The most important near-term use of artificial intelligence in tax administration is not a machine issuing a final tax judgment.
It happens earlier.
Tax authorities can use data analytics and AI to connect records, identify anomalies, rank risks and decide which cases deserve human attention. That changes the economics of enforcement even if the final assessment, audit decision or legal conclusion remains with an official.
The practical shift is from finding a problem manually to using machines to decide where humans should look first.
Key takeaways
- AI is already widely used for tax risk assessment and fraud detection. The strongest documented use case is triage, not autonomous final decisions.
- The mechanism is data matching → anomaly detection → risk scoring → case selection → human review. Better targeting can make limited audit resources more effective.
- The same tools create governance risks. Poor data, opaque models and false positives can direct scrutiny toward the wrong taxpayer, making explainability and human review essential.
Facts: AI is already inside tax administration
OECD data show that AI is no longer experimental for many tax administrations.
In its 2025 work on digital transformation, the OECD reported that 72% of surveyed tax administrations were using AI. Among administrations using AI, 74.4% reported use in detecting tax evasion and fraud and 64.1% in risk-assessment processes.
The same survey contained an important limit: none of the participating administrations reported using AI to make final administrative decisions.
That distinction should define the debate.
The immediate question is not whether a machine replaces the inspector. It is whether software changes which taxpayer reaches the inspector’s desk.
Austria: predictive analytics as an audit-selection layer
Austria provides a concrete example.
The Federal Ministry of Finance operates a Predictive Analytics Competence Center, or PACC. Its official description includes real-time risk assessment, centralised audit-case selection, automated checks and analytical support for tax and customs work.
In August 2025, the Ministry reported that PACC risk models had examined around 6.6 million cases during 2024 and that audit activity connected with those models resulted in approximately €354 million of additional tax revenue.
The Ministry is also explicit about the institutional boundary: PACC provides information on high-risk cases, while tax adjustments are imposed only by the responsible audit bodies.
That is the model in its clearest form:
machine-assisted selection, human administrative action.
Greece: cross-checking, profiling and prioritisation
Greece provides a second example.
The Independent Authority for Public Revenue has publicly described AI as useful for detecting tax evasion and smuggling, prioritising audits, predicting violations, analysing business tax behaviour, identifying suspicious transactions, carrying out mass cross-checks and applying risk-analysis techniques.
Its broader digital infrastructure reinforces the mechanism. The myDATA platform digitises business transaction information, while the authority’s new integrated tax-information system explicitly includes risk analysis, compliance, case assignment and support for audits.
The point is not that every Greek taxpayer is being “audited by AI”.
The point is that the administration is building a much denser information environment in which machines can help determine what deserves attention.
Interpretation: the audit begins before the audit
Traditional tax enforcement has always involved selection.
No administration can inspect every taxpayer manually. Authorities therefore decide where to concentrate staff.
AI changes the cost and scale of that selection process.
A system can compare historical filings, payment information, third-party records, transaction data and other information across far more cases than a human team could review individually.
The important change is therefore procedural.
The audit can effectively begin with a risk model long before a taxpayer receives an audit notice.
The mechanism
The transmission mechanism is:
data collection → entity matching → anomaly detection → risk indicators → prioritisation → human case selection or review → formal compliance action.
Each stage matters.
Matching determines whether records referring to the same person, company or asset are correctly linked.
Anomaly detection identifies patterns that differ from an expected baseline.
Risk scoring prioritises those signals.
Human officials then decide whether the signal justifies a question, intervention or audit.
A model does not need to understand the whole tax law to change enforcement materially. It only needs to improve — or appear to improve — the allocation of human attention.
The strongest countercase
Risk models can be wrong.
The OECD identifies poor data quality, bias, lack of transparency and weak explainability as material risks in tax-administration AI.
An internationally mobile person may legitimately have addresses, income sources, companies and bank relationships in several jurisdictions. A restructuring can create a large statistical discontinuity without implying abuse. A misspelled identifier can create a false mismatch. Historical enforcement data can embed historical bias.
A model optimised only for detecting unusual behaviour can therefore confuse complexity with non-compliance.
That is why the absence of final autonomous decisions in the OECD survey is important.
Human review is not merely a temporary inconvenience. It is part of the legitimacy of the system.
Scenarios, not forecasts
Under better-governed triage, tax authorities use AI to narrow large datasets while maintaining documented human review, audit trails and challenge procedures.
Under black-box targeting, increasingly complex models influence who is selected without taxpayers or even officials being able to explain the signal adequately.
Under contestable automation, administrative law, courts and governance frameworks force authorities to provide stronger transparency, validation and correction mechanisms.
Observable triggers include published AI-governance frameworks, audit-selection rules, court cases, data-protection decisions and official reporting on model performance.
Practical consequences
The practical response is not to make a tax profile look statistically ordinary.
It is to make the underlying facts coherent and explainable.
Residences should be documented. Company ownership and management should be consistent with filings. Major movements of capital should have an evidence trail. Third-party reporting and internal records should not tell contradictory stories without a legitimate reason.
International complexity is not itself a problem.
Unexplained inconsistency is.
The more efficiently authorities can identify anomalies, the more valuable good records become before any formal audit starts.
Sources
- OECD, Tax Administration Digitalisation and Digital Transformation Initiatives, 2025: https://www.oecd.org/en/publications/tax-administration-digitalisation-and-digital-transformation-initiatives_c076d776-en.html
- OECD, Tax Administration 2025 — Compliance Management: https://www.oecd.org/en/publications/tax-administration-2025_cc015ce8-en/full-report/compliance-management_988ac964.html
- OECD, AI in tax administration, 2025: https://www.oecd.org/en/publications/governing-with-artificial-intelligence_795de142-en/full-report/ai-in-tax-administration_30724e43.html
- Austrian Federal Ministry of Finance, Predictive Analytics Competence Center: https://www.bmf.gv.at/en/topics/combating-fraud/anti-fraud-units/pacc.html
- Austrian Federal Ministry of Finance, €354 million additional tax revenue through AI methods, 13 August 2025: https://www.bmf.gv.at/presse/pressemeldungen/2025/august/pacc-ki.html
- Greece IAPR, IAPR at BEYOND Expo 2025: https://aade.gr/en/news/events/ekdiloseis/iapr-beyond-expo-2025-metropolitan-expo-45642025
- Greece IAPR, myDATA — IAPR e-books: https://www.aade.gr/myDATA-IAPR-e-books
Disclaimer
This Insight provides general information and analysis. It is not tax, legal, data-protection or administrative-law advice. Tax administrations use different systems, legal safeguards and data sources, and the treatment of any taxpayer depends on the applicable law and facts.
