Transforming tax functions with Artificial intelligence
It’s been hard to keep up with the hype surrounding artificial intelligence (AI) in recent months. Many of us are feeling the pressure to get to grips with AI.
As a tax leader, you will need to embrace it – and fast. Because in its latest form, AI has the power to completely transform tax functions.
By ‘its latest form’, I mean generative AI: the technology behind ChatGPT, which has pushed the hype surrounding AI into overdrive. As the name suggests, generative AI can generate content, using algorithms that have been ‘trained’ on vast quantities of existing digital content.
Like in so many fields, generative AI has many potential uses in tax functions. So much so, that one academic paper recently heralded the ‘rise of the robotic tax analyst’.
AI models and tax applications
To understand how generative AI might impact tax operations, let’s look briefly at the three categories that impact the tax function:
1. Rule based automation
Rule based automation solutions include applications like robotic process automation, decision trees, dashboards and automated reporting. These work well for routine tasks with a known dataset and single, defined decision to get to – and therefore a simple logic and clear route to the desired outcome.
For example, applying the right VAT rate to a given transaction. The data and algorithm are fixed, and you know the rules to follow to get to the answer you need: where your business is located, where the buyer is, the price, and so on.
2. Discriminative machine learning models
Discriminative machine learning model solutions include machine learning. They can be trained to take on elements of decision-making. As with automation, the relevant data is fixed; but some exploration is needed to work out the rules to follow in order to find the answer.
That’s useful when categorising assets for tax purposes that the business has purchased, and identifying exceptions. Data on millions of previous decisions can form the training data for a machine-learning tool. By inputting huge quantities of ‘right answers’, it can be trained to work out how to categorise future transactions.
3. Large language models
Large language models such as ChatGPT can get to the right solution when none of the data, rules or decisions are clearly defined. They rapidly process enormous numbers of data points (from multiple sources) and possible outcomes, then infer the rules from this training data.
From a tax perspective, image-classification models could classify waste materials according to the environmental regulations that apply to them – from just a simple digital image of the produce in question.