Kirsten Fish, Second Commissioner, Law Design and Practice
CPA Tax Forum 2026
Sydney, Thursday 20 August 2026
(Check against delivery)
Good afternoon everyone.
I would like to begin by acknowledging the Gadigal people of the Eora Nation, the traditional custodians of the beautiful land on which we meet today. I pay my respects to their Elders past, present and emerging, and all Aboriginal and Torres Strait Islander peoples here today.
Artificial intelligence is a topic that comes up in almost every conversation, and the context of tax administration is no exception.
AI and automation are already part of our tax system. The real question now is how we can use this new technology to improve services, make compliance easier, strengthen the integrity of the system and support a better experience for the community. And, importantly, how do we do that lawfully, transparently and in a way that preserves the things that make the system work: trust, judgment and accountability.
AI will change how tax administration and tax advice are delivered, but it will not change the fundamentals of the system. Taxpayers remain responsible for the information they provide. Tax professionals remain responsible for the judgment they exercise. And the ATO remains responsible for administering the law fairly, lawfully and transparently.
AI did not start with ChatGPT
When ChatGPT became widely available in late 2022, it changed the public conversation about AI almost overnight. AI moved from specialist forums into everyday work conversations. People could see it, test it and imagine how it might change their own jobs.
But tax administration has always been shaped by technology. Long before anyone talked about generative AI, technology was already central to the way Australia's tax and superannuation systems operated.
Consider the scale of the system.
- millions of taxpayers
- millions of businesses
- billions of transactions
- an enormous volume of information moving through the system every year.
Technology has always been essential.
Inside the ATO, the emergence of large language models felt more like the beginning of a new chapter in a story that had already been running for quite some time. Starting with automation, moving to machine learning and advanced analytics, and now generative and agentic AI.
Each era has built upon the previous one: generative AI has not replaced analytics, analytics did not replace automation. Instead, each layer adds another capability.
The technology changed, but the underlying questions remained largely the same: how do we help people comply more easily and help the tax system operate more effectively and with greater integrity.
Why this matters for the tax system
When we talk about AI in tax administration, the technology is only one part of the story.
The more important question is what kind of tax system it helps us build.
From the ATO’s perspective, there are 4 practical outcomes we are looking for.
The first is making compliance easier.
Pre-fill is a very familiar example of how we use rules-based processes, data matching, digital lodgment and workflow automation. Most taxpayers no longer need to manually enter large amounts of income information because it is already available through information reported to the ATO. That is not generative AI. It is not machine learning. But it is technology reducing effort, reducing mistakes and making compliance simpler.
Real-time prompts are another good example.
Historically, somebody might lodge a return containing an error and not hear from us until much later, through a costly and time-consuming review, adjustment or audit.
Now, predictive models operate in real time during the lodgment process to compare a taxpayer’s claim with claims made by people with similar characteristics, such as income and occupation, and prompt the user to check the figure and correct it if appropriate.
This moves the interaction earlier. The taxpayer still chooses what to lodge but receives relevant information while they can still act on it.
In 2024–25, this real-time technology enabled more than 720,000 prompts for taxpayers to check amounts in their individual income tax return. Taxpayers who received a prompt are more likely to review their return and more likely to adjust their data. This protected an estimated $62.5 million in revenue from incorrect returns.
That is not only about detecting mistakes. It is about helping people avoid them. It represents the evolution we are aiming for: shifting administration from retrospective correction to on-time support and assistance.
The second outcome is earlier intervention and stronger system integrity.
AI and analytics make it easier to identify emerging patterns. That may be unusual claims, possible identity theft, fraud networks or broader compliance risks. The earlier we identify an issue, the more options we have to respond proportionately.
Fraud detection, identity protection, risk modelling and network analytics all support the same broader purpose: protecting the integrity of the tax and superannuation systems. As fraud becomes more sophisticated, the tools used to detect it also need to become more sophisticated.
Fraud is increasingly digital, organised and networked. Most often, it does not appear as one obvious event. Rather, the activity may involve multiple identities, businesses, accounts or changes to tax account details.
The ATO has developed a Blast Radius capability, which uses network analysis to identify connections between entities and detect potential fraud patterns that would be extremely difficult to identify manually or when individual cases are viewed in isolation.
Between February 2025 and February 2026, this capability helped the ATO identify and protect 12,830 taxpayers from harm associated with fraud. That means fewer people affected by identity crime, greater confidence in the integrity of the tax system and better protection of revenue for the Australian community.
The third outcome is better use of expertise.
Like every regulator, we face a simple reality: there are always more potential compliance issues than there are resources to investigate them all.
This is where models are used to detect patterns, identify outliers, prioritise work and support staff to focus attention where it is most needed. The ATO has used these capabilities across risk assessment, fraud detection, anomaly detection, case triage and compliance support.
Using machine learning, our risk models help us identify where attention is most likely to be needed. Not to make decisions. Not to determine outcomes. But to help prioritise effort.
There is also a great deal of work in tax administration that involves assembling material: bringing together facts, documents, correspondence, financial data, legal references and prior interactions. AI can assist with that preparatory work.
Our document understanding capability is one example of this.
Work-related expense claims are a significant area of compliance focus. The substantiation material can be extensive and diverse: receipts, invoices, bank statements, logbooks, PDFs, photographs and other records – on average 147 pages per audit.
The technology machine reads and categorises the unstructured documents, extracts key information, matches documents to claims and prioritises a likely best order for consideration by the case officer.
Importantly, the validity of deduction claims are still determined by our auditors. However, the technology ensures our experienced officers spend less time sorting files and more time on the real question of whether the material substantiates the claim under the law.
During 2025 we gave our staff much broader accessibility to AI, expanding access to enterprise-enabled tools, including Microsoft Copilot, with enterprise protections embedded from the outset. This supports staff productivity by reducing time spent on tasks, particularly repetitive administrative activities.
That is the practical opportunity of generative AI in a tax administration context.
Not replacing officers.
Not automating statutory responsibility.
Reducing the amount of time skilled people spend on work that technology can assist with, so they can spend more time on the work that only people can do well - dealing with complex issues, exercising judgment and engaging with taxpayers and agents.
The fourth outcome is better decision quality.
The ATO makes millions of decisions each year. While the vast majority are correct, the reality is that some decisions will be inconsistent or wrong. Today, we often discover those issues through objections, complaints, reviews or appeals after the fact. AI gives us the opportunity to identify them much earlier.
By reviewing decisions at scale, AI can help identify patterns, inconsistencies and outliers that would be impossible to detect manually. This is not about AI making decisions. It is about helping us understand where decisions may warrant closer human review and where our practices may be drifting from established standards.
That creates a powerful capability for continuous improvement. We can identify specific case types, issues or decision-makers that may need additional support, and provide targeted coaching, training and guidance. Rather than waiting for errors to emerge months later, we can intervene earlier and lift capability across the organisation.
The result is better decision-making, greater consistency and fairer outcomes for taxpayers. While much of the discussion about AI focuses on productivity, one of its greatest long-term benefits may be improving the quality of administrative decisions. In a system as large and complex as ours, even small improvements in decision quality can deliver significant benefits for taxpayers, the ATO and confidence in the integrity of the system.
Technology does not displace administrative law
One question that increasingly arises is whether AI changes the legal framework within which tax administration operates.
The short answer is no.
Administrative law principles apply whether information comes from a paper file, a spreadsheet, a business rules engine, a machine learning model or a generative AI tool.
Decisions still require lawful authority.
Relevant considerations still need to be taken into account.
Irrelevant considerations still need to be avoided.
Procedural requirements still need to be followed.
People still have review rights.
Decision makers still need to be able to explain their decisions.
Technology may change how information is gathered, organised or presented.
But it does not change the legal requirements for a valid decision or exercise of power. It does not change the legal character of the decision. It does not move accountability from the officer to the tool.
That is particularly important where decisions affect rights, obligations or entitlements.
For the ATO, that means our systems need to be designed so officers can understand what role AI has played and can make a decision that is lawful and defensible.
AI risk and governance in the ATO
The ATO’s approach to AI governance starts with a simple recognition: the risks are practical, not theoretical.
AI can deliver enormous benefits, but in tax administration its risks affect real people. That means our governance approach needs to be proportionate to the consequences of getting it wrong.
The first risk is privacy and confidentiality. The data we hold is entrusted to us by the Australian community. It is not our data. The emergence of powerful AI capabilities does not lessen our obligations, if anything, it heightens them. That is why we place such emphasis on access controls, data stewardship, metadata, lineage and governance. We need confidence that data is secure, appropriately sourced and fit for purpose before it is used in analytical or AI-enabled processes.
The second risk is error, including error at scale. Generative AI can produce answers that sound plausible and confident but are wrong. In tax administration, that matters because decisions can affect refunds, liabilities, penalties, entitlements, cash flow and sometimes the viability of businesses. Technology can scale good administration, but it can also scale poor administration. That is why transparency, oversight and accountability are so important. We publish an AI Transparency Statement, we require approved tools for approved use cases, and we ensure AI-enabled processes remain explainable, auditable and subject to scrutiny.
The third risk is over-reliance. One of the most practical challenges is the temptation to assume that because a system generated an answer, the answer must be correct. But in tax, small differences in facts, timing, purpose, evidence or legal context can produce very different outcomes. Human judgment remains essential. Our policy is clear: accountability rests with people, not systems. AI can assist, support and surface insights, but responsibility remains human. Decisions that adversely affect taxpayers’ rights will always be made by a person, and taxpayers retain their rights of review regardless of the technology used to support those decisions.
Ultimately, our governance approach reflects the nature of the risks we are managing. The greater the potential impact on privacy, taxpayer rights or public confidence, the greater the need for controls, oversight and accountability. For us, the principle is straightforward: if we cannot explain it, defend it and audit it, we should not use it in a tax administration process. AI governance is therefore not separate from good administration. It is how we ensure innovation strengthens, rather than compromises, trust in the tax system.
Interestingly, those same principles are increasingly relevant for tax professionals, because advisers and agents are confronting many of these same opportunities and risks.
What AI means for tax professionals
Whenever AI is discussed, one question inevitably arises.
What does it mean for the profession?
My sense is that some activities will change significantly.
- searching for information
- summarising documents
- preparing first drafts
- assembling materials
- researching legislation.
These activities are increasingly being supported by AI.
A task that previously required hours may now take minutes.
That does not remove the need for a tax professional.
On the contrary, the value of professional skill increases as information becomes easier to access. While everybody has access to tools capable of generating a technically plausible answer, only a professional can determine whether it is correct, relevant and appropriate.
The skills of tax professionals will necessarily need to evolve, and where and how professionals add value will change.
Tax professionals will increasingly need to challenge outputs, validate conclusions and understand the quality of the underlying information. This requires the knowledge to identify irregularities and the professional judgment to navigate uncertainty and competing perspectives.
Increasingly tax professionals must understand not just what the data is, but what it means. They must collaborate with other experts across industries. They require expertise and experience to drive insight that will create commercial value, and honed communication skills to effectively deliver it, cutting through uncertainty and complexity.
And ethics and integrity will remain fundamental. In recent times, aspects of professional conduct and integrity across parts of the profession have come under scrutiny and criticism. Australia is reliant on tax to fund government services, and the efficient running of the tax system is reliant on the existence of tax professionals. We, as a country, need tax professionals that are trusted and trustworthy.
AI is a market transformation.
Taxpayers and clients will use AI.
Agents and advisers will use AI.
The ATO will use AI.
The practical challenge is ensuring those uses contribute to better compliance, better advice and better administration.
What comes next
If I step back and look at the trajectory, I see a fairly clear progression.
First came automation. Then machine learning. Now copilots.
The next phase involves increasingly agentic systems and more structured AI support for end-to-end workflows.
Within the ATO we are already exploring use cases across our operations and in our tax administration.
Firstly, those that will assist us to run more efficiently as a large employer organisation.
These include an AskHR chatbot to make HR information easier to access for staff and assist with routine processing tasks, coding assistants and reverse-coding assistants to maintain systems more effectively and interpret legacy code more easily, and a finance focused tool to support and complete repetitive reconciliation and analysis tasks with greater consistency.
Secondly, we’re developing use cases that will assist us to be more efficient and accurate in our taxpayer interactions and administration. We are about to roll out a pilot on call transcription and summarisation. This will deliver significant benefits with reduction in call times, improved record quality and an ability to scale our analysis of call drivers and how we can improve our service.
We’re exploring AI in the triage of taxpayer requests to identify complexity earlier, recognise related matters, flag urgency and suggest the most appropriate pathway for handling. This will help route requests faster, improve consistency, reduce manual sorting and ensure matters requiring specialist attention are identified sooner.
We’re developing a case profiling assistant to bring together relevant information from multiple systems, including taxpayer history, prior interactions, documents, correspondence, risk indicators, map facts to legislation and guidance and prepare material for officer review.
If our officers can begin with a clearer picture and spend less time assembling background material, they will have more time to understand the issue, engage with the taxpayer or agent and make a properly supported decision.
And we’re exploring the use of AI for drafting and reasoning support to help officers structure and draft position papers, correspondence and referenced reasoning documents. This should produce clearer, more consistent and better supported explanations for taxpayers, while preserving human decision-making, accountability and responsibility.
Each of these use cases has been deliberately chosen to enable us to build reusable platforms, embed assurance and ethics checks and prove AI can be delivered lawfully, transparently and at scale.
Conclusion
AI and automation are already changing tax administration, tax compliance and the tax profession.
The opportunity is to use AI to make the system easier to comply with, more responsive, more consistent and stronger in integrity.
The challenge is to do that in a way that preserves trust.
The future of tax will not be about choosing between technology and expertise.
It will be about combining them well.