AI in tax – revolution, risk or somewhere in between?

By Joe Walker & Philip Claridge 

If endless LinkedIn posts are to be believed, the future of outsourcing your tax positions entirely to AI has arrived. AI will have your tax return done better, faster and cheaper than your old fashioned human tax advisor (subscribe to my newsletter for more!).

Based on our own conversations with tax and finance teams, we understand that a wholesale AI tax revolution remains subject to certain timing differences. However, where there is smoke there is fire and many teams are increasingly adopting AI tools as part of their everyday work, and finding that, used well, AI can be a powerful productivity tool.

That raises an obvious question: if AI is increasingly being used to help prepare tax positions, what happens when the answer is wrong, and how will Inland Revenue view taxpayers relying on it?

Can taxpayers rely on AI-generated tax advice?

The short answer to this is “no”. Ultimately, taxpayers are responsible for the correctness of their tax positions. In and of itself, reliance on an advisor (human or AI) does not prevent Inland Revenue from reassessing a taxpayer’s self-assessment. However, if the reassessment stands, the process by which the taxpayer formed that self-assessment, including the use of AI, is relevant to the penalties that apply to any tax shortfall that was corrected.

Inland Revenue recently commented on this briefly in interpretation statement 26/04 (IS 26/04), which discussed the application of shortfall penalties for “not taking reasonable care”. The message of IS 26/04 in relation to AI is practical and unsurprising: reasonable care depends on how the technology is used. When dealing with a complex tax issue, a single prompt to a general-purpose AI tool, followed by uncritical reliance on the answer, will not usually be enough. Inland Revenue expects taxpayers to use accurate facts, check sources against trusted materials such as legislation, case law, Inland Revenue rulings and other guidance, and review outputs for accuracy and relevance to the taxpayer’s circumstances.

On the face of it, this is a reasonable approach. In our experience, there are risks with relying on AI tools in a tax context. For example:

1. AI tools can be wrong in ways that look right. It may produce a fluent answer that blends correct principles with outdated law, foreign tax concepts, or invented authorities. On the face of it such ‘advice’ appears logical and well thought out, even to a professional, and it is only on closer examination that the issues are evident.
2. AI tools correctly answer the question posed, but either:
   a) The ‘wrong question’ is being asked – a common scenario encountered by advisors; or
   b) Clarifying questions are missed, such that the answer provided relies on assumptions that are not appropriate to the context.
   

This is not to say that AI tools are unhelpful. On the contrary, the performance of state-of-the-art AI systems is remarkable. They may well surface issues that would have otherwise been missed or highlight areas that require further consideration or advice. Further, as the performance of AI models improves, the types of issues noted above may become less common. However, as Inland Revenue emphasises, notwithstanding how impressive AI tools may seem at face value, uncritical reliance should be avoided. For tax positions involving judgement, material amounts, uncertain law, or unusual facts, taxpayers should treat AI output with particular care.

AI tools and Inland Revenue search powers

Prompts, uploaded materials, model outputs, and other data associated with AI tools are often retained, at least for a period, by one or more parties involved in providing or using the service. The security and governance of this data is already on the radar of most organisations as a risk management consideration. However, as part of this, taxpayers should bear in mind that Inland Revenue has extremely broad statutory powers to compel taxpayers and third parties to provide documents and information, including records held electronically.

As the use of AI tools expands, it is conceivable that Inland Revenue may seek to use its search powers to access records/data generated by AI tools as part of its enforcement activities. For most end users this paper trail is largely invisible, however it may extend well beyond the ‘chat history’ visible to them. Requests to provide information could be directed to taxpayers themselves or, in some cases, to technology service providers directly. While we are not currently aware of Inland Revenue making requests of this nature, taxpayers should nevertheless consider what tax-sensitive information is being retained, for how long, and by whom.

Certain information may be protected from disclosure to Inland Revenue, including material that qualifies as legally privileged or as a “tax advice document” under New Zealand’s tax legislation. There might be some circumstances where information generated by an AI tool falls within the scope of these protections, but this should not be assumed. These rules are complex, and taxpayers should obtain specific advice on their application.

Deloitte comment

AI tools can be a valuable assistant: helping to frame issues, test thinking, uncover relevant sources and improve efficiency. However, Inland Revenue’s comments are a timely reminder that AI does not change the fundamentals of tax compliance. Taxpayers remain responsible for the positions they take and cannot outsource this responsibility to an AI tool.

For tax and finance teams, the practical challenge is to realise the benefits of AI without allowing speed and convenience to dilute quality or create other risk exposures. This should be considered as part of an organisation’s broader AI and data governance settings, including how AI tools are approved, what information may be entered into them, how outputs are reviewed, what records are retained (and by whom). Judgement is required as to whether tax-specific policies are required, as this may depend on the organisation’s existing controls.

Organisations may also need to revisit and adjust their approach as AI tools evolve. For example, the ability for AI agents to orchestrate long-running workflows (potentially autonomously or with limited user input) sharpens the challenge in relation to:

1. Ensuring appropriate reviews and controls are in place before increasingly complex outputs are incorporated into tax positions.
2. Managing the proliferation of tax sensitive data, particularly where this may be being documented without user visibility. 

AI is a technology with the potential to fundamentally shift how organisations operate, including finance and tax teams. In our view, tax is not unique - AI is an opportunity to be embraced not avoided. However, as with other technology its use should be intentional, with the risks understood and appropriately controlled.

If you have any questions about AI and tax please contact your usual Deloitte adviser.