A practical look at where AI strengthens tax computation, where it creates risk, and how iLevvy keeps professional judgement in control.
By Sharanne Au | Tax Intelligence
The conversation around AI in tax software has largely been promotional. Vendors promise transformation; practitioners remain sceptical; somewhere in between, the real picture gets lost.
This article attempts a more honest account, one that takes seriously both what AI genuinely enables in tax computation and where it fails, distorts, or creates new categories of risk. It also sets out the design principles that a well-built AI tax system should embody and how iLevvy approaches each of them.
Part I: What AI Gets Right
1. Reasoning About Ambiguity, Not Just Applying Rules
Rule-based computation engines are built for deterministic problems. Tax is rarely deterministic. The question is not always "what does the statute say?" It is "what does it mean in this specific fact pattern, given current IRAS guidance and prevailing practice?"
A company with a lease arrangement straddling FRS 116 and hire purchase treatment cannot rely on a rule engine to characterise the arrangement. Characterisation is a judgment call and it determines everything that follows. A well-designed AI system surfaces the interpretive question, presents the competing positions, and maps the risk profile of each before the computation runs.
This is the foundational shift that AI makes possible: from applying law to reasoning about it.
iLevvy's approach: Ambiguous classification events in iLevvy are not silently resolved by the system. They are surfaced as reviewable judgment nodes. The preparer sees the competing treatments, the applicable provision, and the system's recommended position, and signs off explicitly. The decision becomes part of the workpaper, not a hidden assumption.
2. Eliminating the Low-Value Hours
A substantial portion of every tax engagement is not professional judgment, it is data wrangling: tracing figures from accounts to tax schedules, reconciling trial balance items to the correct treatment buckets, cross-checking inputs against prior-year workpapers. These hours are expensive, error-prone, and largely unrewarded by clients who do not understand why they cost so much.
AI-assisted import pipelines can classify ledger entries, flag anomalies, and populate tax schedules at a fraction of the manual effort with full audit trail transparency.
iLevvy's approach: iLevvy's import resolver classifies trial balance line items against a three-layer taxonomy: account type, tax treatment category, and schedule destination. Classifications are logged, reviewable, and overridable. The preparer's time shifts from populating schedules to reviewing judgment calls.
3. Scaling Expertise Without Scaling Headcount
For boutique practitioners, the economics of tax compliance have always been difficult. Sophisticated analysis, group relief elections, loss carry-back strategies, foreign tax credit optimisation, requires deep technical capability that is expensive to develop and retain. Clients who need it often cannot afford the advisory time it requires.
A well-designed AI tax engine can surface relevant provisions, model alternative positions, and present the decision to the practitioner for sign-off. The practitioner provides judgment and accountability; the AI provides analytical reach.
iLevvy's approach: iLevvy's loss item utilisation module and further deductions engine are built to systematically surface every available position capital allowances, donations, group relief, carry-back elections rather than relying on the preparer to remember to check each one. Nothing falls through the cracks because a busy practitioner forgot to ask.
4. Producing Workpapers as a By-Product, Not an Afterthought
Tax workpapers are the evidentiary backbone of any filing position. Good ones take time. Under deadline pressure, they get cut short. When a Revenue authority queries a position three years later, the quality of the contemporaneous workpaper is often the difference between a swift resolution and a prolonged dispute.
AI-assisted computation can generate workpapers as a structural output of the computation itself, every adjustment traced to its statutory basis, every judgment call documented with the reasoning captured at the time it was made.
iLevvy's approach: iLevvy produces its own workpaper and filing reference sheet formats, not IRAS form replicas. The output is designed to be used as an internal computation record and a basis for professional review, with statutory references embedded at the adjustment level.
Part II: Where AI Falls Short, And the Risks That Come With It
5. AI Hallucinates, Including on Tax Law
This is the most important risk in the field, and it is underreported. Large language models do not retrieve authorised text. They generate plausible-sounding text based on patterns. In tax, that distinction is critical. A model asked about the Section 14C deduction or the arm's-length standard under the OECD Transfer Pricing Guidelines may produce a response that sounds authoritative and is factually wrong, citing the right statutory reference for the wrong provision, or confabulating a case that does not exist.
The danger is not that AI is uncertain. It is that AI is confidently wrong in ways that are difficult to detect without independent verification.
iLevvy's approach: does not rely on AI to generate statutory interpretation from scratch. The tax logic is hard-coded against verified statutory text and IRAS e-Tax guides, with AI used for classification and reasoning support within defined parameters, not as a freestanding legal authority. The system is designed to be sceptical of its own outputs.
6. AI Creates an Illusion of Completeness
A well-functioning AI system processes what it is given and produces a result that looks complete. It does not know what it does not know. If a preparer fails to flag that a subsidiary changed its functional currency mid-year, the system will not ask. If an intercompany transaction is miscategorised at source, the AI will compute correctly on wrong inputs.
Garbage in, garbage out is not new, but AI-assisted systems can make garbage look unusually clean and well-documented.
iLevvy's approach: includes a structured data completeness checklist as part of the engagement setup flagging common omission patterns (currency changes, related-party transactions, new plant additions, group structure changes) before computation begins. Completeness is the preparer's responsibility; the system is designed to help them discharge it.
7. AI Erodes Skill If Used Passively
When AI handles the mechanical work, practitioners who never learn why it does what it does become dependent on outputs they cannot critically evaluate. Over time, the profession risks producing reviewers who can approve or reject AI outputs but cannot independently produce them. This is a real structural risk, particularly for junior practitioners whose technical formation has historically depended on doing the mechanical work themselves.
iLevvy's approach: is designed around transparency, not black-box automation. Every output includes a visible computation trail. The intent is that practitioners using iLevvy understand what the system is doing and why, so that the tool builds technical fluency rather than replacing it.
8. AI Cannot Bear Professional Liability. The Practitioner Can
When a filing position is successfully challenged and penalties are assessed, no AI vendor absorbs the liability. The practitioner does. This is not an argument against using AI, it is an argument for understanding the allocation of responsibility clearly. AI is a tool. Professional judgement, sign-off, and accountability remain entirely human.
The risk is not using AI. The risk is using AI without maintaining the professional oversight that makes that use defensible.
iLevvy's approach: is built with the sign-off structure front and centre. Every judgment node, every override, every election requires an explicit preparer action. The system does not proceed silently. There is no ambiguity about who made the call.
Part III: The Bigger Picture
AI in tax computation is neither the revolution its proponents claim nor the threat its critics fear. It is a tool, a powerful one, with genuine capability to reduce the low-value hours, surface missed positions, and produce documentation that would otherwise not exist. It is also a tool with real failure modes: hallucination, false completeness, and the passive erosion of expertise in practitioners who never learn to question the output.
The question for practitioners is not whether to adopt AI-assisted tools. That decision is increasingly being made by competitive and economic pressure. The question is whether the tools they adopt are designed with honest scepticism about their own limitations or whether they are designed to look impressive while obscuring the risks.
iLevvy is built on the premise that the practitioner is always in charge. The system reasons alongside you. It does not decide for you. That distinction, between augmentation and automation, is the one that matters most.
This article reflects the views of the author and is intended for professional discussion. It does not constitute tax advice. iLevvy is a tax computation software product currently in development.
