Bringing explainability (XAI) to AI in tax
“Why did the system decide to apply a reduced VAT rate to that transaction?”
As the future of tax technology will increasingly lean on AI and machine learning, we look at how we bring explainability (and with it, auditability) to AI in tax.
The rise of AI in tax determination
Tax determination used to be built on hard-coded rules: a matrix, often with two key parameters. While rigid, this gave tax teams a certain comfort. They could trace every decision through rule tables and logic trees.
But often, the business needs something more – an added parameter, or custom rulesets, and so traditional determination frameworks gets stretched. Enter machine learning.
AI models promise smarter, faster tax determinations. They can:
- Learn from historical transaction data to predict tax treatment,
- Identify unusual patterns that might signal errors or fraud,
- Handle nuanced scenarios where traditional rules fall short.
Sounds like a dream. Except for one problem: a risk of black box decisions(?). How do we convince a tax authority under audit that the decisions made were the right ones, and that we know how they were made and why? How do we minimise the risk of hallucination, or identify it where it has taken place?
The black box problem
Certain AI systems, especially those based on complex neural networks, can produce results without an obvious explanation. They might tell you:
“Transaction A is standard rated VAT.”
…but not why.
For tax professionals, that’s a showstopper. We work in a field where every decision must stand up to:
Auditors: Prove how you reached your conclusion.
Regulators: Demonstrate compliance and proper treatment.
Business stakeholders: Answer, “Why are we paying this much tax?”
Without explainability, trust evaporates. Even the most accurate model becomes unusable if no one can defend its decisions.
Enter explainable AI (XAI)
This is where explainability (XAI) comes in. It’s a field dedicated to making machine learning transparent.
In the tax world, XAI can help:
- Highlight the key factors that drove a tax determination decision.
- Visualise how different inputs (e.g. product type, jurisdiction, customer status) influenced the result.
- Offer human-readable explanations for each decision.
Imagine if your tax engine could tell you:
“Applied reduced VAT because the product is classified as medical equipment and the buyer is in Country X, where medical goods attract a lower rate.”
That’s actionable and auditable.
Techniques for explainability
Several tools and techniques have emerged (and are emerging) to help:
- SHAP values (SHapley Additive exPlanations): Break down each prediction and show how much each input contributed. For example, SHAP might reveal that “product category” influenced 70% of the VAT outcome.
- LIME (Local Interpretable Model-agnostic Explanations): Creates simplified models around specific predictions to explain them in plain terms.
- Rule extraction: Converts black box models into a set of understandable rules that approximate the original model’s behaviour.
These aren’t perfect. Sometimes explanations can themselves become complex. They come with their own assumptions and limitations. But they’re a critical step forward in bridging tax expertise with AI.
Why this matters now
Global tax authorities are increasingly digitally-savvy. Real-time reporting, e-invoicing, and data-driven audits are becoming the norm. If your tax determination logic can’t be explained, you’re exposed to compliance risks.
Moreover, transparency builds trust with business stakeholders. In regulated industries like tax, explainability often wins. It’s better to be slightly less predictive than unable to justify your decisions to tax authorities or auditors. Layered with human tax experts, we supplement the AI driven results.
Without explainability, AI in tax risks remaining an interesting technology, but ultimately sidelined by tax leadership due to overarching concerns on its accuracy and auditability.
The road ahead
AI will undoubtedly play a growing role in tax determination. But unless we solve the explainability problem, tax professionals will continue to ask:
“Can we really trust the machines?”
The answer lies in blending cutting-edge AI with rigorous, transparent explanations. It’s not just good technology. It’s essential for keeping tax compliant and defensible in an increasingly complex world.
Is your tax technology ready to explain itself? Or are you still facing black boxes in your determination logic? Get in touch (directly or in the comment section) – we’d love to hear how ready your team is to embrace AI, and the challenges you’re seeing in bringing this tech to the table.
P.S. If you’d like to learn more about XAI, I found these videos helpful
– https://www.youtube.com/playlist?list=PLV8yxwGOxvvovp-j6ztxhF3QcKXT6vORUl.