Tax functions are built on controls. Teams know where numbers come from, how calculations are performed and how results can be reproduced when an auditor asks. Those principles still apply once AI is in the picture. Large language models are probabilistic. Their outputs can vary, and it can be hard to trace which fields were used, how data was joined and why a conclusion was reached. In tax, that traceability is a compliance requirement. A transfer pricing position, a Pillar Two calculation or a VAT return can be examined years later, and the team has to reconstruct how each figure was produced.
The practical question this session asks is where AI genuinely improves tax work, and where deterministic, reproducible analytics remain the better choice. A mature tax function knows the difference.
A five-level maturity model
Level 1: Experiment. Tax professionals use AI individually for quick wins like summarising a ruling, drafting a memo or researching a treaty question. The gains are real, but results vary, confidential client or company data can be exposed, and there is little record of how an output was produced.
Level 2: Understand. The tax team starts to separate different uses of AI and analytics. Summarising new legislation with an LLM is a different task from asking it to map transactions to VAT codes or calculate an effective tax rate. Teams learn what happens between source data and conclusion.
Level 3: Control. AI and analytics become part of controlled processes. Data sources are defined, access is managed, and procedures are documented. Outputs can be traced back to source data. This is where AI use starts to fit inside a tax control framework. The question shifts from “can AI do this?” to “can we show how this result was produced?”
Level 4: Reproduce & Validate. The same source data and the same logic produce the same result. Data lineage is visible, transformations are documented, and AI-generated scripts and conclusions are validated before the team relies on them. Sometimes the more mature decision is to keep the calculation itself deterministic. Tax provisions, GloBE calculations and transfer pricing adjustments are typical examples.
Level 5: Rely. AI, deterministic analytics and human judgement work together. Automation runs where processes are controlled, AI is used where it adds value, and deterministic analytics handle the work that needs precision and auditability. Tax professionals stay responsible for challenging conclusions and deciding whether a result is reliable enough to file. The goal is controlled reliance. The person who signs off a return stays accountable for every number in it.
Six questions every tax team should be able to answer
Reproducibility: will the same data and logic produce the same outcome tomorrow, and in three years when the authority asks?
Audit trail: can we reconstruct the steps, filters, calculations and transformations behind the result in a form a tax auditor could follow?
Data lineage: do we know which source systems, files and fields fed the calculation or the return?
Explainability: can a tax professional explain the outcome to an authority without saying “the AI found it”?
Data integrity: can we show that joins, mappings, aggregations and exceptions were handled correctly? Think of GL accounts mapped to tax codes, or intercompany transactions matched across entities.
Accountability: who owns the outcome and who signs off when AI produces a plausible but wrong answer?
What you’ll take away
A framework for deciding where AI belongs in the tax function
A clear line between deterministic and AI-driven analytics
A view of which tax calculations should stay deterministic
Six questions for judging whether an AI-driven analysis can be relied on in a filing or an audit
A view of maturity measured by how well tax teams understand, control, reproduce and explain their numbers, rather than how much work is handed to AI