When AI Gets Tax Wrong: Guardrails for Complex Tax Reasoning

Aden: So today, we're here to talk about when AI gets tax wrong: guardrails for complex tax reasoning. I'm going to hand over to Andrew Noble, founder of LodgeiT. Andrew, welcome, and thanks so much for being here. We'll let you share your screen, check that's all working, and off you go.

Andrew Noble: Thanks, Aden. And yes, that's LodgeiT. Folks, you'd think that with a topic like this, I could have secured sponsorship from a professional indemnity insurance company, but needless to say, at LodgeiT, what we're interested in is building guardrails for complex tax reasoning, so you don't need to leverage insurance. And moreover, we're also very interested in how we collaborate and work with the profession, because we believe they are the right people to bring into the mix to help us get this right. So I am going to be introducing Jenny, Sami and Dev in a moment, but I just want to put some background around what it is that we're going to be talking about today.

Andrew Noble: So, we have based today on a true story, and that is where we've had an issue with the LodgeiT system, and a worksheet that was calculating in a way that we thought was right. We used AI to help us decide whether the worksheet was calculating right or wrong, and we had one of our users, who is a practitioner, who's here today, who's Jenny, who kept coming back and saying, no, your worksheet is not right. And actually, we had based that worksheet on what the AI… on what the ATO had given us. And as it turned out, Jenny was right, and we were wrong.

Andrew Noble: So. What do we know about AI? It recalls from memory. Essentially, it's a big memory system that has been trained on a lot of information, and it's got a good chance of paraphrasing and getting the information mixed up, especially when it comes to complex tax scenarios. It also has a habit of filling in the gaps. What do we mean by that? Try it yourself. If you've got a tax scenario and a set of problems around that scenario, what you'll find is that the AI will drop in a missing fact. For instance, if you don't specify residency, the AI will probably make an assumption that you're asking about a resident. But it's this kind of gap filling that leaves you exposed.

Andrew Noble: And also, at the end of the day, it doesn't leave a trail. So, when it finally derives a solution for you to a problem, let's say it's a numerical solution, there's no chain of reasoning that you can follow all the way through. And that, in and of itself, presents a problem, especially if later on you're going to have to show up in front of the ATO to explain how you reached your conclusion. So, our solution at LodgeiT to this is to introduce something called Rules as Code, and that is where we write the Act in something called Logical English. Not the whole Act. In this case, we focused on a very specific area that I'm going to be introducing our guests to talk about in more detail, which is Division 36.

Andrew Noble: But the thing with writing rules as code is that regardless of how often you run the process, how often you present the facts to the machine, you'll always get the exact same answer out. That's what a rules engine essentially does. It makes sure that the solution is always the same, given the inputs. And the good thing, obviously, on the flip side with AI, which we're all used to, is it's great at reading your scenarios in plain English, pulling the facts into the AI, and then allowing those facts to enter into the rules engine where the rules can be run. So, it feels like a seamless experience, and I'll show you in a moment what we're talking about in the actual product.

Andrew Noble: So now, for me to introduce our guests, who are tax professionals. We've got Jenny Haines, chartered accountant who runs Natural Numbers Accounting in Goonellabah in northern New South Wales. Jenny is the practitioner from our opening story. When our software carried forward too much tax loss, she knew the figure was wrong, and didn't let it go. She then reviewed the rules we wrote as a result, and corrected some of our section references along the way. Welcome, Jenny! We've also got Sami Charaf, who's director of PRACMD, where he provides tax advisory and business valuation services to the healthcare industry. He is a lawyer and a fellow of the Institute of Public Accountants. Mathematics, accounting, law and AI are right in his area of interest. Sami reviewed our FBT car calculator and our Division 36 rules, and his findings shaped the rebuild. Welcome, Sami.

Andrew Noble: And then we've got Dev Magoon, Certified Practising Accountant, works in public practice. He has reviewed our CGT small business concession rules and our Division 36 rules, and he has a knack for finding the quiet errors, the ones a calculator would ship without anyone noticing. Welcome, Dev. So, I'm gonna hand over to Jenny now, and ask Jenny a question. Jenny, what made you so sure the figure was wrong, and what kept you pushing forward when we kept coming back and saying, no, no, you're wrong?

Jenny Haines: Hi, Andrew, thank you. Basically, I knew I was right, and the reason I knew I was right is I was applying human thinking, and had read the Tax Act. So, the problem really seemed to be… one, I had no idea I was arguing with AI, I thought I was arguing with humans, and I checked the legislation, made sure I was right, because I don't like being wrong, and so I made sure that I was totally correct. And when your guys kept coming back and saying, no, no, we're right, I kept pushing because I knew that it was incorrect, and I was applying the spirit of the law, rather than just the hard fact. I knew the spirit was, if you have other income, you lose your tax losses. And regardless of how the program was applying that, or the AI was applying that, I knew the result had to be: this guy had other income, it had to eat into those losses.

Jenny Haines: So I kept pushing and pushing, and as you know, probably from your support people, I'm a nag in the background, and I do tend to keep pushing. But the thing was, AI thought it was right. It quite confidently said it's right, but I knew, after so many years of experience, that it wasn't right.

Andrew Noble: You went back to the legislation then, and you read the legislation again to check that, and formulated your result from the actual legislation, rather than just guessing through from memory.

Jenny Haines: Yeah, and I started out with legislation in a bookcase, and on paper. And so that's always been where I start from. AI may not always go that far. It will go, okay, I've found an answer. It must be the right answer, and it often isn't. And yeah, my personality is, I keep pushing until everyone agrees with me, and then we're all happy.

Andrew Noble: And Jenny, just in terms of reading the Logical English rules, you were one of the five people that worked with us on the project of going through and instantiating Division 36 into Logical English rule code. What was that actually like?

Jenny Haines: I found it a little bit of a challenge, because I've been coding since before there were PCs. That old. So, I started off with very much having to have a structured language, and an if-then sequence, which doesn't always work, because an if-then sequence doesn't take into account every possible scenario, only the ones the program has thought of. So, to me, it was a little bit of a challenge, but once I got the hang of it, and your support guys helped me understand it, it makes sense, because it's less rigid. But it also has to know when to stop and when not to stop in its calculation.

Andrew Noble: Yeah, so everyone in our team you dealt with gave you a fluent answer, and none of them could show you which step of the Act they'd applied, which obviously you then worked through. So that leads us on to Sami. Sami, you were paid to read the law as rules, not to check a calculator. What's that like, and what did you find? What was your experience?

Sami Charaf: So, it was actually quite a novel experience. My background also initially came from mathematics programming, then I did law and accounting. I practised in both law and accounting. For me, reading the legislation comes naturally, reading code comes naturally, and reading the numbers comes naturally. But doing this was not any one of them, but a combination of all three. Because what you had to actually do is say, hold on, AI isn't actually reasoning through this. It's working off correlations that it has built into its matrix, essentially, through all the training that it's had. It doesn't really know if something's correct or not. We need to tell it what is correct and not and how to do it. Otherwise, it'll just go off and do whatever it wants to do, based off pure correlation, essentially. So, the difficulty is that you have to go in and read the law, take everything you know about tax and accounting,

Sami Charaf: and then try and put that into pseudocode for a machine to go through. And this is actually what Logical English is. It comes from Prolog, and it was developed by Bob Kowalski at Imperial University. It's essentially trying to turn tax rules, legislation, not just tax, but legislation generally, and rules, into something that a computer can execute deterministically and give you an answer. And where it doesn't have an answer, it can tell you exactly why it doesn't have an answer, what needs to be worked out. And that's the challenge. You think that what you are doing is simply translating a set of rules into Logical English, and it reads logically, like the example on your screens right now. That is supposed to be something straightforward, but realistically, it has so many failure modes. And we don't necessarily think through them, because we just naturally assume we know what we're talking about.

Sami Charaf: When we say, hey, you want to calculate a loss? Yeah, of course we realise that the whole point of Division 36 is that if you've got a loss… to get a loss, you have to have income, assessable income. You use up your deductions. If there's more deductions than what you had, you end up with a loss, but you've got to offset this loss against your exempt income, your net exempt income. And if you don't do that, then you can't just carry forward that loss. You have to do that first, and then it gets carried forward. And same thing with having a loss in future years. You just instinctively know that you've got to apply them from the oldest year first to the newest year. You instinctively know as a tax agent that there's limits on certain deductions. You can't just go ahead and generate massive losses off things like deductible gifts that you've made to DGR entities, or, for instance, personal superannuation contributions. There's a limit on that. There's a limit on pensions, gratuities.

Sami Charaf: These come naturally to us. They are in the Act. They all have legislative standing. They don't automatically get flagged as correlated to that rule, because when an AI goes in and looks at this, it might not be in that section, and it needs all that background knowledge and information of what happens in this case, in that case. It doesn't have that. We need to program that in, and make sure that the rules reflect both the legislation, as well as the bigger picture, in terms of how we interpret the legislation and the tax Act. And again, it's not just the legislation. In this case, it was getting it from an ATO worksheet, which, again, is not legislation. One of the hardest things I've found, especially in the accounting space, from a legal perspective, is we accountants tend to find it hard to go straight to the legislation. It's not second nature as a graduate. It's something you build over time, whereas it might be second nature for a lawyer.

Sami Charaf: But at the end, tax is really an area of law. When it comes to applying for and getting an assessment from the ATO, applying rules, these are laws. The ATO worksheets are not authoritative. The ATO guidance on their website, unless it's in a binding ruling, is not authoritative. When you're before the tribunal, you're before the court, you need to actually come up with the rules. And that's where this comes in. What we went through is all the legislation, double-checked that everything was translated into code that the machine can execute, and then we made sure that we thought of all the edge cases, where it's not immediately apparent, but where it might fail the limits of the system.

Andrew Noble: Yeah, fantastic. Thanks, Sami. I guess that leads us on to Dev. So Dev, the error wasn't in the arithmetic, it was in what the rules left out. And errors like these don't announce themselves. Why are the quiet ones the dangerous ones? And moreover, what was your experience like? Because you have actually moved on to our second level of reviews across the Div 36 rules. So maybe tell us a little bit about those questions and what the experience was like.

Dev Magoon: Yeah, thanks, Andrew. I guess the quiet ones, it's because nobody goes looking for them. Because the thing is, if it's just a clear error or a loud error, it gets caught on the first review. Like, the figure looks odd, someone asks, and then it gets fixed. But a quiet error, it looks good on the surface, and it only fails on the client that you didn't think about. So in normal cases, it would work, but because tax law is so wide, and because there's so many different types of scenarios, it wouldn't work on the unusual case, so I guess that's why it can get missed quite easily. And one of the earlier bounties that I was working on, it was regarding the CGT active asset test, which is quite normal for everyday practitioners, because people keep buying and selling properties. And when I was using it, the calculator gave sensible answers, but then when I read the rules behind it, I found three things.

Dev Magoon: One of them relied on a subsection that had been repealed. Another one was that it applied to a situation that it didn't cover, and one of them actually misread the definition of what an active asset is under section 152-40. So, that actually matters because the active asset test is often the gateway to other small business concessions. And, you know, when a client sells or buys a property, whether they can pass, it can be the difference between whether they get a large tax bill or a small one, so it's quite important. And I was also reviewing Division 36, as you mentioned earlier, and it was pretty much the same story. A lot of the scenarios, it reproduced to the exact dollar, and on the surface, the test passed, but catching those quiet errors is often in the gaps, like, in the steps itself, rather than the output. So, in summary, I would say that quiet ones are quite dangerous, because they get through an initial review, and then

Dev Magoon: often, years later, once someone lodges a tax return, it's often the difference between whether the ATO would review it and find it later on. And yes, that's why it's important to always check it in each different type of situation, because it can get missed.

Andrew Noble: And just around the work scenarios, how did you find our work scenarios? Because as you work through our rule sets, you obviously had to review the scenarios themselves, because we produce the rules, and we also produce the scenarios, because there's no point having rules unless you actually test them with real-life scenarios. So, how did you find those scenarios?

Dev Magoon: Well, that's the thing, those scenarios, you have to run them not just once, but you have to keep running them, because it does produce the answer that you are looking for, but then, oftentimes, once you run it through multiple times, then you find… because everything has to link back to the Act and the tax law, so it often takes time to find out that, oh, actually, this might be missing this specific step, because a lot of tax law is all about judgment. So, you just have to run it more than once to actually know if it's right or wrong, and apply our own professional judgment to it.

Andrew Noble: And in terms of those rules, can you give us some kind of sense of the complexity of something like Div 36? Because I think it's not well appreciated that there are sections of the Act that are quite deeply nested in terms of the way the rules are structured and hung together. Can you give people a sense of what that is like around Div 36?

Dev Magoon: Yeah, so Div 36, it's a bit complex, because it goes very much into detail. Like, one small thing can be the difference between having an answer produced differently. For example, things like having split receipts, because if one of the receipts is entered as an exempt amount, or one of them's non-assessable, like, how the calculator would want to get that, whether you would put it as a whole amount or based on split receipts. You have to be very mindful and specific, because one small tweak can really get a wrong answer. So I think that's quite important when you're testing out the scenarios in the software. And the AI, unfortunately, it can't get it right every time, because a lot of it, as I said before, is about professional judgment, and having worked through different types of scenarios.

Andrew Noble: Yeah, okay, fantastic. So, folks, I'm gonna push on a little bit, and let's have a look at what the rules give you. This is obviously what our objective is when it comes to building the rules engine into Tax Genii. With the optimal outcome, you give the AI your scenario and your prompt, or your query, and you get a figure, and the proof. So that would be one outcome. The second is, you would want the AI to let you know if there's an additional necessary fact that it needs. A good example of that is residency, and we'll have a look at that again in a moment. And then the third case is, there's certain situations, like in the case of Division 36, where there's no point asking about Division 36 in the context of a company, because a company doesn't have any relationship with Div 36. So you'd want the AI to say, hey, that's outside of what we need to talk about here. You can't talk about companies and Div 36 in the same sentence.

Andrew Noble: So, this is what we've now got in Logical English, embedded inside of Tax Genii. And I will show you in a moment how the rules actually are displayed in terms of how a conclusion is reached. Let's actually jump in now and have a look at what this is like in Tax Genii. So, if you come to the LodgeiT website, you can sign up and have a play with this. I won't run all of it now, because we've got a limited amount of time, but I'll take you through some examples that I've saved here. We will start with this. This is the example of putting in a question relating to a company and Div 36 in the same prompt. And our Tax Genii, which, by the way, you can bring and plug in your own AI, so this is really a framework and a method for you to use our search engine capabilities in the background.

Andrew Noble: There is actually quite a few players in the market already that do similar things, whether it's with a vector database in the background, or a graph database. None of that really matters. If you're in the tech space, that kind of stuff can be put together pretty quickly and easily. The thing that matters is how do you ensure that the knowledge and the information that's coming out is accurate. And in this case, you can see, because we're running the rules in this, the AI here is then seeing that from the rules. It correctly incorporates the calculator refusal verbatim, stating plainly that this specific calculation is out of scope for the automated calculation, because the query concerns an Australian private company, excluded as the file covers only non-corporate Australian resident individuals. References to Division 36…

Andrew Noble: …and so on, follow standard legal structure, with no apparent hallucination of laws. But you can see here that we will state clearly when there are rules involved, if the calculator's declining to answer further. Now, as we were saying before, we've only just got started on building these rules. There's probably thousands of rules that we're going to have to get to, and the way that we're dealing with that is we're running these bounties. Dev and Sami and Jen are some of our first collaborators who have come in and taken part in our bounties. These are the bounties that we have been running so far, and we've obviously got lots more planned, but what it essentially means is that we work through and build out not just the rules, but also the test scenarios that go with the rules. So, Div 36 is a good start. In this case, the calculator declined to carry on any further.

Andrew Noble: This is a separate case. An individual, and you can see here, has assessable income. Same prompt, but essentially, the AI stops and goes, well, I'm not going to go any further, because clarification is needed whether the taxpayer is an Australian resident or not in that financial year. So this is a case where the average AI will just give you a solution and an answer, but it won't tell you that it's essentially synthesised some information to put into the missing slot for whatever information should have been there. And then, this is a solution that actually worked. The deterministic calculators ran. And so, in this case, it has been stated that the person is an Australian resident, and then you've got your issues, the relevant law, links to the relevant law, the application of the facts, the conclusion, and then, again, how the structured advice has been reviewed.

Andrew Noble: And in this case, the deterministic calculators have been applied to this. And as I was saying before, one way to limit how AI makes mistakes is to make sure that it only works across the knowledge that is relevant. So, what our search engine is doing in the background is pulling out chunks of the legislation. You can see here, there's a chunk of the legislation that's pulled out and becomes essentially the ground truth for this AI to work with, rather than just letting it run across an unlimited amount of vector space and essentially synthesise its own knowledge. So, we limit what it can gain access to, and also, any piece of knowledge that it uses gives you the jump-off point to the actual ATO legislation. One thing you'll notice here is, because this deterministic calculator ran, you then get the Logical English script, the raw proof code for how that calculation was derived. Yes, you'd want to take your time to read through this, but then we also use an AI to read these rules and put it into an easier format for you to read.

Andrew Noble: So, you've essentially got a set of guardrails there that you can use to help you in understanding and applying the legislation. This doesn't mean that you don't, at the end of the day, still carry your obligations for making sure that you fully understand how the legislation is working. We're simply giving you some toolkits that make this easier, and I guess give you some additional degree of confidence that some other professionals spent a lot of time working with the rules to get this right. Other things in terms of how Tax Genii works: go to the LodgeiT website, you can try this yourself, it's all free. You will notice in here that we give you some free models to play with, but if you want, you can bring and plug in your own models. So, you're welcome to come in and put in your own top-of-the-range Claude, Gemini, all the latest models you can plug and play.

Andrew Noble: You can also use the MCP server, if you want to just run this directly without even coming to our website. What you can do here is, we give you the ability to configure three agents across the top of this, so that you end up with a normal agent that does the first run, a second agent that would work as the analyst, and then the third agent that would work as the critic. So you could add your own keys, or you could just use the same model for each of those. But again, from our research, even having an analyst, a critic, and a judge won't, without rules, prevent the AI from doing things like filling in the gaps, or, when there are complex areas of the legislation and numerical calculations, making mistakes all the way through.

Andrew Noble: So, talking about numerical calculations, if you come along to the LodgeiT website, we've just launched a lot of calculators that you can call with your AI. We're putting together a calculator constellation, and just like with the Logical English rules that we're building for tax reasoning, we encourage others to build calculators, or share with us the entry points for those calculators. Our long-term vision is that there will be thousands of calculators globally that AIs will be able to leverage so that they don't hallucinate. At the moment, you'll see we've got in the range of 25-odd calculators. We've got fringe benefit calculators, Division 7A calculators, we've got a double AASB depreciation calculator, HP calculators, and we've got some tax depreciation calculators coming in the next few days. So, you can call our calculators with your agents through MCP. You've got Tax Genii, which you can, again, access through the MCP server with your agent, and then

Andrew Noble: we have our accounting toolkit, which gives your agent the ability to think more like an accountant, and reason deterministically across accounting information. Right now, you can test it with the datasets that are there. Later on, we'll make it so that you can work with your own GLs from source systems like Xero, QuickBooks and MYOB, and any other system. But essentially, this SBRM engine checks the logic of accounting, and it works by checking balance sheets balance, equity roll-forwards roll forward correctly, cash rolls forward correctly, cash flow sections add up, any notes add up. So it's really a toolkit for agents to work deterministically, rather than hallucinating and guessing. And then finally, if you're not familiar with LodgeiT, we have got Client Relay, which is a set of tools that involve engagement, getting paid using Stripe. At this stage, you've got QuickBooks in the background that you can use for the accounting side of things.

Andrew Noble: KYC AML checks built in, and then also client forms, where you can put together the forms that you need, structured in a way that allows you to collect information from your clients seamlessly, and with conditions in them. So if the client says they've got, for instance, a car deduction, then it'll say to them, do you want to go forward with the cents per kilometre method or the logbook method? So you can actually build it out based on conditional logic, and you can also use your agents through the MCP server to do that. We'll be putting in our own agents to make it easier for those people who don't want to think about the agent part of all of this. But anyway, if you haven't looked at LodgeiT, come and have a look. We're bringing out, as I'm showing you here, lots of toolkits and capabilities that you can use with your existing agents. You don't have to come and use our agents, we're very happy for you to plug and play your agents, and lots of free stuff here for you to play with.

Andrew Noble: So come and give this a go. Then, finally, I want to sort of go, well, what's in it for you? I'm gonna get back on very quickly to working with us around our bounties, because we actually do pay people to help us read our rules and our scenarios. Jumping back in with Jen first. Jen, what do you think's in it for you as a practitioner here?

Jenny Haines: Well, I'd like to have a lot more accuracy that I can rely on, so I can do more interesting stuff, rather than just tax.

Andrew Noble: Okay, so hopefully our Div 7A calculators and the other calculators that we're putting together will expand out into the accounting space for you with that.

Jenny Haines: Yeah, yeah, and tax and compliance is the boring part of accounting.

Andrew Noble: Okay, so you…

Jenny Haines: …move on to rely on the accuracy of what's happening there, so that I don't have to spend as much time on it.

Andrew Noble: Yeah, makes sense, and that's obviously, at large, what we want to work on with you. So, going across to Sami, what do you think's in it for you, working on these things with us, and this kind of technology?

Sami Charaf: I think you get to be part of the next evolution of tax, so I really think this is where it's headed. Because we can't just have it hallucinating, we need to give it guardrails and a framework for it to work within. And your input doesn't just help one client, but becomes a framework for future responses, and it helps provide advice to thousands of tax agents over thousands of scenarios, in a scalable and defensible way.

Andrew Noble: Fantastic. Now, next we will ask Dev. Dev, what do you think's in it for you, collaborating with LodgeiT on these rules and what we're working on generally?

Dev Magoon: I guess it's figures that you can trace back, and having something you can review and defend, because you know that it's been reviewed by a practitioner. And also, I guess, the bounties pay you to do what we do every day, which is reading the law carefully and applying scenarios to different situations, and so the tools are right there for you.

Andrew Noble: Yes, that's fantastic. Well, thanks, Jen, Sami, and Dev for collaborating and working with us at LodgeiT to get these rules right, and coming on today and giving us your thoughts and input. Any questions out there? Let's just have a quick look and see if there's any questions. Where are the calculators in LodgeiT? Okay, so if you want to come in and have a look, come through to the LodgeiT website, and you will find AI that calculates. You can start here, and connect through to the calculators. And you'll actually see we run this on a separate website, logic.org, because the long-term vision for our calculators is… we're getting started. You can see here we've got our first external contributor who built a calculator, and we linked to it. It's a coal LSL levy calculator. That's Ryan Dugood. He's come on and built that calculator, as long as the calculators meet our

Andrew Noble: design and publishing standards, and meet our rounding conditions. So this is the sort of thing with building a global ecosystem of calculators. You want those calculators to all work in a very standard kind of way, so rounding becomes an important part of what it means to put the calculator together in a standardised manner. So, if you want to learn about how to put your calculators together, if you love building with AI… these calculators are all built with AI and tested with AI, and with humans testing as well, but come along and give them a try. As I said, we've got a full array of fringe benefits calculators. And we are putting up other calculators and tools, like a FANO classifier. If you've used LodgeiT, you'll be familiar with the fact that we've got an amazing classification engine that helps put your accounting balance sheet and profit and loss line items

Andrew Noble: into our financial reports, and worksheets, working papers, and also tax forms. So come and have a look at this, but at the end of the day, we're looking to build a marketplace. We've got a vision for how collaborators who introduce calculators will actually earn revenue from the calculators that they build in the ecosystem. We certainly don't see ourselves being able to build all the million calculators and Logical English rules across all the English tax jurisdictions: Commonwealth, and other English tax jurisdictions of the world. So, this is just a start, and part of this is an open source project, so if you feel like you can contribute, jump in, reach out, and have a crack.

Andrew Noble: It is possible now to have your AIs help with constructing those Logical English rules, but the key is to have professionals like Sami, Dev, and Jen actually go through the rules and the scenarios to make sure that those are accurate before you put those rules out. So that's what we're doing at LodgeiT. Thank you all very much for attending, and thanks, Sami, Dev, and Jen for joining today.

Dev Magoon: Thanks, Andrew.

Aden: Thank you, Andrew, and thanks for that talk today. Thanks to everyone that's come along to our session.