Tax Applications

Al can be utilized for a variety of tax functions. For example, Al can be used for extracting critical data from tax documents, classifying sensitive transactions, identifying possible deductions and tax credits, comparing pricing structures for transfer pricing, and tax forecasting (Van Volkenburgh, 2019).

Technology services firm CrowdReason provides software called MetaTaskerPT to classify documents by asking human workers to

Applications of Al in Accounting 29 identify the document type. The software recognizes a correct answer when there is consensus by three separate people. A similar process is used for defining the taxonomy of a document by asking questions such as: Is there an account number? How many payments on the tax bill? When is the due date? Much like other ML algorithms, the more data MetaTaskerPT analyzes, the better able it is to recognize documents.

Once the system understands the document, it can extract key information and automatically input it into a software system (Van Volkenburgh, 2019). Al can be used for classifying tax-sensitive transactions that generally take a lot of time for human staff to accomplish. For example, algorithms can be used to identify assets that are incorrectly booked to accounts based on historical classifications made by staff. By using machine learning, the number of manual reviews can be reduced from 50,000 transactions per year to less than 300 (Van Volkenburgh, 2019).

Al can also be used to review tax notices received from various tax jurisdictions to determine if they are informational or require action. ML can be used by tax agencies for predictive modeling to identify potential fraud or tax avoidance. Given how lengthy and complicated the Internal Revenue Code is, Al is well suited for identifying possible deductions and tax credits. For example, H&R Block uses IBM’s Watson system as part of its questionnaire engine.

Deloitte uses Al for a variety of tax applications. For example, the firm uses natural language generators in its tax practice to provide targeted financial advice (Nickerson, 2019). It also uses a supervised machine learning tool with NLP to automatically extract clauses from legal documents such as contracts and deeds. It can review high volumes of trust agreements to automatically classify the type of trust for tax purposes, saving tax professionals a multitude of work hours each year.

Furthermore, Deloitte has a global practice that specializes in recovering refunds related to indirect taxes, such as sales taxes or value-added tax (VAT). Due to the complexity of tax laws in various jurisdictions and the vast volume of data, indirect tax recovery is a challenge. Deloitte uses CognitiveTax Insight™ (CogTax) for indirect tax recovery. CogTax can analyze the full population of accounts payable transactions by applying optical character recognition (OCR), machine learning algorithms, and analytics to identify overpayments and reduce the potential for over or underpayments in the future (Deloitte, 2019).

EY developed a tool called the capital allowances automated review tool (CAART), which uses machine learning to assist with identifying potential allowances (i.e., tax relief) that are often overlooked (Duffy, 2019). The CAART tool analyzes large volumes of fixed assetcost data (e.g., buildings under construction) at rates much faster than humans to assign the correct tax treatment. This analysis requires an understanding of tax legislation, case law, and construction terminology. The CAART tool uses ML to learn the relationship between text descriptions and the corresponding tax treatment. The raw data comes from the client’s accounting system or from external cost consultants and is then uploaded into CAART. This tool then predicts the tax treatment based on what it learned from the training data set (and this only takes a matter of seconds).

After the output is downloaded, CAART provides a probability score of citing how confident the system is that it predicted the correct answer. If necessary, corrections are made to the output, and once approved, the results can be included in the training data set so that the system learns and will have increased accuracy for the next case. The tool also has the capability and flexibility to be programmed with additional rules to accommodate legislation changes and “what if” scenarios.

The Wall Street Journal recently reported that governments around the world are increasingly using machine learning and data analytics to identify tax evaders, respond to inquiries, and generally become more efficient (Rubin, 2020). Brazil is using these technologies to detect anomalies, which has resulted in a 30% increase in inspections. Canada will be launching Charlie the Chatbot to respond to taxpayers’ inquiries. In the United States, the Internal Revenue Service is using Al to examine notes recorded by IRS agents when responding to queries. The IRS also uses Al to determine the optimal combination of notices and contacts that are most likely to result in the taxpayer settling the amount owed. The IRS analyzes data from inside and outside the agency to identify thousands of high-income individuals who did not file returns. The article notes that the IRS criminal investigation unit uses Palantir Technologies, a data-mining firm, to identify potentially fraudulent cases.