AI Opportunities

Tax & Revenue Administration

AI can help tax administrations protect revenue earlier — improving the decisions that determine who should be registered, which returns warrant intervention, whether a refund should be released, and where scarce officer capacity should be focused.

The biggest opportunity is not finding more AI use cases. It is improving the decisions that determine whether revenue is collected, lost or recovered.

taxrevenue integrityVATGSTfraude-invoicing
← All AI Opportunities
Tax & Revenue Administration

The opportunity

Move revenue integrity upstream

Tax administrations already have rules, controls, audit programmes and experienced investigators. The opportunity for AI is not to replace them. It is to bring additional context to decisions while there is still time to change the outcome.

That distinction matters. An incorrect refund identified before release can be stopped or verified. The same refund identified afterwards becomes a recovery problem. Tax charged by a trader that disappears may never become a refund at all; the opportunity is to identify the trader or supply chain before the unpaid liability becomes uncollectable debt.

This changes the question from where can AI be used? to where are we discovering risk after the opportunity to prevent the loss has passed?

The opportunity is not more detection. It is earlier action.

VAT & GST illustrates the scale

Revenue loss is large, adaptive and not confined to fraud

€128bn

2023 EU VAT compliance gap cited by the European Commission.

~€50bn

Estimated annual carousel-fraud losses across EU Member States.

$2bn

GST refunds paid in the Australian fraud case before containment, with a further $2.7bn stopped.

Fraud is only part of the revenue-integrity problem. Insolvency, honest error, complexity and administrative error also contribute to the gap. The common requirement is better evidence at the point where an administration can still act.

Where AI can create value

Four shifts can change the economics of revenue administration

The specific tax, system and legal context varies by administration. But across the lifecycle, the highest-value opportunities tend to share one characteristic: better information changes a consequential decision.
01

Move integrity upstream

Identify suspicious registrations, trading relationships, returns and refund claims before loss becomes a recovery problem.

02

See the taxpayer in context

Connect taxpayer history, suppliers, customers, invoices, shared identities and prior findings to reveal risk that is invisible in a single return.

03

Put judgment where it matters

Prioritize verification, audit and investigation so scarce officer capacity is spent where intervention has the greatest expected value.

04

Turn outcomes into learning

Capture what officers discover and feed those outcomes back into rules, models, network analysis and future decisions.

Exhibit 1 · The decision map

A tax return is not one decision

Between registration and recovery, an administration makes a sequence of decisions. AI has value where additional context can change one of those decisions — particularly before funds leave or debt becomes difficult to collect.
01
Registration Is the business genuine and entitled to register and claim?
02
Invoicing Is the invoice valid, reported and consistent with the trading relationship?
03
Return & assessment Is the return complete, correctly treated and consistent with known rules?
04
Pre-refund integrity Refundable — but does the wider context warrant verification?
05
Release or intervene Release, hold, request evidence, verify or apply an appropriate control.
06
Audit & recovery What was wrong, what can be recovered and what should the administration learn?
Once an incorrect refund is paid, prevention has become recovery.
Exhibit 2 · Context changes the decision

Valid on its own. Different in context.

Rules answer explicit questions well. The harder opportunity sits between them: activity that passes individual checks but looks different when relationships, history and peer behaviour are considered.

What the transaction says

  • The invoice exists.
  • The return is correctly formed.
  • The arithmetic reconciles.
  • No known rule is broken.
  • The claimant looks plausible in isolation.
→

What the network may reveal

  • The supplier charges tax but never remits it.
  • Goods or invoices circle through connected traders.
  • Multiple registrations share bank or identity attributes.
  • A new trader claims credits far outside plausible scale.
  • Behaviour changed abruptly before the claim.
For some schemes, the unit of risk is the chain — not the trader.
Exhibit 3 · Modernisation changes the evidence

E-invoicing makes AI more useful — not unnecessary

E-invoicing and e-reporting move administrations from periodic return totals toward structured transaction-level evidence. Some questions become deterministic. Others become richer contextual problems.

Returns only

Evidence arrives periodically and is often aggregated.

Matching may happen after the period. Supply-chain behaviour has to be inferred across returns. Peer groups are broad and intervention may come weeks later.

→

With transaction-level evidence

Invoice relationships become visible earlier.

Matching can become a rule. Network features become richer. Peer comparisons become more precise. Risk can be assessed while a short-lived chain is still active.

Validation asks whether the invoice is well formed. Integrity asks whether the trade was real.
Exhibit 4 · Different AI for different work

Predictive, network and generative AI play different roles

There is no single tax-fraud model. Different decisions require different techniques, and generative AI is most valuable where the work is reading, writing and assembling evidence rather than deciding whether activity is fraudulent.
Discover
Network & anomaly analysis Find unusual traders, relationships, chains and emerging patterns across the population.
Detect
Rules & predictive models Score the current decision using return, taxpayer and network context.
Respond
Generative & agentic assistance Assemble cases, explain reason codes, draft evidence requests and summarize submitted material.
Learn
Outcome feedback Turn verified officer outcomes into labels and candidate patterns for future decisions.
AI should return officer hours to judgment — not take the judgment.

Where to start

Start with the decision where earlier action changes the outcome

The strongest AI opportunity is rarely the broadest one. It is a specific decision where the administration has relevant evidence, an actionable intervention and a measurable economic or operational outcome.

For VAT and GST, the pre-refund integrity decision is a powerful example: release, hold or verify while the refund is still actionable. For another administration, the better starting point may be registration, audit selection, debt prioritisation or taxpayer error prevention.

The discipline is the same: find where value is currently leaking, identify the decision that controls it, and determine whether better intelligence can materially change the outcome.

Discuss the opportunity

Related reading

All insights →
Tax & Revenue Administration

Insights

From Periodic Returns to Continuous Transaction Intelligence

For decades, VAT and GST administration has been organized around the return. Businesses record transactions, aggregate them into periodic declarations and file. E-invoicing and electronic receipts now change that sequence. They make each transaction visible while it is still current, and AI turns that visibility into service, compliance and enforcement decisions taken while economic activity is still moving. The return is becoming the conclusion rather than the beginning.

Tax & Revenue Administration

Insights

From Universal Scrutiny to Adaptive Compliance

Tax authorities have always treated noncompliance differently depending on the case. What is changing is the speed, precision and breadth with which they can do it. As e-invoicing and digital evidence make economic activity visible, AI lets authorities move away from broad controls applied after filing and toward treatments shaped by the taxpayer, the transaction and the reason a risk emerged. The shift is from universal scrutiny to adaptive compliance.

Tax & Revenue Administration

Insights

The VAT Refund Is Becoming a Real-Time Decision

A VAT refund is often treated as the final administrative step after a return is filed. In reality it is one of the most consequential decisions a tax authority makes. Pay too slowly and legitimate businesses finance government uncertainty. Pay too readily and fabricated transactions become public money. Digital transaction evidence and AI are turning that blunt trade-off into a precise decision about what can be released, what must be verified and what warrants investigation.