduct them? We have an awesome system in product (QuickBooks Self-Employed) which actually makes it quite easy. 

So when you're doing your taxes, all you do is connect your credit cards or your bank accounts, whatever you use to spend on your self-employed business. And we'll find the relevant expenses for you based on your industry so you can plan throughout the year. 

And if you don't, we've built software and AI models to go and find the specific deductions that would apply to you and apply them to your return. When they use our system, we find 30 percent more deductions than folks typically track. Mostly because there are deductions they weren't aware that they could be taking for their specific industry. 

ChatGPT training for accountants - book here

Digital First: Just looking at the metrics; I know you guys are big on NPS and other things. You probably are tracking time-to-complete for forms in TurboTax. Have there been any noticeable improvements in those times since turning on Intuit Assist?

Belsky: We do track time and we've got a feature in the product that literally shows you the time that it will take to complete your tax return. But that's not necessarily one of the leading indicators of success that we're looking for. For Intuit Assist, in particular, what we're really trying to measure is confidence. 

So, you know when someone who's used to working with an accountant or a bookkeeper starts to do the taxes themselves, they make that leap. They're encountering a lot of things that they've never experienced before. Things like actually having to do data entry or understanding why their outcome is the way that it is. 

And so when we look at things like how we explain an outcome – here's how much you owe, here are the major factors that drove that outcome – we see that Intuit Assist almost doubles the helpfulness rating. A lot of these experiences have a “thumbs up, thumbs down” (rating for) explanations, which is really exciting to see. And we have started to see them move forward in the product, even for negative outcomes like owing taxes, for example. 

We've got an accuracy check experience where in real time Intuit Assist is monitoring the return for either missing fields or anomalous entries. Things that are just way outside the norm of what we would normally see. And Intuit Assist flags it for the customer and they're able to just fix it. Even if they've moved on from that field or that screen they can just fix it in line. 

We used to save up all of the accuracy issues or errors to the end and then make the customer go back and fix them. And now the system is monitoring in real time, customers are engaging at very high rates and actually fixing the issue. That then leads them to finish the product with no issues and they're ready to file.

DigitalFirst: What other ways are you using generative AI to improve the experience?

Belsky: We use Intuit Assist for two types of questions; identifying opportunities for reducing tax (Which deductions and credits did you qualify for?) and explaining outcomes (Why do you have a refund or a balance due?).

We also use Intuit Assist for self-help questions; accuracy issues, which I just talked about; and document understanding.

We've also been using it for dynamic translations. This will be the first year that TurboTax is entirely available in Spanish as well as English. Every complexity, every service level, all of the help articles or tooltips, is fully translated. And we've been able to lean into these large language models to help translate big parts of our system that weren't as ready to be localised. 

This is a really cool middle layer where content is passed through and translated and then cached and served up to the consumer. We're really excited about what this unlocks. Spanish is the second most spoken language in the United States. But it means we can better speak all of our customers' languages. 

And then going even deeper into the stack, there are about 15,000 tax forms that are issued every year between the federal government and all 50 states in the US. And every year, our team of tax analysts have to look at last year's form and compare it to this year's form and spot the difference. That's something that generative AI does really well – to figure out the differences, explain the differences and make a recommendation on any changes that need to go into the product in terms of how we explain new qualifications, thresholds or anything like that. It will cut out a ton of time from our tax analysts that could be better spent in other parts of the business. 

So all the way from the consumer experience to deep down into our stack, we can leverage Intuit Assist and generative AI to make the tax filing. Either how to run this business more effectively, or make the tax filing experience much easier for consumers.

DigitalFirst: Intuit seems to be talking much more about working with accountants. One of its five big bets is “working with experts”. But how does this automation with generative AI work with your accountants? 

Belsky: We have an expert network of accountants and CPAs, enrolled agents, JDs, who are all working on both tax and bookkeeping. They have their own tools that we built internally that help them serve customers. We talked about the document extraction and application work that jumpstarts the entire return. So they don't have to do kind of onerous data entry, they can move on to either the more nuanced parts of preparing a return or simply reviewing and verifying that everything's correct. 

If you're just taking calls from consumers who are trying to do their own taxes and they have a few tax questions, we use Intuit Assist to ensure quality and consistency of service. This means serving up the right answer and even recommendations on how to lead the conversation. Those come through Intuit Assist.

Intuit Assist is really helpful in generating the case notes from the call as well. The expert doesn't have to spend time documenting everything that was discussed with each customer. This is something that usually takes our experts a lot of time – either in product support or tax. If someone calls back a second time, we want to make sure they feel a continuity of service, even if they're speaking with someone different. 

DigitalFirst: Similar to conversation summary features in Microsoft Teams and apps like Fireflies and Otter? 

Belsky: Yeah, exactly like that.

Digital First: This really does look like the future of tax filing. A highly automated, seamless process, you have in-app video calls in the tax software, and so on. It seems to be a much better experience than an email-based process.

Belsky: That's our hope. That's what we're doing. We look at the old way and we go, why are people still doing this? We really try to understand the main pain points on both sides because we want all accountants and tax preparers to want to be on our network because our tools are so great. 

The opportunity with us is so great because you really can focus on things like serving customers rather than marketing yourself or billing or the things that you really did not become an accountant to do. We're trying to build these products to attract customers as well as accountants. 

Digital First: Are we likely to see anything like this in Australia or New Zealand or the UK?

Belsky: Obviously, QuickBooks is in Australia and Intuit Assist is a baseline feature of our products so a version of it will be available (but I’m not sure when). Right now we only do taxes in the United States and Canada. That could always change.

DigitalFirst: Is there a global strategy to do tax in every country?

Belsky: It really depends. I mean, we're not making a blanket statement like that yet. Also, tax is not as complicated in every country as it is in the US.

What drives all of our global expansion is that it's got to be that match, right? A big customer problem that exists that we can solve well. We're very prudent too. What is the right time as we move into new countries? There's still a lot of work we can do here in the States before we feel like we have fully saturated this market.

DigitalFirst: Do you have an end state in mind for Intuit Assist? When you have say 12 months under your belt of generative AI built into the system? What new features will be possible?

Belsky: I think more and more where we're trying to go is a “done for you” experience where consumers are probably just verifying our work and then hitting “File”. And we're getting there. The work that we're doing around documents is the place to start. Because we need to start with the data and get your data into this return. 

That's our challenge. How can we get more of that data sooner from you in a way that you feel confident we're going to use it effectively? 


These are the levels of automation that Intuit is bringing to its own tax return business. While TurboTax is mainly focused on consumers and self-employed, it is moving into small business with an Uber-style model where the software pulls in an Intuit-employed accountant towards the end of the process. 

It seems inevitable that elements of tax planning will start to appear in accounting software. For example, business owners could ask Intuit Assist for its advice on key spending decisions and how to optimise them for maximum tax efficiency. Intuit is gaining a lot of experience in answering questions for “lookalike audiences” – if you’re a carpenter in Los Angeles, TurboTax will have a picture of the likely expenses you will make and the most tax-effective way to make them. 

One of Intuit’s goals is to convince accountants to use its software to provide the fastest and most accurate tax returns. Will firms build their own AI-based systems to compete? Or will they share their bread and butter with software companies?  

Image credit: Intuit

Subscribe to our newsletter

Subscribe to receive the latest stories and new guides to your inbox. No spam, we promise.

By subscribing you agree to our Privacy Policy.