Follow the money: Australian Taxation Office sets the pace for government adoption of AI

Powered By

The generative query would go something like this: “the top 10 recommendations for tax expenses I can claim in Australia and how to prove them”.

That’s the rather awkward scenario the Australian Taxation Office (ATO), arguably one of the nation’s leading data science trailblazers, is facing as it once again cautions taxpayers not to make up or falsify claims. It’s only a matter of when, not if, they are caught.

Two years ago, the nation’s revenue agency was one of the most sought-after speakers on Australia’s packed tech conference circuit for its exemplars of how it was using advanced analytics, pattern recognition, machine learning and robot customer service avatars to guide customers as needed.

None of that has changed, but what has changed is that generative AI (ChatGPT and competitors) are now rapidly mining, ingesting and regurgitating any and all the publicly facing data and content they can grab to serve it back up to a transfixed public marvelling at the novelty of machine-driven labour reduction.

The very hungry caterpillar

Generative AI is both a blessing and a curse for agencies like the ATO.

In enterprise AI, data sources and data hygiene are tightly controlled within a confined environment to produce results that are trusted and tested thoroughly. Think chatbots, next-best automated decisioning to provide appropriate recommendations or suggestions.

Commercial generative AI is the opposite.

Left unregulated and untrained, self-learning machines can grow up ingesting all the worst information, clickbait and conspiracy in the same way that any number of people have self-validated their theories on world government, secret networks and a general withholding of the truth. And let’s face it, not everyone likes paying tax.

But generative AI is here and growing exponentially in a repeat of the technological wild-west scenario that accompanied the first dotcom boom, and then the explosion in social media that left broadly left governments as followers rather than leaders.

Engaging with such platforms is fraught with risk because agencies can potentially become beholden and dangerously dependent on them, like when Facebook started unplugging government content when it was first legally asked to start paying for news.

The march of popularist technologies like social media has for at least two decades directly challenged the established communication channels of government. Both policymakers and policy practitioners are acknowledged as being slower, indeed reactive to these phenomena.

The price of trust

The most recent example of this is the wave of so-called TikTok-fuelled GST frauds that culminated in Operation Protego.

The ATO was forced to partially de-automate GST refunds because of wholesale abuse by way of first-person tax fraud by the claiming of false GST expenses to manufacture fraudulent tax refunds that could always be traced back to the claimant, but are uneconomical to individually pursue.

Billions have gone out, but it’s likely the cost of recovery is higher than the amount paid because those claiming fake refunds are of few means or assets to seize in reprisal. It’s a game of Whac-A-Mole on an industrial scale.

The sad but real scenario is that many people thought tax fraud was worth the risk, and thus gamed the system, betting the government will not throw good money after bad if it can help it.

Call it a distributed fraud attack, call it a mass movement scam, but the fact is that the agency charged with the collection of revenue found itself blindsided by a foreign (arguably hostile) comms channel and has subsequently seen its take and efficiency go down.

This single attack unwound years of work educating, building trust, and fostering a basic understanding in migrant communities of how tax works, its benefits, and obligations.

The Protego phenomenon was propelled by social media rather than AI, but the two are starting to blend and blur as established Big Tech players muscle their way into the AI bazaar.

There is now a real question as to whether agencies like the ATO need a ‘rapid reaction’ or ‘flying squad’ to deal with such threats to public revenue leakage given the size of the hit the ATO sustained and cannot economically recover.

Rebuilding integrity

The irony is that the ATO has for the last decade or so made it far simpler, more efficient and economically sustainable for businesses and individuals to stay on the right side of the law without a major expenditure of effort and labour.

At the hearings of the Royal Commission into the Robodebt Scheme, taxation officers senior and middling easily defended their own agency’s integrity as others faltered.

Questions were asked. Challenges were made. The efficacy and operation of robodebt was vigorously contested.

In simple terms, Tax ‘called out’ the issue only to be told, by politicians, to get with the program.  And so it was.

Fortunately, Tax saw beyond the myopic, regressive and utterly cruel robodebt scheme. Internally, it is known that the ATO challenged the policy premise of debt manufacturing to create savings, and with good reason.

The first is public trust. It’s very hard to explain why bogus retrospective debts can be applied by one agency and not another. A LOT of people owe Tax money because the agency is statutorily tasked with collection.

There is a view, and not an isolated one, that the abusive nature of robodebt eroded social norms and boundaries that ricocheted and hit Tax. When the government acts unethically and illegally, it’s hardly surprising sections of the community apply the same rules.

This propensity to fraud and fiddling in turn creates friction that in turn reduces efficiency and threatens to increase the Tax Gap, this difference between what is estimated to be owed in total and what is collected.

Before COVID hit, Australia had an enviable tax efficiency rate of 93% thanks in large part to the level of automation, clean data, solid architecture and years of public education and careful nudging.

Lean and seen

The efficiency rate will necessarily have dropped post-COVID. There are still thousands of zombie businesses sitting on deferred and uncollected debts that are yet to wash through the tax system and to do it suddenly risks systemic shock.

It will be tightly held information, but the ATO will have sharp visibility of this risk thanks to the billions it has invested in analytics and modelling capability over the last two decades that has been neatly integrated with its customer-facing front-end systems.

As big US technology vendors hype up the capabilities of their proprietary AI, Tax has been quietly, methodically and systematically bedding down and optimising its own estate so it is not blindsided by persistent waves of tech mega trends that come and go like the seasons.

As constant AI-powered hallucinations and snafus have demonstrated – think Amazon’s infamous men-only recruitment tool, or the unfortunate incident of Tay, Microsoft’s tool that soon resembled a deranged cooker after gorging on public social media feeds — a lot of public generative AI can go wrong, and when it does it’s all out in the open.

It’s against this backdrop that Tax has to keep its own systems evolving but also prepare for the inevitable ingestion of its own public-facing data into AI bots and software and try and turn those interactions into something positive.

Discreet but determined

A lot of the work Tax does using cutting-edge AI remains necessarily out of public view, not least because sophisticated tax avoiders — and its no-secret tech platforms themselves are under scrutiny in multiple jurisdictions for their aggressive tax minimisation behaviour — are always looking for means or technology to counter detection or to find a way through the net.

The breach of such necessary confidentiality, indeed the abuse of public trust, is at the heart of the enduring PwC scandal that has cast a light on the ethics and tactics of the Big Four, with Tax’s sin-binning of miscreants having kicked off the great introspection.

What is coming into view though, even if scarcely reported, are the years of work to create a solid and future-ready data architecture through the ATO’s Smarter Data program that’s regarded as one of the best training grounds in the country for gaining rock-solid skills and reasoning that underpin critical systems.

“Our Smarter Data program works with one of the largest data collections in the Australian government. Our data is a critical asset. It presents complex challenges for how we manage and use it for the benefit of staff, clients and partner agencies,” Tax says.

Its key application is around better client service, early intervention to stop problems like ballooning tax debts and a stated goal of prevention rather than correction. That last goal is the polar opposite of the way robodebt was applied.

Touch one, touch all

Perhaps the best example of this long-term, ethical development is Single Touch Payroll (STP) that pulls not just money but key data off business software to bring the tax system and economic management as close to real-time as a payroll cycle can get.

This year, for the first time, the Australian Bureau of Statistics will use Single Touch Payroll to extract labour force and wage data, unburdening businesses and governments from digging through their own system to fill in extensive forms.

Importantly, the data will be direct and holistic rather than sampled, sharpening the picture and increasing granularity.

These are the foundations for a trusted, ethical and functional government application of AI that promises to improve services and make them more affordable and targeted.

After a fairly rough few years, a good run for the public sector’s technology projects will be warmly welcomed by all, and it cannot come soon enough.

Six of the best

It’s perhaps not surprising that a tax agency isn’t shy about advocating for rules around data, given it likes to enforce rules around money that is now almost entirely electronic. Those rules extend to AI and the data it uses.

“We use a common form of artificial intelligence known as machine learning. Machine learning algorithms can consume large amounts of data and provide timely analysis and assessment. Machine learning means that a process which could take months, if done manually, can be done in days, saving time and resources,” Tax says.

“Machines are trained to use historical data to make improvements and are faster than ever at consuming vast quantities of information. This has led to better allocation of workloads and resources, which leads to better services for taxpayers.”

But it’s the culture that really cuts through.

“The ethical use of data and insights is embedded in our corporate culture and values,” Tax says. “In 2020, we pre-filled over 85 million pieces of data with the use of our data and analytics technology.

“We hold ourselves accountable to six ethical standards to ensure you have confidence in how we collect, manage, share and use your data.”

These are the ATO data ethics principles:

1. Act in the public interest, be mindful of the individual
   We administer and ensure the integrity of the tax, superannuation and business registry systems for the Australian community. We recognise our actions impact on the community and individuals, so we will be clear about our intent when we collect, manage, share and use data.
2. Uphold privacy, security and legality
   We respect privacy. We ensure that the individual and community information we hold is kept safe, protected and shared securely as authorised by law.
3. Explain clearly and be transparent
   We are open and will communicate our activities involving data in a way that is accessible and easy to understand.
4. Engage in purposeful data activities
   We only collect, manage, share and use data where necessary to perform our functions and to deliver and enhance our services.
5. Exercise human supervision
   We oversee and are accountable for our activities involving data and the decisions we make.
6. Maintain data stewardship
   As data stewards, we protect the data in our care and respect the stewardship requirements of other agencies. When we acquire or share data, we will agree on how the data will be used and kept securely.