Impact of AI on Tax

A year ago, almost no one was talking about generative Artificial Intelligence (AI). Now, it is hard to avoid the subject. Given that AI has been around for many years, what is it about the extra word “generative” that has caused it to become such a hot topic? In a webinar on 17 October – which you can now watch on-demand – 20:20’s Innovation's Sharon Cooke explored the question with Stephen Edginton, Chief Product and Technology Officer at Dext and tax technology guru Paul Aplin.

In a nutshell, the big step forward with generative AI is that it creates new content. Traditional AI can be trained to recognise, say, a lion by showing it multiple images of what is and is not a lion. It can be asked “is this a lion?” and will give a response based on the data it has been trained on. Generative AI can be asked to describe a lion or to generate an image of a lion (or even to write a poem about one in the style of a particular poet). In the context of tax and finance, it can be asked technical questions in plain English and will generate answers in plain English.

While some finance professionals remain unconvinced about the likely impact that generative AI will have – some firms have banned staff from using it, or imposed significant restrictions – others are already using it in earnest. Several leading law firms for example are routinely using it to analyse and check complex documents while others are using it to answer specific questions and to draft reports.

This raises significant questions. Who, for example, is responsible if the answer or the advice contained in a report produced by generative AI is wrong? The context may be new, but the question is not: it is one that professionals face every day.

Twenty-five years ago, when tax-return software started to become popular, many practitioners checked the computer-generated computations in detail; now, most simply sense-check or accept that the chance of error is extremely low. We will go through the same process of testing the quality of AI generated output to build confidence, to assess risk and to know where the limits lie.

That leads on to another question: should we automatically accept output from a “black box”, or do we need some degree of transparency to understand the results and decide whether we trust – and take responsibility – for them? Again, professional judgement is needed. For members of professional bodies signed up to PCRT, there is a clear requirement to exercise professional judgement in giving tax advice and to document the rationale behind it. Although PCRT was drafted before generative AI hit the headlines, that requirement still applies.

There are important issues to consider around client confidentiality and GDPR and firms will need robust policies and procedures in place (a policy document template is available to 20:20 Innovation members).

Inevitably, clients will use tools like Chat GPT to investigate tax questions themselves and probably to check advice they have been given. While the tools now available are far more powerful than anything clients previously had access to, being asked “are you sure, because I read an article in the newspaper that said…” or “my friend in the pub says he doesn’t pay tax on this, so….” Is familiar territory for any professional adviser. We just have to be prepared for those questions to be far more sophisticated in future.

The webinar also looked at the likely use of AI by tax authorities. A recent OECD report disclosed that 98% of tax authorities are either using or plan to use data analytics and 72% are either using or planning to use AI. One obvious use case is the employment of chatbots, to answer taxpayers’ queries without the need for a human adviser on a helpline. HMRC is not alone in facing pressure to do more with less and to move progressively away from human interaction to digital channels.

The other obvious use case is in identifying tax risk and targeting cases for enquiry.

The French tax authority has done this by using AI to look for swimming pools in aerial imagery and to compare the results with property tax returns. Additions such as swimming pools can create extra property tax charges. While there was a reported 30% error rate (with other structures being incorrectly identified as swimming pools), over 20,000 undeclared pools were discovered in the exercise and the additional tax for 2023 is estimated at around €40 million. In Brazil, the tax authority has used AI to predict taxpayer behaviour and to send out targeted intervention letters.

It is easy to see how attractive both the query resolution and tax risk targeting use cases are and it is inevitable that AI will become an increasingly familiar tool in tax administration and compliance. We – and our clients – need to be aware of that.

Generative AI is a game-changing technology. Because it learns, it is getting better every second of every day. It is already being used by major firms and by tax authorities. It will rapidly become as familiar to us as Google has. A quarter-century ago when the Google search engine hit the scene, many said that it was not a tool that you would ever use in a professional practice. The fact is, however, that many of us do so routinely, but we exercise professional judgement on the reliability of the sources it directs us to just as we always did when selecting publications in a traditional library. We will do the same with generative AI.

So, a revolutionary new technology that is likely to transform the way tax advice has arrived. But, for sure, it is one that throws up some very familiar issues.

The one thing we can be sure of is that the pace and scale of change will not slow down. The train has left the station, and we need to be on it.

Find out more about Dext via their Resource Centre page