How SARS uses machine learning for compliance

SARS has stated that it integrates advanced data analytics, machine learning and artificial intelligence into risk detection. Its 2024 to 2025 annual report explains that data from domestic and international sources is used to identify submissions that may be incomplete, nonexistent, misrepresented or fraudulent. The report also says that standard audit cases can be created automatically when sufficient third party information is available.

Earlier SARS reporting provides a particularly useful illustration. In a 2021 statement, SARS said that data from sources including banks, retirement funds, medical insurers, property records, company records, motor vehicle records and population information was being used as input for machine learning models, risk profiling and case selection. SARS reported that about 33 percent of R115 billion collected from compliance activities at that point resulted from automatic risk profiling using data and machine learning.

Supervised and unsupervised learning

Two broad categories help explain ML models for tax non-compliance. Together, these approaches form the practical foundation for ML models for tax non-compliance in a revenue authority.

Supervised learning tax compliance uses historical examples where the outcome is already known. For example, a training dataset might contain previous cases that were ultimately confirmed as compliant or non compliant. The model learns relationships between taxpayer characteristics and the observed outcome. It can then score new cases according to their estimated probability of non compliance.

This approach is useful when an agency has reliable historical labels. Possible applications include predicting whether a return may require verification, prioritizing audit cases, identifying potentially improper refund claims, or estimating the likelihood that a debt will remain unpaid. These are practical settings in which supervised learning tax compliance can support consistent case prioritization.

Unsupervised learning does not require a predefined outcome label. Instead, the system searches for structures, clusters or unusual observations within the data. In tax administration, this can help analysts find taxpayers whose behaviour differs substantially from comparable groups. It can also highlight new patterns that may not resemble previously detected cases.

SARS publicly describes machine learning models, risk profiling and detection capabilities, which is useful when considering ML models for tax non-compliance, but its public material does not provide a complete technical inventory of the individual algorithms used in production. Therefore, it would be inaccurate to claim that SARS specifically uses a particular classifier, clustering method or neural network unless SARS publishes that detail.

Common model types used by revenue agencies

Across government analytics, several model families are suitable for tax compliance prediction model development. Choosing the right tax compliance prediction model depends on the target outcome, data quality, legal context and need for explanation. Logistic regression is often useful when analysts need an interpretable probability of a defined outcome. Decision trees and ensemble methods can capture nonlinear relationships between variables and may perform well when many risk factors interact. In practice, a tax compliance prediction model should be assessed on both predictive performance and administrative usefulness.

Random forest and gradient boosting approaches are examples of ensemble methods that combine multiple decision structures. They can be useful for ranking cases by estimated risk. Neural networks can model complex relationships in large datasets, although their complexity can make explanation and governance more difficult.

Anomaly detection methods are especially relevant to unsupervised analysis. They can strengthen ML models for tax non-compliance when historical labels are limited. They can flag records that differ strongly from expected behaviour without requiring a historical label for every case.

Feature engineering from filing and third party data

A predictive system does not normally work directly from raw tax returns. Data scientists transform raw records into variables that a model can interpret. This process is called feature engineering.

For tax administration, useful features can be created from filing history, declared income, deductions, tax payments, refund activity, changes between filing periods and relationships between information reported by different sources. Third party information can provide additional context. For example, differences between reported income and information received from financial institutions or employers may become risk indicators.

Industry and peer comparisons can add another layer. A business can be compared with similar businesses using variables such as turnover, expense ratios, filing frequency or refund patterns. These comparisons should be designed carefully because legitimate businesses can have very different circumstances.

SARS has emphasized the value of third party information and data integration. During Filing Season 2024, it said nearly 150 million third party data records had been harvested and used with artificial intelligence and enhanced machine learning algorithms to support automatic assessments, improve return integrity and simplify the taxpayer experience.

From prediction to compliance action

Predicting tax evasion machine learning is only one part of the process. In practice, predicting tax evasion machine learning should be treated as a risk signal rather than a finding of wrongdoing. A risk score becomes useful when it supports appropriate administrative action.

A revenue authority can use model outputs to prioritize cases, request additional information, conduct verification, select audits, manage debt or identify potential refund risks. The objective should be targeted intervention rather than treating a model score as proof of wrongdoing.

SARS has described this approach through data driven risk detection and automated case creation. Its recent reporting also describes debt propensity models, AI driven risk profiling and refund fraud risk management as components of compliance activity. In 2025 to 2026, SARS reported R103 billion from tax verifications where risks were flagged through AI driven risk profiling and related analytical systems, alongside R94 billion from resolving outstanding debt cases supported by debt propensity models.

Data Tables & Compliance Indicators

The following figures are drawn from official SARS and OECD reporting and illustrate reported compliance outcomes and automation trends.

What this means for South African tax administration

The South African case shows that machine learning is most valuable when it is connected to a wider data ecosystem. Filing information alone may reveal inconsistencies, but third party information can provide additional evidence for comparison. Historical compliance outcomes can support supervised models, while unsupervised methods can help identify unfamiliar patterns.

For taxpayers, better risk targeting can also reduce unnecessary intervention. This is another reason predicting tax evasion machine learning should be paired with review and evidence. SARS has repeatedly stated that its technology investments are intended not only to detect non compliance but also to make compliance easier for taxpayers who are meeting their obligations. Its Auto Assessment programme is an example of how data integration and automation can reduce the work required from individuals.

Governance, fairness and human oversight

Data should have a clear administrative purpose and should be protected against unauthorized use. Models should be monitored for performance and unintended bias. Officials need procedures for reviewing questionable predictions and correcting inaccurate information.

This is why supervised learning tax compliance should be treated as decision support rather than an automatic substitute for professional judgement. A fourth use of supervised learning tax compliance is therefore not another model, but a disciplined process for reviewing model outputs. Human officials should be able to examine relevant evidence, challenge model outputs and apply the applicable law.

The future of predictive compliance in South Africa

SARS is continuing toward a more integrated digital administration. Its Modernization 3.0 material describes an ambition to use data science, artificial intelligence and advanced analytics to improve voluntary compliance, efficiency and transparency. Its 2026 VAT modernization initiative also describes AI enabled risk detection and analytics as part of a future digital VAT ecosystem.

For organizations studying machine learning tax compliance South Africa, the key takeaway is that successful implementation depends on more than model accuracy. In a wider global context, machine learning tax compliance South Africa also illustrates how local data and institutional design shape implementation. It requires high quality data, appropriate features, strong validation, clear governance, skilled officials and a careful connection between predictions and lawful administrative action.