Bringing tax advisory services to a higher level: legal research with tax-i™ Bookmark has been added
Bringing tax advisory services to a higher level: legal research with tax-i™
AI-driven research tool for legal rulings pertaining to tax legislation
The application of AI to legislation is significantly increasing throughout the world. Deloitte has established a key milestone with tax-i™: a virtual tax legislation research assistant. tax-i™ provides insight into legal documents that facilitates more efficient research, statistical substantiation and the evaluation of tax cases. Seven million legal documents have been read into tax-i™ and this number is growing rapidly. ‘In the future, everyone will be able to conduct online searches of court cases and obtain guidance,’ Marc Derksen, tax-i™ Operational Lead, foresees.
Two years ago, tax-i™ began as a start-up under the wings of Deloitte’s Tax & Legal Indirect Tax branch at the initiative of Michel Schrauwen, Frank Nan, Roderick Lucas and Marc Derksen. They saw lots of opportunities in the field of legal technology and artificial intelligence (AI) to make legal advice more efficient. They read in legal databanks and then classified the court cases for analysis. Their experiments focused on predicting the outcome of court cases. ‘We succeeded in doing this with a certain degree of accuracy,’ says Derksen. ‘In addition, we found out which information is needed to be able to make such predictions.’
The experiment, named tax-i™, began with technology as the starting point and focused on European tax legislation. The first step was to read in all tax cases of the European Court of Justice. Because EU member states are obliged to publish their rulings and make them accessible to third parties, a great deal of data is available. This made it possible for tax-i™ to use this data to predict the probable outcome of a ruling of the European Court of Justice on the basis of the case details put in by the tool’s user. tax-i™ makes these predictions within a few seconds using a machine learning algorithm that has been trained to recognise patterns in tax cases and create an advanced model. Using network visualisations, summaries, extensive search functions, statistical analyses and deep learning, tax-i™ subsequently provides legal insight.
This way, tax-i™ automates the legal preparatory work, which is largely of a repetitive nature. It includes tasks such as searching for relevant jurisprudence, analysing court rulings and evaluating whether a tax case is probably going to be successful. tax-i™ is accessible to any Deloitte employee throughout the world, as well as to clients and professionals doing their own legal research. However, interpretation remains imperative, says Derksen. ‘To safeguard certain quality criteria, it is important to involve a tax advisor capable of maintaining an overview of the whole. Even with a self-driving car, you still need to keep your hands on the steering wheel.’
AI and human collaboration
‘Ever since Deep Blue defeated Kasparov in a chess game in 1996, there have been suggestions that AI is going to replace human labour,’ Derksen says. ‘But Kasparov saw things differently: instead he foresaw a future with collaboration between AI and humans. He started up Centaur, a new form of chess in which human players use AI to predict the outcomes on the game board. AI does not provide a human line of argumentation. For example, a prediction must also be traceable so that a human can see why a certain outcome is being suggested. For clients it is essential that any advice can be substantiated and that is the work of humans. However, AI can provide excellent support in this respect. tax-i™ supports lawyers in finding relevant arguments, in substantiating advice and with statistical analyses. Moreover, the latter is most certainly important in countries such as the United States, where we started reading in legal documents as well. The jury trial system, for example, makes it interesting to trace back political preferences and profiles. The percentage of rulings provided by a specific judge can also be an interesting characteristic in countries with common law.’
Local conditions in each country demand local collaboration. ‘Within Deloitte we look for a legal subject-matter expert for each country, who can help us gain deeper insights into local legislation, specific interests and customs. For example, the case structure in Britain is different from that in the Netherlands or France. Each country will have its own team, which will carry out analyses and make predictions specifically for that country. The core team that hosts the platform and performs generic tasks is located in the Netherlands. The number of job openings is increasing’, Derksen explains. He expects tax-i™ to experience significant growth: ‘In the past, people would immediately visit a doctor in case of physical complaints. Today, we can look up almost all information online before visiting the GP. Providing advice has become more important than informing. I foresee the same thing for legal research. High-quality advisory work and contact with clients is going to gain ground on research work in terms of time spent.’
Recognising and avoiding bias
Furthermore, tax-i™ helps remove human bias. Derksen observes that bias can occur especially among the more experienced specialists. ‘A VAT partner would normally ask an analyst to research several matters of which he is top-of-mind aware that they are relevant. But his experience does not cover all aspects, which is why he tends to only consider the well-known aspects. However, if you use tax-i™, you consider all possible aspects that can be relevant to your research. In other words, foreign legal areas, dated cases or simply cases that, while they only briefly deal with the particular topic, are nevertheless decisive.’
There is bias on the AI side as well. That bias occurs because AI is not all that good at contextual thinking. AI is very good at simple repetitive tasks in a single knowledge domain, but when domains simultaneously converge, AI misses the input required to correctly interpret the context. ‘The dataset used to train the algorithm itself can also contain a bias’, Derksen says. ‘For example, if a dataset exclusively contains data about a specific demographic, the algorithm can only make a prediction based on that demographic. This can result in undesirable situations when people are not aware of this. Through means of checks and balances, we are taking this into consideration throughout the entire tax-i™ development process. Yet another reason why AI cannot do without human intervention.’
tax-i™: modelling an advanced decision model
tax-i™ fundamentally creates a data model that reflects the status of the law at a given point in time. This starts by reading in the data, which is subsequently subjected to advanced analytical methods. Derksen gives a specific example: ‘Suppose you have 100 court rulings relating to VAT exemptions of which 20 resulted in a VAT exemption being applicable, while 80 result in the inapplicability of a VAT exemption. tax-i™ creates a model on the basis of the specific characteristics of the textual content of these cases. With the help of advanced statistical models, a decision tree is created that can be used to predict the outcome of new cases.’
Aside from its predictions, tax-i™ also has a number of supporting research functionalities. For example, using an interactive line diagram, tax-i™ visually displays how different cases are related. The size of a dot indicates the relevance of the case based on the number of references. In addition, tax-i™ provides an AI-generated summary of all cases. ‘The problem with legal research is that the number of results is often considerable. This leads to great pressure for lawyers. By mapping out all of the relevant cases visually with AI, you can immediately see which rulings and arguments can be used to create the underlying substantiation. This way, tax-i™ makes legal research many times more complete and efficient.’
Roadmap for tax-i™
tax-i™ currently contains more than seven million legal documents, including jurisprudence and legislation for the European Union, the Netherlands, the United Kingdom, the United States, Germany and Spain. This number will continue to increase, Derksen suggests. ‘At the moment, tax-i™ still focuses on tax matters, but as platform we are seeing opportunities to apply it to all forms of jurisprudence. Over the long term, tax-i™ is intended to become accessible to all tax and legal advisors in the world. In this context, we are now working on a legal translation tool to enable us to work with data in multiple languages.’
‘Despite the fact that we are experimenting a great deal with the opportunities provided by modern technology, customer perception will be leading in the tool’s further development. This is why we are testing the various value propositions of tax-i™ with clients, so that we maintain our focus on what constitutes added value for them,’ Derksen explains.
In the meantime, Deloitte has established a reputation as AI expert in the academic world. The tax-i™ analyses have been evaluated by professors associated with two Dutch universities. In addition, Deloitte’s tax-i™ team is also facilitating the very first master’s tax technology degree at a university. ‘It is important that the tax advisors of the future will be able to find their way in a tax world that will be increasingly dominated by technology,’ Derksen concludes.