AI in tax: How to make it safe and scalable

More from the series

AI in tax: What is it?

Read the previous article

AI investment is accelerating, so governance and ROI discipline matter 

As investment in AI accelerates, expectations rise alongside it, boards and management teams want faster cycle times, better consistency and clearer insight, without creating new categories of risk. For tax functions, that makes it important to be disciplined about where AI is used, how outputs are controlled, and how value is measured beyond early pilots. 

Governance matters because tax is a “trust” domain. AI-assisted work should be designed so that it remains defensible (clear sources and reasoning), auditable (versioning, review notes and decision logs where appropriate), and confidential (controls over what data can be used, where it is processed, and who can access it). Just as importantly, responsibilities should be explicit: AI does not own conclusions, people do. 

Recent high-profile incidents in professional services also illustrate why controls matter. For example, a Big4 firm reportedly agreed to partially refund the Australian government after a report was found to contain fabricated references and an incorrect quotation, with subsequent updates disclosing the use of a generative AI tool in preparing parts of the report. The lesson for tax functions is straightforward: if AI is used to accelerate drafting or analysis, organisations should hard-wire validation, source-checking and clear approval responsibility into the workflow, so speed does not come at the expense of accuracy and trust. 

ROI discipline matters because many AI benefits are easy to describe but harder to evidence at scale. A practical approach is to start with a small number of priority workflows (e.g., technical research and memo drafting, document review, routine correspondence, data extraction for compliance), define quality gates, and track outcomes over time. 

In a tax context, useful measures often include: turnaround time (from issue raised to reviewed output), rework rates (how much must be corrected in review), technical quality indicators (citation completeness and accuracy), consistency across deliverables, and risk outcomes (fewer late filings, fewer follow-up questions from stakeholders, stronger audit trails). Over time, these operational measures can be linked back to cost-to-serve, capacity and risk reduction, helping leaders make informed decisions about scaling AI use cases. 

Practical steps for tax leaders to adopt AI safely 

For most organisations, the near-term impact of AI in tax is not “hands-free automation” of complex positions. It is augmentation, faster drafting, better retrieval, more consistent documentation and improved triage, provided the organisation sets clear boundaries for use and builds a control environment that is fit for purpose. 

Define where AI is allowed, and where it is not: Separate low-risk productivity use (e.g., summarisation, drafting support) from high-risk decisions (e.g., technical conclusions, filing positions, transaction recommendations) that require human sign-off. 

Build a defensible audit trail: Require citation to primary sources (legislation, IRAS guidance, treaty text, contracts, working papers) and retain prompts, versions and reviewer notes where appropriate. 

Strengthen confidentiality and data controls: Confirm what data can be shared with third-party tools, apply access controls, and align usage with your internal policies and contractual obligations. 
Design quality gates into workflows. Use checklists, sampling and structured review to detect hallucinations, incomplete analysis and misinterpretation before output reaches a filing, board paper or external audience. 

Update the operating model and skills mix: As routine tasks reduce, invest in reviewers/interpreters who can validate AI outputs, manage data quality and translate results into decisions. 

Where AI is genuinely useful in a tax function today 

For many tax teams, AI delivers the most consistent value when it is applied to clearly bounded tasks, supported by strong source materials and followed by structured review. Examples include document summarisation, first-draft writing, checklist generation and issue-spotting prompts to reduce omissions, provided outputs are verified against primary sources and the facts. 

In other words, the first process AI tends to tackle is outsourced, standardised work – the sort of templated returns and vanilla processing some firms had already shifted offshore long before AI models were an option. The work that remains in-house, and especially in a financial centre like Singapore, is the work that is hardest to codify.

Use AI where it is genuinely strong 

Research shows that AI is most reliable when tasks are: 

• Narrow, well specified and low-risk (e.g., drafting emails, basic summaries, simple comparisons).
• Shorter-horizon, not multi-week, and not requiring long-term institutional memory.
• Tolerant of occasional errors, where human oversight is cheap and fast.

In a tax practice, that maps to things like: 

• First-cut summaries of Budget changes, rulings or new circulars, which you then verify against primary sources and your own notes.
• Generating alternative phrasings of advice once you have decided on the technical position and the risk level.
• Creating checklists or skeletons for transaction-steps, due-diligence questionnaires, or training slides, which you then refine and contextualise.
• Brainstorming potential issues or structuring options as a starting point, not as client-ready output.

Used this way, AI can offload some of the repetitive drafting and formatting work while your team retains control over the technical position and risk decisions. The discipline is the same as any quality-critical work: no output should be treated as client-ready (or filing-ready) without professional review and verification against source documents.