Updating taxation: Can AI and data analysis help with HMRC’s human resources problem?

Generative AI and data scientists

To address some of the challenges HMRC is currently facing, it has been recruiting data scientists to develop generative artificial intelligence (GenAI) tools that will assist tax advisors with their workload.

“We recognise that GenAI has enormous potential, and we’re exploring a range of possible ways to use it, while being careful to manage the risks to public trust from new and rapidly evolving technology,” says an HMRC spokesperson.

“Where AI use could impact our customers, we ensure the result is explainable, has a human in the loop, and is compliant with data protection, security and ethical standards.”

There are two ways that generative AI could be used to assist HMRC. The first is through automation by using GenAI to automate the simpler tasks of tax advisors, such as basic data entry, so they can focus their time and energy on the more challenging aspects of their role.

We recognise that GenAI has enormous potential, and we’re exploring a range of possible ways to use it, while being careful to manage the risks to public trust from new and rapidly evolving technology

“HMRC just do not have the people to answer the phones. No one is saying they’re slacking off, but they just do not have the resources to do it,” says Chris Thorpe, a chartered tax advisor and technical officer for the Chartered Institute of Taxation. 

“If you increasingly rely on automation, you’re going to need the people to pick up the phone because a lot of the time the software won’t work or people won’t be able to access it, and they just need to speak to someone. By relying on automation, you need to have the support there at the same time, and that’s unfortunately where they’ve just not been able to keep up.”

AI could also be used to review tax self-assessment records as they are submitted and scan for any anomalies. These discrepancies might be simple mistakes, but they could also be signs of potentially fraudulent practices. Anomalies could occur for any number of reasons, as real life can be inherently chaotic and unexpected events will happen. Consequently, a person’s or company’s finances can be influenced by any number of external factors that could not be predicted, and therefore appear as an anomaly in their tax returns.

HMRC has a historical archive of all the UK’s tax records, which it could use for training AI systems. However, many of these records are stored in legacy systems, which means the hard data for training the AI would be challenging to extract and format in a usable way.

“Old systems do not talk well together,” says Alison Kerrey, a partner of Moore Kingston Smith and chair of the joint Chartered Institute of Taxation and Association of Taxation Technicians Digitalisation and Agent Services Committee. “It might be accurate data, but it might not be in the right place or talking to the right thing for you to be able to use the technology effectively.”

The complex challenges faced by GenAI

Tax law is also incredibly complex. There are baseline laws that act as a foundation, but there is also a lot of case law that can be open to interpretation.  This is compounded by the fact that tax law is constantly evolving. 

With every Budget and Autumn Statement the government announces, the UK’s approach to taxation is changed to meet government policy. Therefore, what was acceptable a few years ago may no longer be viable.

“There’s a lot of legislation, but also there’s a lot of interpretation that you need to add into that,” says Kerrey. “The rules will take you so far in some cases, but then you have to add maybe 40 years of case law and different interpretations, as well as taking things on a balance, and that doesn’t automate very well. It’s quite difficult to build tools to deal with that.”

It is worth noting that although practitioners can be chartered, tax law is not a regulated industry. Therefore, anyone can call themselves a tax advisor, provided they do not falsely claim to be a chartered member of the Institute of Chartered Accountants in England and Wales (ICAEW) or Institute of Chartered Accountants in Scotland (ICAS).

Relying on GenAI for advice on tax law is discouraged, due to the way it can sometimes provide convincing answers that are factually incorrect. If generative AI is unable to find an answer, it may combine similar historical examples to hallucinate an answer that looks correct but has no basis in fact.

“There was a lawyer who used ChatGPT to produce a load of cases in court, and it was discovered that the software had completely fabricated them, and the judge threw the whole case out,” says Thorpe. “That’s a fairly extreme example, but it’s an example of practitioners potentially not looking at it properly and assuming all the figures are right.”

AI will be better than anything we’ve had before, but at the end of the day it will just be a tool that will need to be managed properly. There will need to be some regulation to hold practitioners to account and to ensure they use AI responsibly

The unpredictable nature of GenAI means human oversight will always be required to ensure the information provided is fair and accurate. Although it was not generative AI, the lack of oversight of the Post Office Horizon IT software, combined with insufficient critical analysis of the data it provided, led to hundreds of subpostmasters being charged with theft, fraud and false accounting.

“We’ve already seen what happens, with the whole Horizon and Post Office thing, when you just slavishly rely on computers,” says Thorpe. “As long as people accept that mistakes will be made on both sides, and that you can address those mistakes and people will not be adversely affected, then I do not think there will be a problem.”

Furthermore, generative AI works best when using hard data and absolutes, but it struggles with the nuances that humans can bring to an equation. For example, people can be self-employed or be an employee, but they can also be a part-time employee who also owns a small business – instances like this can cause GenAI to struggle to rationalise the situation.

Another issue is that some companies are not digitally connected. They may operate in remote locations where the internet is unavailable, or simply not wish to use digital tools. For example, there is a timber merchant near Derby who uses a slate and chalk for calculating payments and only accepts cash.

Automating tasks, but with human oversight

The important thing is that the careful use of generative AI can reduce workload through automating simple tasks. This will allow tax advisors to focus on the complex elements of tax law and fringe cases that AI would struggle with.

“AI will be better than anything we’ve had before, but at the end of the day it will just be a tool that will need to be managed properly,” says Thorpe.  “There will need to be some regulation to hold practitioners to account and to ensure they use AI responsibly.”

Ultimately, HMRC has no intention of replacing tax advisors with generative AI. People with experience and knowledge of the UK’s tax laws will always be needed to resolve queries, whether it is by telephone, email or instant messenger. GenAI can help with the demands placed on HMRC by automating simpler tasks and providing basic anomaly detection, but a human element will always be required for maintaining oversight.

“AI can miss complexities and sometimes it can return things that are wrong,” says Kerrey. “Any future use of AI has to be able to work out how to get around that hallucination.”