Writing & perspectives

AI and tax

Where AI Changes Tax Advice—and Where Professional Judgment Remains

AI can accelerate research, evidence review and drafting in tax work. Dependable advice still requires a workflow that preserves context, authority, challenge and accountable professional judgment.

Tax advice is an attractive setting for artificial intelligence. The work is document-intensive, rules-based in important respects and often constrained by time. AI can retrieve authorities, compare documents, identify relevant facts, test scenarios and produce a first draft quickly.

Those capabilities are useful, but they do not turn tax advice into a simple question-answering task. A conclusion may depend on precise facts, the interaction of several provisions, the status and date of an authority, the purpose of a transaction and the distinction between what is technically arguable and what is appropriate to recommend. The quality of the work therefore depends on more than whether an AI response sounds plausible.

The relevant question is how AI changes the workflow through which evidence becomes professional advice—and whether the resulting process gives the adviser better grounds for judgment.

Begin with the instruction and the decision

A research note, a transaction opinion, a compliance review and an initial client discussion are different forms of work. They involve different consequences, evidence and standards of review. The same AI capability may be suitable for one and inadequate for another.

The workflow should begin by defining what the output is for. Is it intended to identify issues, support a calculation, compare possible treatments or form part of advice on which a client will act? Who will rely on it, and what would a material error or omission change?

This boundary matters because AI can easily expand a task. A request to summarise a provision can become a confident view on its application. A search for relevant authorities can become a recommendation without an explicit decision about whether the factual record is complete. Good workflow design keeps the purpose, scope and limitations visible.

Separate facts, authority and interpretation

Tax reasoning draws on several kinds of input: client facts and documents; legislation and other authoritative material; case law and guidance; calculations; and professional interpretation. AI may assist with each, but assurance should not treat them as interchangeable.

An extracted fact can be checked against its source. A statement of law should be traceable to the relevant authority and considered in its proper date and context. An interpretation needs a reasoned connection between the facts and the authority. A recommendation also has to reflect uncertainty, commercial objectives and the consequences of taking a position.

The handoffs between these layers are where apparently minor weaknesses can become material. An omitted fact can make otherwise accurate research irrelevant. A superseded source can support a coherent but outdated analysis. A model can quote the right passage while applying it to a transaction with a different legal or economic character.

A dependable workflow should therefore preserve provenance. Material statements need sources that the reviewer can inspect. Assumptions should be distinguished from established facts. Where authorities conflict or their application is uncertain, the disagreement should remain visible rather than being compressed into one confident answer.

Use AI to widen inquiry, not conceal uncertainty

AI can help an adviser explore a broader set of possibilities. It can identify terms in contracts, compare transaction steps, generate alternative characterisations and locate issues that deserve further research. This can improve coverage and release time for more demanding analysis.

But a longer list of issues is not necessarily a better analysis. Some points will be immaterial; others will depend on facts that are not yet known. The professional contribution includes deciding which questions could change the advice and what additional evidence is proportionate.

Uncertainty should be treated as information. The adviser may have high confidence about the operation of a rule but lower confidence that the available facts place the client within it. A conclusion may be technically supportable while remaining sensitive to a valuation, intention or future event. Presenting these differences clearly is more useful than offering an undifferentiated confidence score.

Define meaningful professional review

“Reviewed by a tax professional” is not a sufficient description of a control. Review needs a defined purpose and the information required to perform it.

A reviewer may need to verify extracted facts, check that authorities are current and relevant, challenge the legal reasoning, recalculate an amount, test an alternative interpretation or assess whether the recommendation is suitable for the client. These are distinct activities. A polished draft can make them harder if it encourages line editing rather than independent challenge.

The order of work matters too. Showing an AI conclusion before the adviser has framed the issue may anchor the analysis. Requiring complete independent re-performance may reduce that risk but remove much of the efficiency. A stronger design might separate factual extraction and research support from the formation of the professional conclusion, with targeted independent work on the judgments that carry the greatest consequence.

The appropriate arrangement should be tested against the actual work. It should examine errors and omissions that survive review, the effort required to correct them and the circumstances in which AI changes the adviser's judgment.

Retain a record of how the conclusion was reached

Professional advice should be capable of reconstruction. Where AI materially contributes, the record should make clear what information was supplied, which sources were relied upon, what output influenced the analysis, what checks were performed and where the professional accepted, changed or rejected the AI contribution.

This does not require preserving every exploratory interaction. It requires preserving the material route from evidence to conclusion. That record supports quality review, later updates and an explanation to the client. It also makes recurring weaknesses in the workflow easier to identify.

Evaluate the combined workflow

Model-level testing remains important. Organisations need to understand factual error, sensitivity to prompts and documents, consistency, security and the limitations of the systems they use. But those tests do not establish the quality of tax advice on their own.

The more useful comparison is between complete processes: current expert practice, AI support under defined conditions and the proposed human–AI workflow. Evaluation should consider factual and legal accuracy, material omissions, traceability, the quality of professional challenge and the total effort required to reach an advice-ready result.

Performance should also be examined across different kinds of case. A workflow that helps with routine research may fail where facts are incomplete, authorities conflict or several taxes interact. These difficult cases are precisely where professional judgment is most valuable.

The opportunity is better professional work

AI can make tax work faster, but speed is not the only or necessarily the largest opportunity. It may improve the organisation of evidence, expose an alternative interpretation earlier, make assumptions easier to challenge or enable advisers to spend more time on the questions that matter to the client.

The aim is not to add a human signature to a machine-produced answer. It is to design a workflow in which evidence is easier to inspect, reasoning is easier to challenge and the final judgment has a clear professional owner.