How AI helps revenue authorities close the tax gap

What the tax gap is

The OECD describes the tax gap as the difference between the tax revenue that could be collected and what is actually collected. It splits that gap in two. The compliance gap comes from taxpayers who do not register, file, declare or pay as the law requires. The policy gap comes from choices written into the law, such as exemptions and reduced rates.

The IMF’s Revenue Administration Gap Analysis Program applies the same idea to VAT, comparing potential VAT collections with what was actually collected. Legislation sets the policy gap. The compliance gap is the part a revenue authority can narrow through its own work, and it is where leakage detection matters.

Where revenue leaks

Revenue authorities describe the same kinds of leakage in their own reports. South Africa’s revenue service, for example, names businesses that are not registered, taxpayers who do not file, false declarations, undisclosed income and fraudulent refund claims, and on the customs side under-declared values and misclassified goods. Its 2024/25 annual report says it prevented impermissible refunds and revenue leakage of R147.9 billion that year.

Most leakage falls into a handful of patterns:

• Trading without registering for the taxes that apply
• Declaring less income or turnover than third-party records show
• Claiming refunds, input tax or deductions that are not owed, often through fictitious invoices or connected companies
• Withholding tax from employees or customers and not paying it over
• Declaring imports below their value or under the wrong tariff heading
• Stopping filing while the business keeps trading

Capacity, not data, is the constraint

Authorities already receive a great deal of information from others. In South Africa, section 26 of the Tax Administration Act lets the Commissioner require third-party returns from anyone who employs, pays or holds assets for another person, and section 46 allows a request for relevant material from a taxpayer or a third party. Banks, employers, company and property registers and customs declarations all describe what taxpayers actually do.

The limit is how many cases people can work. The OECD’s guidance on compliance risk management, published in 2004, sets out the discipline that follows from it: identify the risks, assess and prioritise them, choose the right treatment for each and evaluate the outcome, so that limited resources go where they make the most difference.

What AI adds

AI is useful here because the work is matching and ranking at scale. Revenue authorities are moving in this direction. The South African Revenue Service’s strategic plan for 2025/26 to 2029/30 describes using AI to predict which sectors and entities are at highest risk of non-compliance, and its annual report describes a machine-learning risk engine that scans internal and third-party data. In practice that means:

• One view of each taxpayer, with registrations, returns, payments, customs declarations and third-party records linked to the right person or business
• Comparison of what each taxpayer declares against third-party data and against peers of the same sector and size
• Network analysis that links refund claimants, suppliers, directors and bank accounts, so a syndicate is seen as one case
• An estimate of the revenue at risk in every flag, so the largest and best-evidenced leaks are worked first

From detection to a plan that reaches the target

Finding leakage is half the job. Revenue authorities answer for collections against an agreed estimate. SARS, for instance, measures net revenue collected as a percentage of the estimates agreed with the Minister of Finance. The practical question is which interventions to run, in what order, with the people available.

Different leaks call for different treatments. Small gaps across many taxpayers may respond to a well-worded letter, and the OECD’s 2021 guide on behavioural insights shows how tax administrations test letter wording with randomised trials. Larger and better-evidenced gaps justify audits. Risky refunds can be held: section 190(2) of the Tax Administration Act allows SARS to withhold a refund until a verification, inspection, audit or criminal investigation is finalised. Customs leakage calls for targeted inspections.

A good system compares these options by expected yield and effort, recommends a mix that fits the capacity available and tracks progress against the target. It then records what each intervention actually produced, so the next plan rests on the authority’s own results rather than assumptions.

Safeguards that make it work

Taxpayer information is protected by law. Chapter 6 of the Tax Administration Act requires SARS officials to preserve the secrecy of taxpayer information, and other jurisdictions have equivalent rules. That favours systems that run on infrastructure the authority controls, with access by role and every access logged.

Selections must also be defensible. A taxpayer, a tribunal or an oversight body can ask why a case was chosen, so every flag should show the declarations, third-party records and comparisons behind it, and every decision should rest with a named official. In SARS’s own words, the aim is to make compliance easy and to make non-compliance hard and costly. AI serves both aims when its findings are explained and its recommendations stay with the people accountable for them.