Imagine using a dedicated AI assistant for tax research rather than looking up the legislation yourself. Andrew Noble, founder of LodgeiT, a cloud tax program that works with Xero, QuickBooks and MYOB, recently demo’ed Tax Genii at the recent QuickBooks conference in Sydney.
Tax Genii provides answers based on combined information from Australian tax legislation and the Australian Tax Office’s website using the same model as ChatGPT.
Noble, who also owns an accounting firm in Perth, Western Australia, is an avid explorer of technology. DigitalFirst.com caught up with Noble at QuickBooks Get Connected on 2 November and asked him how the technology works and how he thinks it will change the way accountants will deliver tax services.
This interview has been edited for length and clarity.
DigitalFirst: How long have you been working with AI?
Noble: We've had an interest in AI since 2015 when we made an automated classifier. It brought in accounts from QuickBooks, Xero, MYOB, and then LodgeiT classified them, which automated the filling of tax returns. We were also working on blockchain quite extensively at a research level. Believe it or not, we spend quite a bit on research.
A year ago I started to notice generative AI. Ever since Open AI provided access to their APIs, I've had a guy working on a tax solution we call Tax Genii. Tax Genii is all about using Open AI capabilities over the top of the tax legislation.
DigitalFirst: What does Tax Genii do?
Noble: It lets you chat to a chatbot that calls into the legislation, finds the relevant parts – not just the tax legislation but also the Australian Tax Office website, which is quite extensive and has lots of great resources and examples.
If you're doing a tax advice piece, you can use it to put together your tax advice. It's all about having the references (to legislation and the Tax Office) in there.
DigitalFirst: Does Tax Genii know the whole tax legislation?
Noble: You can't just take your whole tax legislation and slam it into Open AI because it's got a token limit (generative AI translates letters and words into tokens to perform its analysis). You have to pick out the bits that you need, which is the tough bit. We've been on a journey and figured out what it's capable of. It's not just parroting stuff – it is way beyond that. It's reasoning, and that's the foundation of what we do in our roles as accountants. We reason about complex tax legislation, and ChatGPT does a pretty good job. I'm super impressed.
DigitalFirst: Is Tax Genii part of LodgeiT or separate?
Noble: Tax Genii is really about advisory. I don't think we'll put it straight into the tax product. That is something that we'll have to do a bit more research on.
DigitalFirst: Do you have an idea as to how much time you'll save on doing returns using Tax Genii?
Noble: I think it's too hard to say because it will be on a case by case basis. It will depend on the kind of research you're doing and on the complexity of the case you're investigating. There's lots of conditionals that will come to bear on how it might save you time. But it will definitely save you time on drafting your responses and doing the research.
DigitalFirst: What would it look like as a user to have generative AI built into a tax application?
Noble: There are a lot of fields in a tax form that can be filled from granular data sets coming in from QuickBooks, Xero or MYOB. That is your first level of fill. And then you've got information in a typical CRM database like your addresses.
So that's easy, but there are other fields of a tax return. For instance, company tax returns have a field that says are you a base rate entity or not? It takes a bit of reasoning (compared to) how much income is coming from investments versus trade, for instance.
There is a threshold where you're either a base rate entity or you're not. And if you're not, then you can't take advantage of the lower tax threshold for a company. The AI can reason and figure that out.
What it ultimately means is this idea of “zero touch tax”. Data comes in from downstream software applications like QuickBooks, Xero or MYOB, and all of those fields are handled automatically. As an end user you don't do tax. Tax does itself.
DigitalFirst: What types of returns are you talking about? Individual, company, trust, or all of them?
Noble: I just gave you as an example of a company tax return. Something like an individual tax return is still going to take interaction with the client around what they have to report and their deductions. If you're a tax advisor on the production side, there's (effort) in talking to a client and saying, “If you're a metal worker, these are the things that you can claim.” The tax advisor for individual tax returns is always feeding back to the client.
DigitalFirst: Presumably, those ideas will be the same each year. The metal worker is not going to have radically different expenses from last year.
Noble: There's a thousand different occupations so you can have a ChatGPT based system that will interact with you, feed you those ideas, and be part of an information-gathering process.
With tax forms, it's always going to come back to what those datasets look like. And who's got access to what and how well they're presented. That's still a record keeping problem that can't be solved by generative AI. The only thing that generative AI can help you with is reaching the right conclusions, and helping you discover the things that you should be thinking about.
DigitalFirst: Your example of the base-rated entities sounds like a maths problem. There are quite a few examples of generative AI not doing that so well.
Noble: I don't think it matters. I'm terrible at maths, I have a real difficulty adding even the most simple numbers together. But that doesn't mean that I don't know how to pick up a calculator and punch in the numbers. Generative AI is exactly the same.
From my experiments it's a case of leveraging the right tools for the right problem. Open AI for instance already has a compute plugin called Wolfram. If you activate that Wolfram plugin, ChatGPT understands how to use Wolfram’s computational capabilities for complex problem solving.
DigitalFirst: It's quite human, isn't it? ChatGPT is good at language and also knows how to pick up a calculator and use it. What other plugins are useful for tax or more generally?
Noble: Well, right now the plugins are limited. With Tax Genii we're taking advantage of whatever is available on Open AI’s APIs. I believe what will ultimately happen is you'll be able to use any range of different algorithms.
My demo today is showing a tax planning problem for trust distribution. The system gets it half right. But that's something that can be solved by using the correct algorithm to do the computation. Now, I've found the correct algorithm in the Wolfram Mathematica suite. I'm just not capable of implementing it myself. I'm waiting for an engineer to put it into play correctly. And then it'll get it right every time. We just haven't done it yet.
DigitalFirst: Given the current pace of change, how far off are we from the zero-touch tax return?
Noble: Well, I guess it always comes back to investment and how much money is in the bucket for these kinds of things. I would say 18 months, maybe even less – 12 months. Within the next tax season you could see a lot of this stuff implemented.
DigitalFirst: What impact would that have on the role of accountants with business clients?
Noble: I think it will depend again on how those business clients are managing their datasets. If you don't keep clean records, it's not going to help with that. That's a problem for the people like QuickBooks and Xero to worry about. But in terms of actual tax advisory, I think that's going to be impacted, probably even before the tax filing process. Certainly, there are a lot of things that go through a tax form where an AI assistant could follow you through your tax process. The system would be a tax preparer and give suggestions and advice as you're working on a tax form.
DigitalFirst: That sounds exactly like a description of Intuit Assist with TurboTax (a consumer tax software in the US). Have you seen that yet?
Noble: No, I haven't. But something like that. And it becomes easier to implement and deploy. Intuit Assist is probably something that had thousands of hours of research put against it before the generative AI models came along.
I think AI democratises intelligence; it brings intelligence into products far more easily. We're a small company, but now we've got access to Open AI. And everyone else does as well. I think it comes back to innovation and the way that companies deploy.
DigitalFirst: Why do you think generative AI will move so fast with tax?
Noble: Tax is a great subject because tax is language based. There are aspects to language that make it conducive to generative AI, particularly tax because tax is what they call a deontic logic. It has a particular structure; it's associated with things that you can do and things that you must do.
You must lodge your tax return if – and there are always conditionals – if you earn a certain amount of income. You may claim motor vehicle expenses if you use a car for work. You're not permitted to claim clothing unless – another conditional – it's a uniform. There are ways to construct tax legislation in a way that is in a controlled natural language.
And that's where I think systems like generative AI really excel in the language space. The more structured the language, the more capable that language model is of producing an accurate result. Open AI produces code because it's seen highly structured examples before. The more highly structured the examples, the more likely it is that you're going to get very concise, logical outputs.
DigitalFirst: Generative AI can sometimes hallucinate, or invent answers. Do you see this as an issue for assistants like Tax Genii?
Noble: From my research, the bigger the model the more and the more data points it's trained on, the more accurate it becomes at resolving to the exact outcome that you're looking for in the real world.
For us, hallucination is more about getting the logic wrong. I've actually got an example of that where the logic is not quite right. But as the models get bigger and bigger, and they have more and more training data, they resolve more and more accurately to an exact outcome.
The story is that by 2030 the models will already be orders of magnitude more capable. They're already more capable than me in a lot of those domains. I can't write code, but these things just figure out that code for a division 7A interest calculation and spit it out.
People say, you know, this chatbot, it's just a parrot. But it’s not – I've seen the reasoning. I can see it doing it. It's like, wow, that's smart. That's smart. It's not dumb.
DigitalFirst: Thinking five years ahead, once generative AI has been in the market for a good amount of time, what do you think the accounting profession will look like?
Noble: I suspect that the firms on the edge of the curve or ahead of the curve will be automating more and more of their compliance processes.
Generative AI could help you summarise information from a combination of tax and accounting. So it's definitely not just about generative AI. It's about all of the algorithms that you can bring to bear – and there are a lot of other algorithms.
And it's on implementing the basic stuff like using high quality bank feeds like they're talking about here with Open Banking.
You need the right kinds of processes and procedures. Otherwise it’s still junk in, junk out. If you don't resolve those problems, you're never going to get high levels of automation.
Generative AI is something that can help you to make decisions, gain new kinds of insights that you might not have had otherwise. It's not the panacea to shitty data. So if you still want to go and bang away and key in data from system A to system B, it's not going to help you.
The cloud stuff started 10 years ago and now here we are. Things can happen quite fast. So probably in five years, you'll have your tax practice in your pocket. And no one's going to be filling tax forms anymore.
I think you'll probably see people still using old desktop software in five years, whereas the cutting-edge guys won't be doing tax anymore. And it will be a kind of strange dichotomy.
I think you'll also see new business models creep in. Accountant websites with advanced Chatbot functionality will provide three things. One, information lure to onboard clients i.e. chatbot tax assist followed by sign-up option; two, customised responses for clients (the client would need to be logged in); and three, harvest information for form filling purposes, then feed forward to systems like LodgeiT.
DigitalFirst: What do you think is the future of audit, then? We’re already seeing systems like Dext Precision and Xbert look for audit errors on a real time basis which you can then correct on the fly. Do you think we will still need audits?
Noble: You mean monitoring in real time, a real-time audit? Managing the exceptions as they pop up, I think that is the right way to do it. Because if everyone had that kind of technical approach, imagine how easy it is to fill your tax form. The typical tax accountant doesn't have time to go back into the books. I don't want to go and look in QuickBooks and Xero. Bring the data in, fill my tax forms.
(A business just wants an) accountant to take care of all that stuff for me. That's the business model for accountants.