Bridging the Gap: Artificial Intelligence for Tax Functions
Artificial Intelligence simplified
Simply put, AI is software that learns from large data set and rules, and based on that, it can perform tasks you ask it to do.
Picturing AI as multi-layered, like rings of an onion, consisting of several different types of AI components working together to create a task that we are instructing it to do. If we look at the most widespread AI components and structure them different in the layers of the “onion” on how they perform tasks, the first outer layer of the “onion”, or AI component, is Machine Learning (ML) that operate on rules, takes decisions and makes predictions based on the data given (an advanced decision tree algorithm). The next layer towards the middle is Deep Learning (DL) that works like a human brain that thrives on data (for tax could be tax legislation, court cases and other rulings) and learns from examples and patterns and it gets better with practice. In the centre of the “onion”, we have Generative AI (GenAI), which is the language models we know as ChatGPT, CoPilot, Gemini, or the European version, Le Chat. GenAI simply put, is like a super smart copycat. It produces new content based on patterns learned from large data sets from the Internet up to a given date. From this it refers to algorithms that can then create content, including text, images, video, simulations, code, audio, and more. Tax related examples could be generating a first draft of a tax memorandum, using a “what-if” tax projection based on specific inputs, or reviewing and comparing a transfer pricing documentation, tax policy statements, scanning procurement contracts for tax related items, or reviewing intercompany agreements and intercompany transactions at large scale. Le Chat might worth considering from a data security perspective if the company policy is to keep data in the Europe.
If we are to use AI for tax work, for example, to do a control on the tax determination, if we set up an AI component (could be ML) to review a set of invoice data for correct tax determination. Prior to the review we have been using a set of data to train the AI on how to allocate correct VAT codes to a certain type of transaction, whether that is purchase or sales (external or internal). If the AI, in reviewing the invoices, discover a transaction that it does not recognise (from the data, rules and instructions we have given it) AI will “park it” and we can review the transaction and determine the right VAT code, and we then update the data set the AI to use so next time it encounters the same combination it knows how to tax determine this type of transaction. The finding could be a need to correct the ERP system or the shared service colleagues to determine the transaction correct from the start.
Another example of how we might need multiple AI components when the task is to do the annual review and update a Transfer Pricing (TP) documentation. The “simple” steps to do this could be 1) we retrieve a document from the document archive system or a SharePoint folder, 2) we review the documentation and compare with new information or changes in the business/organisation, 3) we update the document with the new company data we have available, and 4) we save the updated TP documentation to the archive with a new version number. In the world of AI, this is several different tasks that might require different AI components to be orchestrated correctly. It surely will also require tax professional intervention to ensure that AI has updated the Transfer Pricing documentation correctly. So here we might need a combination of AI components to get the file, retrieve new business data, and external requirements perhaps, get it to review and suggest updates to the document that is reviewed by a TP expert before the final document is stored back in the achieve. If we want to automate this process, we will need to consider Robotic Process Automation (RPA) for getting the documentation and relevant updated company information before running AI to review and update the documentation.
As mentioned earlier when using GenAI and the output does not produce the expected result for example when asked to interpret a new tax legislation, there can of course be several reasons. A couple of examples could be the GenAI tool used; firstly, is it a general GenAI tool such as CoPilot or Le Chat they might not be finetuned with specific tax context in mind or have access to the latest tax legislation, court cases or rulings. Secondly, what do we ask GenAI, how do we specify the task. Thirdly, do we supply enough data, and the right data, are we specific enough in our request for it to perform the task we expect. We must not assume it knows, what we know! The prompt needs to be Clear, Accurate, Relevant and Specific.
Further, in setting expectations for what AI can be utilized for, we are still not in a place where we can rely fully on AI to deliver the quality of that of a tax professional. We should regard it as a support tool that can deliver on a task or part of a task, such as drafting a memo, reviewing or analysing numbers, or performing a tax control. We will always need to review the result before it is used or submitted. An interesting case of where AI might have been trusted a bit too much, can be found in a LinkedIn article where AI was used to deliver a report to the Australian Government.
Summing up, AI should not be regarded as one tool or one solution. It consists of multiple components or layers, each good at performing (a) specific task(s). If we, as tax professionals, request AI to support our job / work, we need to prepare through reviewing and documenting the specific process / task, analyse and structure the different steps into categories of actions and identify the different AI components, needed to carry out steps and not least ensure we, collect relevant information to give it the best possible foundation for a good result. We also need to consider what AI component that is best for the tax work we are asking it to do – does it need to have specific tax context knowledge or is it just a “moving” data job? And finally, remember we cannot rely on AI (yet) to replace a tax professional, but it can be a perfect buddy to support us.