How AI Can Enhance Global Tax Systems

In the latest In the Pages video, Alexandra Karadima, a senior tax manager based in Luxembourg, and Robert Goulder of Tax Notes discuss how artificial intelligence is changing international tax administration.

In the Pages is a video series produced by Tax Notes. This transcript has been edited for clarity.

Robert Goulder: Hello, I’m Bob Goulder, a contributing editor with Tax Notes. Welcome to the latest edition of In the Pages, where we take a closer look at the commentary in our print and online publications.

Our topic for this week is the role of artificial intelligence in tax administration, and our featured author is Alexandra Karadima, a tax professional based in Luxembourg, and her article will be appearing shortly in Tax Notes.

Alexandra, thank you so much for joining us.

Alexandra Karadima: Thank you so much for having me.

Robert Goulder: Now, some countries are already using AI tools in their tax systems. I tend to think of it as this thing that’s going to happen eventually. Many years from now, there’s going to be AI. But some countries are already doing it in the form of a virtual assistant. I think maybe your article mentioned Spain and Greece were doing this.

Can you tell us about these virtual assistants? Are they reliable? Do people enjoy them? Do people know they’re interacting with AI as opposed to someone who works for the revenue body?

Alexandra Karadima: Yeah, in fact, a lot of countries already use AI in their tax systems. Two of the examples are indeed Spain and Greece. The Spanish Tax Agency has incorporated AI-driven virtual assistants to assist taxpayers with frequently raised questions with respect to the tax filings, the deadlines, and the understanding of the VAT, personal tax, even tax credits.

The AI chatbots provide prompt responses to those questions, reducing the time the taxpayer typically spends waiting for assistance. And even beyond routine queries, Spain’s tax system also uses AI to prefill tax forms using data that the agency already holds, for example, data — income that has been already reported by the employer — to reduce the errors, but also the time spent by the taxpayers. And finally, it also uses the AI for fraud detection, identifying related patterns in financial data.

The same is for Greece as well. Greece’s Independent Authority for Public Revenue recently incorporated AI, also adopting this AI-powered virtual assistant to provide real-time support to taxpayers. It’s reliable and they provide reliable answers to several questions. This chatbot, in fact, helps with tax-related queries, but also documents, deadlines, administrative — and therefore they reduce administrative bottlenecks and support taxpayers.

In addition, Greece leverages AI to conduct automated checks for errors, and they try to identify discrepancies in tax filings, enhancing compliance by allowing taxpayers to correct issues before they escalate. So the use of AI by Spain and Greece is a reflection of a larger technology-driven transformation of the tax administrations worldwide. AI supports both countries in maintaining consistent data. Then it automates error detection, but also enables the tax authorities to compare year-over-year data and identify anomalies, prioritizing typically high-risk errors that needs higher scrutiny, like increased scrutiny.

And finally, they improve satisfaction of the taxpayer via real-time support. And this is the nice part of AI. AI in tax compliance enhances transparency and simplifies processes in a way that builds trust for the taxpayers and eventually fosters a proactive compliance culture, encouraging a more collaborative relationship between the taxpayer and the public administration.

Robert Goulder: It’s hard to argue with that. Those are all good things. It sounds like there’s no downside to it, but maybe we’ll worry about that later.

Your article uses a term that I had never seen before: “predictive justice.” Asking AI tools, “Here’s my situation as a taxpayer. How would a court or a judicial tribunal look at this issue?” Fascinating. It almost sounds like you could train AI to be judges, Tax Court judges, in the future. Can you tell us about this application for the judicial area?

Alexandra Karadima: Of course, predictive justice refers to the application of the artificial intelligence and data analytics to forecast legal outcomes such as how court or tribunals might rule on specific cases and issues. This concept has gained traction in legal tech, where tools and algorithms analyze past court decisions and historical data to identify patterns and trends that can inform predictions about future rulings.

Now, how is this possible? Predictive justice simply utilizes large datasets, including past court decisions, but also legal briefs or even judges’ rulings. And the machine learning algorithms can analyze this data to identify correlations and trends that may indicate similar cases — how similar cases could be resolved in the future.

AI technologies often incorporate natural language processing to interpret legal language and terminology. This capability allows algorithms to actually analyze the text of legal documents, understand context, and extract relevant information, and by examining those data, AI can identify patterns regarding how different factors, for example, the type of the case or the parties involved or the jurisdiction involved, influence outcomes.

For example, it might reveal that certain judges rule more favorably in specific to certain areas of law. And predictive justice has quite a few benefits, but also a lot of challenges. Predictive tools can assist lawyers to craft strategies on data-driven insights, and they can help them allocate their time on cases that actually require more attention — and they’re more complex — and therefore gaining some time. But also for the courts because there are certain cases where the lawyers, based on the data, they decide not to pursue a litigation, and therefore, justice is being helped.

It also democratizes access to legal information to individual and smaller firms because typically, up until now, it was the larger and rich data firms that had this kind of information. In the internet, a lot of people talk about consistency in rulings because of predictive justice. I do not necessarily agree on that because for me, the evolution and change in positions of rulings reflect the evolution of the society to a great extent, so I will not take that as a benefit.

There are a lot of challenges and ethical concerns with respect to predictive justice. AI systems are only as good as the data they are trained on. If historical data reflect biases such as socioeconomic, racial, gender biases, predictive tools may perpetuate or even amplify these biases, and this would result in unfair outcomes. Many algorithms also operate as black boxes, which means that decision-making is not easily understood, but in the context of justice, the decision-making and the methodology that has been followed is crucial. So predictive tools can be a bit blurry with respect.

I imagine also that a potential drawback is that legal professionals may over-rely on predictive justice, which would undermine the legal professional judgment. Therefore, while AI has and may play a significant role with respect to predictive justice, it must be implemented with caution. And ensuring that this tool is transparent and ethical is crucial to maintain integrity in the legal field.

Robert Goulder: That is just amazing, that the AI tools are so sophisticated, they can even narrow down to the actual judge and say, “What are this judge’s tendencies?” That’s just fascinating.

So we’ve talked a little bit about compliance. Now what about enforcement? And I’m thinking about audits. There’s a lot of sensitivity as to who is selected for a tax audit and who isn’t, because audits can be very, very disruptive and so forth. What is the role for AI in identifying audit risks?

Alexandra Karadima: Yeah, that’s very interestingly a part of AI tax audits. Actually, another example of an administration deploying AI is the U.S. If I’m not mistaken, in September 2023 the Internal Revenue Service announced that it will deploy AI in an effort to increase fairness in tax compliance and audits, with the intention being more on high-income earners, partnerships, and large corporations.

The same with the example in Greece in December 2023 — the Greek authorities announced that they will deploy AI in the fight against tax avoidance and tax evasion. As AI systems may collect useful information from various sources like banking transactions, digital platforms that we use on a daily basis, or even social media, the AI may detect, in real time, suspicious movements which could indicate tax fraud, evasion, or avoidance, unusual discrepancies in the income of individuals and businesses, but also unusual transactions or even money transfers.

AI enables revenue authorities to develop high-risk profiles for taxpayers, based on historical data, but also behavioral patterns, and this allows the tax authorities to prioritize audits of high-risk cases, optimize resource allocation. At the same time, they can automate simpler tasks like review of legal docs, for example, and therefore enhance their efficiency. Most importantly, AI may detect, as I said before, potential fraudulent activities, suspicious claims on a real-time basis. But therefore revenue authorities may improve accuracy and efficiency, and that would lead to better tax collection results by default. However, careful attention again should be paid to data privacy.

Robert Goulder: Now, I have to ask you about transfer pricing because that is an area that we think about a lot, and it seems like in transfer pricing litigation, you have endless squabbling over whether intragroup dealings were done consistent with the arm’s-length method, which is admittedly sort of an amorphous concept. It can be very hard to pin down what is actually arm’s length. Is this something that AI could help with? Telling us what’s arm’s length and what’s not, looking at comparables? Is that something AI might be good at?

Alexandra Karadima: Yeah, I like the question; it’s yes and no. So I would expect that the deployment of AI would play a significant role in [the] transfer pricing field. By utilizing AI-driven analytics, tax authorities could”? “field by utilizing AI-driven analytics. Tax authorities could potentially process and analyze datasets related to intercompany transactions, thereby identifying discrepancies but also transfer pricing manipulation. For instance, machine learning algorithms could benchmark pricing against similar transactions within the same industry, very quickly. AI tools could quickly gather relevant data and access compliance and identify test compliance with respect to local but also international transfer pricing regulation. This not only accelerates the tax audit process, but also enhances the accuracy of transfer pricing regulation.

AI can develop also predictive models to forecast potential pricing strategies. They can also use the natural language processing tools to quickly analyze the contracts and agreements around intercompany transactions and identify those terms and conditions that are crucial for transfer pricing compliance. And they also can analyze the communications and context so that they identify the motivation for the pricing strategies.

However, the effectiveness of AI in the field of transfer pricing heavily relies on the quality, but also quantity of data available, which means that inconsistent or incomplete data can lead to inaccurate analysis, and also transfer pricing often involves complex transactions [so] that comparables are very difficult to identify. And therefore, I expect that the AI could not assess, efficiently, unique arrangements. As a result, while AI could enhance analysis and efficiency in transfer pricing, human expertise remains essential in transfer pricing to have accurate results.

Robert Goulder: I wanted to ask you about languages. When you think about the OECD and their transfer pricing guidelines or their model tax conventions, they tend to publish everything in French and English, but the world includes a whole lot of other languages.

Are there any linguistic issues related to AI? I mean, is it going to help us overcome those issues, I guess, and am I correct in thinking this could actually affect how mutual agreement procedures are handled, that we could have the inventory of MAP cases could go down if AI becomes an effective tool there? What are your thoughts on that?

Alexandra Karadima: Certainly, the complexities of international taxation are indeed compounded with linguistic barriers, especially when it comes to interpreting multilateral agreements like the multilateral instrument or the multilateral convention by the OECD. These instruments are initially drafted in English and in French, as you said, and then require translation in numerous languages for the implementation across the world. AI can play a significant role with respect to these translations and the precision of these translations, improving the efficiency of the processes. The same holds true also for the MAP, the mutual agreement procedure, which is more of a later stage. It’s more like between the two tax authorities, competent authorities. AI-driven machine translation tools can handle complex legal terminology and context [more] effectively than traditional translation methods.

AI can assist in developing a comprehensive glossary of legal terms and phrases using the context of international taxation. This ensures consistent and precise translation across various languages, reducing ambiguities that may arise from different interpretations. Advanced AI systems also incorporate contextual understanding, allowing them to consider the nuances of legal language. Thus, by leveraging advanced translation technologies, tax authorities can overcome language barriers, promote clear communication, and ultimately resolve disputes more effectively. However, it remains essential to balance the use of AI with human expertise to ensure that the translation maintains the legal integrity and the contextual relevance.

Robert Goulder: Undeniably, there’s going to be just a lot of efficiencies here, and AI is going to be springing up in all sorts of applications, some foreseeable, some maybe less foreseeable. But what about taxpayer rights and also other statutory rights like the General Data Protection right that the European Union has? Your article mentioned the GDPR article 22 could be relevant here. Work us through some of those issues.

Alexandra Karadima: Yeah, actually, this is the most tricky question. It’s a difficult question — the field of the fundamental rights. The rise of artificial intelligence indeed raises important questions regarding taxpayer rights and the protection of individuals in the context of data privacy and automated decision-making. In Europe, frameworks like the GDPR, the General Data Protection Regulation, and also recent judicial rulings provide critical context. The GDPR, and the article 22 that you mentioned, addresses automated individual decision-making, including profiling.

Actually, it grants individuals the right not to be subject to decisions based solely on automated processing. This provision is crucial for protecting taxpayer rights as it ensures that decisions impacting individuals cannot be made solely based on AI algorithms without human oversight. If an AI system generates a decision that adversely affects an individual, for example, a tax penalty or increased tax scrutiny, article 22 mandates that the individual has the right to contest this decision and seek human intervention. This protects taxpayers from potentially erroneous or biased outcomes generated by automated systems.

There is another, actually, case law from the Court of Justice of the European Union where the Court set some limits to the tax administration with respect to the list of high-risk taxpayers. So the Court has asserted that although the taxpayers may manage taxpayers’ data in their own way in an effort to combat tax fraud, there are certain limits and those instruments should always be in line with GDPR regulations. So I think that this is also relevant with AI because a lot of AI tools actually seek to establish this list of high-risk individuals and businesses in the fight against tax fraud.

Robert Goulder: Well, there you have it. A lot to absorb, and I would encourage everyone to read the article when it becomes available in Tax Notes. Alexandra, thank you so much for joining us.

Alexandra Karadima: Thank you very much for having me once again.