Artificial Intelligence comes of age for VAT
Many tax professionals have assumed (errr…hoped) was simply too jurisdiction specific, too nuanced, too dependent on context for machines to handle well.
That assumption has shifted quickly. In the last 6 to 12 months, especially with the latest version of the coding advanced in Claude last month,
AI has improved at a pace that is hard to ignore, particularly in coding and structured analysis. Tasks like extracting facts from documents, drafting memos, comparing treatments across countries, or spotting potential risk issues increasingly take minutes rather than hours.
It is tempting to frame this as the rise of autonomous agents: systems that plan, decide and execute tax work end to end with minimal human involvement. But that framing misses where progress is actually coming from. The step change is not primarily “smarter models”. It is smarter workflow design around the models.
VAT determinations are rarely a single decision. They are a chain. Classification feeds place of supply. Customer status and evidence requirements affect liability. Exemptions and special schemes distort the “default” logic. Invoice data quality and system mappings determine whether the right rule even gets applied. When all of that is collapsed into one prompt or one black box decision, errors become invisible until they surface downstream in returns, reconciliations or audit queries. The model can sound right while being wrong.
Structured, multi step workflows change the economics. Planning before execution. Data gathering before synthesis. Drafting followed by review and revision. Intermediate checks that force the system to show its working and give humans clear points to intervene. In practice, this approach can make even less capable models perform better than newer models used in a single step. The workflow, not the model, becomes the main lever.
Examples of where VATCalc has been early adopters of AI evidence this:
- Accuracy in item classification. AI can suggest VAT categorisations quickly, but always within the boundaries of expert-maintained global VAT data.
- Confidence in advice. Advisor generates draft guidance that reflects verified content, minimising the risk of hallucinations or incomplete analysis.
With tax’s zero-tolerance for error, the real constraint is reliability
This matters because the real constraint in indirect tax is not intelligence. It is reliability. Tax has little tolerance for probabilistic behaviour. A small error repeated across high transaction volumes becomes a material exposure, and multi step agent systems compound failure risk at each stage. Moving from “often correct” to “trustworthy at scale” is the hard part. That last mile is governance: testing, auditability, exception handling, and controls that make outcomes explainable.
This is where VAT tax engines still sit at the centre. Deterministic rules engines remain essential because consistency is non negotiable. AI does not replace that core. It augments it. The most valuable near term applications sit around the edges: accelerating implementation work through faster coding and test generation, reconciling ERP outputs against engine determinations, flagging anomalies for human review, monitoring regulatory change and highlighting the likely impact on configured logic.
So rules engines stay central with humans in the loop
The practical takeaway is straightforward. Human in the loop is not a temporary compromise. It is the operating model. Clean source data, a single repository of truth, and workflows that keep knowledge in the system rather than in inboxes are what make AI powerful rather than risky. Tax will not disappear. Judgment, governance, controversy management and strategic structuring remain deeply human. But the cost base of tax knowledge work is shifting fast, and the teams that redesign their VAT workflows now, with reliability as the goal, will compound advantage.
