For the last few years, the world has been captivated by the power of generative AI and the seemingly limitless business potential it presents. But while much of the hype generated by AI tools like ChatGPT is fully justified, these technologies still have a long way to go before their capabilities catch up with the ambitions of many business leaders.
Media coverage of misadventures with AI involving the likes of Google and Air Canada show that the implementationand rollout of AI solutions is far from straightforward, but getting it right remains absolutely critical to overall success.
A round-the-clock resource
So, for the tax and accounting industry, what is the best lens to view GenAI through in its current stage? I think it would be helpful for most organisations to think of it as the ‘infinite intern’ – i.e. a round-the-clock resource that is happy to do anything thrown its way, at scale, and has the ability to turn around jobs extremely quickly albeit being prone to frequent errors and therefore requiring close monitoring.
When GenAI is approached with this mindset, the tasks it is best suited to start to become much more apparent. For instance, while it shouldn’t necessarily be trusted with one-off, high value tasks, it excels at automating routine tasks and reconciling large amounts of data that would otherwise take many hours or even days to get through. Doing so not only slashes the time taken to complete such tasks, but it also frees up tax professionals to focus on strategic activities that are much more valuable to the organisation as a whole. It is no surprise that finance companies are adopting GenAI applications at pace for this purpose.
Objective and trainable
Another great benefit of GenAI, when compared to human employees, is its consistency and objectivity towards tasks, regardless of the time or day of the week. A famous research study by Ben Gurion University in Israel and Columbia University once found that judges – whose job it is to remain impartial – granted 65 percent of requests they heard at the beginning of the day’s session and almost none at the end. Right after a food break, approvals jumped back to 65 percent again. Jonathan Levav, associate professor of business at Columbia, said that the judges could just be grumpy from hunger, but they probably also suffer from mental fatigue. AI suffers from neither of these things.
Similarly, studies have found that humans are less productive and make more mistakes on Fridays, compared to other days of the week. Again, AI solutions are not affected in this way, meaning users need not worry about where and when tasks are set.
GenAI also responds to training in the same way that interns and employees do – the more you provide, the better it will become at its assigned tasks. Comprehensively training these tools with high quality data and information over an extended period of time will enable them to start producing valuable insights that might otherwise be missed. This can then be fed into real-time reporting and decision-making, with AI-generated insights allowing organisations to make more agile, informed decisions as a result.
AI technology is still far from perfect
GenAI is developing quickly but, that being said, the current generation of GenAI tools shouldn’t be left to operate on their own without regular monitoring or management. In the absence of frequent course corrections, small, simple errors can quickly become much bigger ones, causing major problems further down the line.
Accuracy also remains a concern, which is why they are best kept away from certain front-line tasks. Granted, each of the most popular GenAI tools use some form of disclaimer to alert users to the likelihood of errors, but some of the well-documented accuracy problems appear to be getting worse over time, not better. According to research from Stanford and UC Berkeley, for example, when tested on a range of tasks, ChatGPT’s performance in some areas had become “substantially worse over time.”
Even as accuracy will increase as general large language models become more advanced in the future, there is no scenario where HMRC will consider AI-powered tax returns that are, for example, 98% correct as ‘good enough’. Similarly, what impact could an error rate of 2% or more have on individuals or corporations relying on generative AI to represent their financial affairs accurately? This, and other similar questions, remain unanswered at the time of writing.
While some would like us to believe the days of fully automated, AI-powered tax returns are just around the corner, the truth is that such capabilities are likely still many years away. But that doesn’t mean GenAI, in its current state, can’t already play an invaluable role in optimising tax and accountancy processes. In fact, it’s already doing so in many instances. The key lies in playing to its current strengths as an infinite intern, rather than elevating it to positions that it isn’t yet ready for within the organisation.
Russell Gammon, Chief Solutions Officer at Tax Systems