Banks must be trusted to hold our money and transfer it electronically, but they lend it out in waves of credit bubbles with barely a fraction in reserve. We have to trust them with our privacy, trust them not to let identity thieves drain our accounts.”
Memory Machines - Data centers have proliferated across Ireland, at great cost
MATT TAIBBI: If AI Thinks George Washington is a Black Woman, Why Are We Letting it Pick Bomb Targets?
AI auditors let the ATO find millions in unpaid tax and super
Paul SmithTechnology editor
The Australian Taxation Office’s deputy commissioner of smarter data Marek Rucinski says he can measure the return on his team’s efforts to deploy artificial intelligence across the Australian economy to the tune of hundreds of millions of dollars.
Few private company or public sector executives are as open as Rucinski in discussing the financial spoils available from increasingly smart software, and he says strong controls are needed to make sure humans remain a crucial part of the system, as the bots get more powerful.
Rucinski wowed attendees at The Australian Financial Review’s Government Services Summit in 2023, when he explained how AI was a central part of the ATO’s efforts in trying to recover almost $45 billion in unpaid taxesowed by Australians. He says efforts have continued apace since, as he balances growing data and digital capabilities with a measured approach to ensure tech smarts are deployed in a safe and responsible manner.
“In the ATO, we would only use AI where it is legal and ethical to do so, and where humans are ultimately responsible for the decision-making that impacts clients,” Rucinski says.
“We keep abreast of emerging data and analytics technologies and techniques to identify those that may potentially provide value in administrating the tax, super and registry systems end-to-end.
“The returns are indirect and we expect it to be demonstrated through increased operational efficiencies, client satisfaction and revenue outcomes, such as through the early detection of fraud.”
Measurable returns
Whereas some organisations’ measurement of the benefits of AI adoption are a little rubbery and relate to the perceptions of how many hours of work is saved by automation, the ATO can measure returns in dollars and cents.
Rucinski says natural language processing has been used to help turn vast quantities of unstructured data into valuable data feeds for intelligence analysts, by highlighting the most value documents to review and even where and what to look for in those documents.
He cites the example of the leak of 11.5 million files from Panamanian law firm Mossack Fonseca in 2016 – popularly known as the Panama Papers – as a notable example. By using its AI systems to analyse the vast troves of documents, the ATO has dug into a goldmine.
With figures last updated at the end of 2023, its bots have raised more than $242 million in liabilities, collected more than $60 million in cash and finalised more than 535 audits and reviews.
Rucinski says graph analytics – which uses tools and algorithms to analyse the relationships between data points in a graph database – has been used by the joint-agency Phoenix Taskforce to identify fraudulent pre-insolvency advisers and unscrupulous liquidators who may be promoting and facilitating illegal phoenixing activity.
This taskforce raised $304 million in liabilities and $108 million in cash collections in 2022-23.
Ideas for deployment of AI in the ATO come from both “top-down” and “bottom-up” approaches, Rucinski says. This means some are identified through strategic planning processes, but others are grassroots ideas that come from agency staff who notice a business need.
Super uses
He says one particularly successful initiative has been in using stratified sampling to create AI models that have proved highly capable at identifying employers who are most at risk of underpaying super contributions.
“By targeting the highest risk employers, we’re able to more efficiently use resources to ensure employers are paying the correct amount of super for their employees,” Rucinski says.
“The accuracy of our models also means we are far less likely to contact or undertake reviews of employers doing the right thing, saving those businesses time and resources.”
He says that, since the 2018 financial year, these models have helped raise around $295 million in liabilities and have a success rate of around 90 per cent in detecting underpayment.
Despite his heavy use of AI, Rucinski remains a staunch advocate for strong human oversight and accountability for the actions of AI algorithms. He says the ATO has published data ethics principles, and that all of its staff are trained on them before they use AI tools.
“As AI becomes more powerful, the ability of humans to explain the models diminishes. This makes it more imperative to ensure there is human oversight and/or decision-making,” Rucinski says.
“In the ATO, we ensure humans are always in the loop and we provide clients with a right of review of decisions.”
Gain insights into the week’s biggest tech stories, deals and trends. Sign up to The Download newsletter.
