Lessons From Applying AI in a Traditional Tax Business
By Harvey Lewis, Chief Tax Data Scientist, EY UK & IRELAND
Harvey Lewis, Chief Tax Data Scientist, EY UK & IRELAND
There is a concern amongst some quarters of the professional services sector and, for that matter, many other sectors that Artificial Intelligence – or AI – will ultimately replace jobs that are today carried out by humans. Despite having accumulated many years of training and applied experience, an increasing proportion of a typical professional’s time, today, is still spent poring over data and making essentially low value, routine decisions, which could be automated.
CIOs and Finance Directors can see the opportunity for using technology to increase efficiency and reduce costs with tools like machine learning and Robotic Process Automation (RPA). Professionals, on the other hand, are trying to figure out how – or even if – they can move up the value-chain to deal with more complex and nuanced issues that the machines cannot yet master.
What we’ve learned at EY is that technology is rarely a perfect substitute for human judgement – even in routine cases. Of course, AI can process data at lightning speed and achieve relatively high accuracy overall but what about the edge cases, or where there is insufficient good quality data for training? How do these imperfections change the way that technology is designed and implemented in a professional services environment? And what’s the impact on the people?
EY has been building AI solutions and delivering them to traditional professionals for several years. We’ve learned one crucial lesson: in areas like regulatory compliance, it’s essential to keep the human in the decision-making loop.
“In areas like regulatory compliance, it’s essential to keep the human in the decision-making loop”
Let’s look at an example: EY Alert is a machine learning application, which our tax professionals or clients use for classifying financial transactions. The classifications are used to apply the correct tax treatment, thus making sure that the appropriate amount of tax is paid.
This is an example of supervised machine learning: the tool is shown a set of ‘training data’, such as a spreadsheet, which contains many transactions and their corresponding descriptions as well as data such as supplier names, business unit names, transaction amounts and currencies – collectively known as ‘features’.
In addition, each transaction has a clearly labelled category, which EY or the client’s tax function has previously assigned.
Machine learning is the process that EY Alert then uses to identify which patterns of features consistently map to a given category. For instance, the application could learn that transactions whose descriptions mention various items, such as “stationery”, “pens” or “card”, bought from the same small set of suppliers for amounts less than £1,000 should always map to a category called “Office sundries”, whereas the same items supplied by other vendors and mentioned in conjunction with “conferences” should be classified as “Marketing costs”. The more examples shown, the clearer these patterns become until the classification ‘model’ can predict categories for new transactions using only their features.
One of the significant advantages of machine learning is its speed: what used to take the tax professional hours or even days can now be completed by an algorithm in seconds. With this working example, though, speed isn’t everything. If the training data is poor, for instance because features are missing, or they lack consistency, then the model, although perhaps still usable, may fall well short of being perfect.
Equally, if the new data to which the model is to be applied is of poor quality or is subtly but consistently different compared with the original training data then even a good model may perform badly.
In areas of regulatory compliance like tax, what matters most is getting the correct classification, not the fastest – the ends, not the means. Therefore, we must design the overall process to include both model-based decision-making and professional judgement while retaining as much of the efficiency saving as possible. To do this, experienced tax professionals first use EY Alert to train and apply a model before using additional tools in the application to help them interpret the model’s decisions and correct the output.
Compared with the way tax professionals are used to working, this technology-enabled review is fundamentally different. Firstly, it is iterative: the professional can now run the application multiple times to finesse the model and further improve performance. Secondly, instead of using a random sampling approach, which is appropriate when humans make essentially random mistakes, the professional can be guided to misclassified transactions by specific model metrics (such as ‘confidence’) and more detailed explanations of how different features in the input contributed to the output.
AI is changing the role of and skills required by the tax professional. Experienced judgement is still needed, though, particularly to support edge cases that the machine struggles to classify. And, in addition, new technology skills and knowledge are needed to use AI tools and successfully interpret their output. Today, we are extending these principles with new active learning methods for monitoring changes to tax legislation around the world.
We’ve discovered that technology is not simply a substitute for people; technologies like AI need people to deal with the complexities of the business world.