Andrew Watson, Deputy Commissioner and Chief Data Officer
Address at the Australian Government Data Summit 2026
Canberra, 22 April 2026
ATO’s AI strategy - From Models to Co-pilots: How the ATO Is Industrialising AI, Responsibly
Opening – The ChatGPT moment
Good afternoon everyone
In the spirit of reconciliation, I want to acknowledge the Traditional Custodians of the land on which we meet today, the Ngunnawal people, and pay my respects to their Elders past and present, and to any Aboriginal and Torres Strait Islander peoples here today.
At the ATO, we believe that in addition to delivering on the outcomes in our Reconciliation Action Plan, if we all do our job well — fairly, respectfully, and in a way that all people can trust — we contribute further to reconciliation by helping governments to fund the services and programs that support our community.
Most people in this room would remember November 2022.
That was the moment ChatGPT appeared and almost overnight, artificial intelligence stopped being something talked about in specialist forums, and became something everyone could see, touch and experiment with.
Suddenly, the question wasn’t ‘What is AI?’, it was ‘How can AI do this for my work?’
For the ATO, that wasn’t the moment that we started our AI journey. Rather, it made the potential of AI more visible, and with that, increased expectations, pace to adopt and scrutiny.
That’s the story I want to share with you all today.
Machine learning has been core, to the ATO for years
Long before generative AI, the ATO was already using machine learning models as part of how we administer the tax and super systems.
These models included:
- Predictive models to assess and prioritise risk
- Classification models to triage and group cases
- Anomaly detection to surface unusual patterns, and
- Models that learn from historical outcomes and improve over time
So, we have operated in an environment where we have had a mix of models that produced deterministic outcomes alongside an increasing number of probabilistic models.
These weren’t pilots or experiments, but production machine learning models, embedded in real workflows and decisions, with humans in the loop.
This is why our AI strategy talks about moving from pockets of excellence to enterprise capability - not starting from zero - because we have done this before with machine learning.
Our strategy also states that AI should augment what we do, not sit by the side, as a novelty.
Essentially, machine learning in the ATO, has been doing exactly that for years:
- Improving consistency
- Supporting prioritisation
- Assisting our staff to focus effort where it matters most, and
- Helping taxpayers to easily get it right from the start.
The evolution of Generative AI didn’t replace this foundation.
Rather, it has added, and continues to build, on them.
If AI is the engine — then data is the fuel
What hasn’t changed since the first model went live is that there is no AI without data.
And there is no trusted AI without trusted data.
The ATO’s data holdings are a national asset. We hold a large amount of data, in trust for the Australian community – collected from taxpayers when they lodge their returns or acquired from third parties.
We use this data, and the insights they provide, to administer the tax and superannuation systems more effectively. This includes tailoring our taxpayer services and making more informed decisions.
That’s why we take great steps to ensure we have strong data stewardship, lineage and governance in place, to protect our data and ensure it is used and shared ethically.
This includes when data is used in automation and AI. These controls aren’t optional extras, and are what make AI safe to scale in the ATO environment.
Without them, poor data quality leads to poor model outcomes. This in turn leads to inconsistency and unfairness that directly affects community trust and can also have significant adverse impacts on individuals.
In essence, AI doesn’t fail, the data does.
Governance, ethics and social responsibility: Trust is the product
As AI became more visible, expectations rose quickly about its potential.
However, like any emerging technology, AI introduced new risks, particularly relating to privacy, fairness, security and transparency.
While there is a challenge in leading the accelerated adoption of AI with responsible use in the ATO, governance and ethics are not an afterthought, they’re foundational to the way we operate.
As stated in our AI transparency statement, ‘we recognise the importance of robust governance, oversight and accountability. This helps us to ensure AI development and use is ethical and safe and delivers fit-for-purpose outcomes’.
Basically, our approach is simple - if we can’t explain it, defend it, or audit it - we shouldn’t deploy it.
Our AI policy makes this clear:
- Only approved tools for approved use cases
- Human oversight is mandatory
- Accountability sits with people, not systems, and
- AI generated records must be properly managed
A key principle that underpins all of this is that accountability does not get automated. While AI can help make more informed decisions, as well as surface insights, ultimately responsibility always remains with us.
As an administrator, we use human judgement when considering taxpayers’ rights and entitlements to ensure a fair tax system for everyone. This is especially so when interacting with taxpayers who may be vulnerable or from diverse backgrounds.
Ensuring the right governance foundations are in place, is not about slowing innovation, rather it’s acknowledging the importance of earning and maintaining trust in our administration, and the important role we have in safeguarding taxpayer data.
A colleague recently shared with me an African proverb I think is relevant here – ‘alone we can go fast but together we can go far’.
By bringing together the technologies, the data, the governance, the skilling, and change management to the right business problems, we believe AI can help us go far at the ATO.
What changed with Generative AI — and what didn’t
So, what actually changed for the ATO after 2022?
We’ve enhanced our AI capability by introducing Natural language interfaces, particularly to identify and stay ahead of potential threats. For example, the Digital Identity System Fraud and analysis capability established by the ATO and part of the broader Australian Government Digital Identity ecosystem to detect, prevent and respond to fraud at scale.
We developed the ability to summarise, draft and synthesise information such as ‘document understanding’ bringing together vast quantities of diversified documents for auditors to review, as substantiation of a taxpayer’s work-related expense claims. This meant building AI capabilities to machine:
- read receipts, invoices and other unstructured documents,
- suggest which ones relate to what deduction claim, and
- prioritise a likely best order for our auditors to consider the documents.
The validity of deduction claims are still determined by our auditors. But rather than staff having to sort through a randomly organised digital shoebox, they have an AI assistant that organises and prioritises documents, saving time and making it easier for them to assess the claims. At the same time this AI solution, then asks for feedback and continuously learns from how it did.
During 2025 we gave our staff much broader accessibility to AI, with Generative AI tools such Microsoft Copilot Chat and Copilot M365, while ensuring enterprise data protection were in place. Given the sensitivity of much of the data held by the ATO, it was really important that the decision to make a consumer level Generative AI tool available to all our staff, considered what, if any, data left the secure ATO IT environment.
In deploying Generative AI tools, what didn’t change for us, is the need for:
- quality data
- explainability
- governance, and
- human judgement
While Generative AI is now more accessible, it doesn’t replace our machine learning capability, nor our highly trained and skilled data scientists who develop and deploy our enterprise AI tools.
Generative AI sits alongside our machine learning capability, supporting staff productivity by reducing time spent on tasks, particularly repetitive administrative work.
Like all large organisations we have seen a broad spectrum in the adoption and use of Copilot from enthusiastic early adopters to healthy sceptics. But overall, it is helping us develop greater AI literacy and capability across the ATO. This positions us well to move from using AI for personal productivity to embedding it in enterprise-wide processes.
We are now moving beyond just using Generative AI tools and are shifting our focus to how we can use AI deliberately, safely and at scale.
From experimentation to industrialisation: 7 priority use cases
This is where the ATO is making a deliberate shift.
In making Gen AI tools available, we are seeing new ideas and use cases being generated across the organisation, and at the same time, the risk of ‘a thousand flowers blooming’ everywhere.
Rather than experimenting all over the place, our Executive recently prioritised seven enterprise AI use cases — not as technology showcases, but as business capability plays. They fall into a few clear themes.
First, decision and case support.
Some use cases focus on helping assemble information, such as:
- Case profiling assistants that bring together relevant facts, map them to legislation and policy, and prepare material for review, and
- Support for triaging taxpayer requests that arrive in large volumes and in a variety of unstructured formats, helping assess complexity, identify related matters and suggest next steps
The decision still sits with a person, while AI does the heavy lifting of information assembly.
Second, drafting and reasoning support.
These use cases support drafting:
- Position papers
- Structured correspondence, and
- Clear, referenced reasoning documents
AI helps bring together decisions made by humans with well-structured material — consistently and efficiently — to draft documents that remain subject to human review.
Third, specialised operational support.
Use cases include:
- Assistants that help assemble evidence and map against policy criteria, and Finance focused Copilot capabilities to support repetitive reconciliation and analysis tasks.
These are areas where processes are complex, repeatable and information heavy, and where consistency really matters.
Fourth, internal productivity and capability uplift.
Use cases such as:
- An ‘AskHR’ chatbot to make HR information easier to access for our staff and assist with routine processing tasks
- Coding and reverse coding assistants to help interpret and draft code in ATO systems. The ATO has a rich history of creating its own code but sometimes when the person who had created and maintained that code over a number of years leaves the organisation there may be gaps in the documentation supporting the code. We want to test the opportunity for AI to draft support materials and suggest what the original requirements for the code were.
These are about removing friction or time-consuming manual tasks, not about cutting corners.
Why these seven use cases were chosen, is that they allow us to:
- Build reusable platforms and help us make decisions on what and how many platforms will best support us going forward,
- Embed assurance and ethics checks aligned with APS and ATO policies,
- Establish clear ownership and accountability – we need AI use cases to be business lead and enabled by technologists – not the other way around,
- Work with our people on how to best manage any change, and
- Prove AI can be delivered lawfully, transparently and at scale.
Ultimately, focussing on these seven will build our organisational muscle to:
- identify AI use cases that can deliver value, and halt work on those that don’t,
- progress successful pilots and proof of concepts to embed AI capabilities within business processes,
- manage change across people, process and technology to maximise the potential benefits of AI
- embed good governance and risk management, and
- establish how we want to deploy Agentic AI.
Agentic AI is the next frontier, and we are starting to work through the right mix of:
- agents that are part of platforms
- agents that are part of software, and
- agents we will build ourselves.
We also have a rich history with business automation, particularly Robotic Process Automation that gives us some good lessons and examples of how to blend technology with our people and in business processes, that can guide us in how we deploy Agentic AI.
Put simply, we cannot scale tools without scaling trust.
Closing – the point of it all
In closing today, AI will inevitably reshape how we work and create opportunities to better interact with taxpayers, strengthen integrity and build a future-ready workforce.
AI will enhance the way we operate by:
Reducing friction
Improving consistency
Strengthening fairness, and
Supporting better decisions
AI — whether machine learning, generative or agentic — is not about replacing people. ATO staff are at the core of what we do. It’s their judgement, empathy and values, that enable trust.
Data has always been the fuel that has driven our analytics and data will continue to fuel expanded use of AI. We need to understand and ensure the quality of the data and only use it where it is fit for purpose.
Finally, Australia’s tax system is acknowledged globally for its high level of voluntary compliance, which in a large part, is built on high levels of trust and sophisticated use of data.
Good data fuels good models, both deterministic and probabilistic, good governance builds trust, and trust is the real value we deliver and need to protect on our AI journey.
Thank you and I welcome any questions.