Andrew Watson, Deputy Commissioner and Chief Data Officer
Address at The AI Summit Australia
Melbourne, 9 September 2026
Finding fraud, identifying income and analysing deductions
(Check against delivery)
Good morning everyone.
I’d like to start by acknowledging I am on Wurundjeri land today in Naarm and pay my respects to their Elders past, present and emerging, acknowledge their ongoing connection to this country and extend my respects to any First Nations people with us today.
When people hear the term artificial intelligence, many think of ChatGPT, Copilot, large language models and virtual assistants.
But the reality is that the ATO's story with AI and advanced analytics did not start with generative AI.
In fact, it started decades earlier. Long before myTax, before smartphones were common, and before AI became a household term, the ATO was already using data and analytics to solve problems at scale.
What has changed over time is not our objective. Our objective remains exactly the same: To protect the integrity of the tax system, make it easier for people doing the right thing, and help us focus our efforts where risk is greatest.
What has changed is the capability of the tools available to us to do this.
Today I'd like to share three examples that illustrate that journey:
- finding fraud
- identifying income
- analysing deductions.
And throughout those examples I'd like to reinforce a simple idea: AI and advanced analytics are not competitors. They're complementary capabilities. The key is using the right tool for the right job.
The ATO's innovation journey
To understand where we are today, it's useful to understand where we've come from.
Many people in this room will remember eTax. For those who don't, eTax was the ATO's downloadable tax return application that existed before myTax and the modern online experience we have today. At the time, eTax was considered highly innovative. Most people saw it as a digital way to lodge a tax return. What many didn't see was the growing capability behind it.
As millions of returns were lodged electronically, the ATO began applying increasingly sophisticated analytics to understand patterns within the data. We weren't just looking at individual tax returns anymore. We were looking for relationships, behaviour patterns and common characteristics that linked activity together.
One of the challenges we faced was the growth of unregistered tax preparers.
These were individuals preparing returns on behalf of others without being registered or operating within the appropriate regulatory framework. Looking at a single tax return provided very little evidence. But when you started analysing patterns across thousands of returns, something interesting emerged.
Returns might share common characteristics. They might exhibit similar deduction patterns. They might be lodged from common devices. They might demonstrate common preparation behaviours.
Individually, none of those signals proved anything. Collectively, they formed a pattern. By combining pattern detection with network analysis, the ATO could identify groups of returns that appeared connected and investigate whether an unregistered preparer sat behind them.
That represented a significant shift in thinking. We moved from investigating individual events to understanding systems, behaviours and networks.
That same principle still underpins much of our fraud detection capability today. The scale is larger. The technology is more advanced. But the fundamental idea remains the same. The signal is often found in the connections.
Today's AI capabilities are not replacing that heritage. They are building upon it.
Case study 1: finding fraud
That takes us to our first example: fraud.
Fraud rarely occurs in isolation. Organised fraud often involves multiple identities, multiple accounts and multiple linked activities. When viewed individually, many fraud attempts appear insignificant. The real insight emerges when we understand how they're connected. This is where advanced analytics, including network analytics, have become incredibly powerful.
A recent example is Blast Radius. Blast Radius uses network analytics to identify clusters of potentially compromised taxpayers by examining a range of relationships between different data sets. Rather than reviewing isolated events, analysts can identify groups that exhibit similar behaviours and understand whether they may be linked to the same fraud actor or attack vector.
The approach starts with what is already known. If we identify confirmed instances of fraud, we can then ask:
- Who else appears connected?
- What similar behaviours exist?
- Which other taxpayers sit within the same network?
- Are there broader patterns we have not yet identified?
This capability has significantly improved fraud detection effectiveness while also helping identify emerging fraud typologies that traditional approaches may not uncover. Thousands of taxpayers have been identified for protection through the application of Blast Radius network analytics.
Alongside this, the ATO operates identity crime models that act as a first line of defence by identifying high-confidence fraud indicators and enabling intervention before harm occurs.
What's important here is recognising the technology being used. This isn't generative AI producing content. This is advanced analytics, machine learning and network analytics solving a very specific problem.
And that's an important distinction.
AI versus advanced analytics
People often use the terms AI and analytics interchangeably. They're related, but they're not the same thing.
Advanced analytics helps us identify patterns, classify risk, make predictions and detect anomalies. It answers questions such as:
- Is this likely to be fraud?
- Which cases should we prioritise?
- Which behaviours look unusual?
- Which risks deserve attention?
Generative AI solves a different problem. It helps us understand and work with information. It can read documents, extract meaning, summarise content and interact using natural language.
Neither is inherently better. They simply solve different business problems.
The real magic will happen when we combine them.
Advanced analytics finds the signal. AI helps humans understand and act on it.
Harnessed together, they're far more powerful than either is alone.
Case study 2: analysing deductions
Let's look at deductions.
Historically, reviewing substantiation evidence has been highly manual. Imagine a case officer receives what I call a digital shoebox. Not a neatly organised folder. A shoebox.
Inside might be PDFs, scanned receipts, invoices, bank statements, spreadsheets, emails and photographs. Some documents are relevant. Some aren't. Some are duplicates. Some are poorly scanned. Some contain critical information hidden deep within hundreds of pages.
The challenge isn't applying the law. The challenge is first working out what's in the box. This is where document understanding becomes valuable.
Document understanding uses AI to read, understand and categorise unstructured content at scale. The ATO has been developing these capabilities to support areas such as risk treatment and large-scale document processing.
Think of AI as the assistant who opens the shoebox first. It sorts the contents. It separates invoices from receipts. It groups related records together. It identifies relevant information. It highlights anomalies and areas requiring attention.
But it does not make the decision. The human still makes the judgement. The human still applies the law. The human still determines the outcome.
AI simply removes much of the effort involved in locating and organising information. Instead of spending hours figuring out what's in the shoebox, our people can focus on what they are uniquely qualified to do: apply expertise, judgement and experience.
Case study 3: identifying income
The final example is identifying income. And this brings me to what I believe is the most important lesson in AI.
People often focus on the model. I focus on the data. I've often said: ‘if AI is the engine, then data is the fuel.’
You can have the best engine in the world. But if you put poor-quality fuel into it, it's not going to perform very well.
The exact same principle applies to AI.
The most sophisticated AI model cannot compensate for incomplete, inaccurate or poorly connected data. Quality data remains the foundation of every successful analytics and AI capability.
This is particularly important when identifying income. To accurately understand taxpayer circumstances, we need data that is trusted, connected and fit for purpose.
We need:
- high-quality data
- strong governance
- clear definitions
- consistent standards
- reliable metadata.
The ATO's investments in integrating and connecting data provide the foundation that allows analytics and AI solutions to operate effectively.
The temptation for all of us is to focus on the latest AI capability.
But we believe the ATO will extract the greatest value from AI not necessarily because we build the most sophisticated models. We want to know we have the best data for the task at hand.
Because without fuel, even the best engine goes nowhere.
Conclusion
The lesson from these three examples is simple.
Successful AI adoption isn't really about AI. It's about solving business problems.
For nearly twenty years, the ATO has been using data and analytics to tackle increasingly complex challenges.
From identifying unregistered preparers through pattern and network detection in eTax. To machine learning models that identify identity crime. To Blast Radius and network analytics that uncover organised fraud. To document understanding that helps our people make sense of digital shoeboxes full of information.
Each generation of capability builds on the one before it.
Finding fraud demonstrates the value of network analytics. Analysing deductions demonstrates the value of AI-assisted document understanding. Identifying income reminds us that none of it works without fit-for-purpose data.
If AI is the engine, then data is the fuel.
And neither is particularly useful without skilled people behind the wheel.
Because ultimately, the future belongs not to organisations that deploy AI for the sake of AI, but to organisations that combine quality data, strong analytics, AI capabilities and human judgement to solve real problems.
And that has always been the ATO's approach.