The South African Revenue Service’s planned use of artificial intelligence could simplify tax compliance and improve enforcement, but experts warn that incomplete data, limited capacity and inadequate safeguards could undermine the project.
The real constraint on SARS in its quest to use AI is the quality and completeness of its data, rather than the sophistication of its software. Where third-party data is comprehensive — as it is for employment income, interest and retirement contributions — SARS is already close to the point where submitting a return becomes a formality, says Chetan Vanmali, a partner in the tax practice at Webber Wentzel.
However, where data is thin or fragmented, including in parts of value-added tax (VAT), customs and the informal economy, AI models have little to work with, he said.
“The commissioner's emphasis on VAT and customs modernisation and on digital identity (a single verified electronic identity for each taxpayer and representative across all dealings with SARS) is the right one for that reason.”
SARS outlines its AI plans
Speaking at the recent 13th Annual Tax Indaba in Sandton, SARS commissioner Dr Johnstone Makhubu spoke about the use of artificial intelligence, SARS’s growing debt book and progress in modernising the tax administration system.
He reiterated that SARS Modernisation 3.0 seeks to build an intelligent tax and customs administration platform powered by data, automation and AI, while upholding strong governance and public accountability.
Key initiatives include VAT, customs and excise modernisation, intelligent case management, digital identity capabilities and the use of AI to improve routine processes and service delivery.
While noting the importance of technology and adapting to changing demands, the commissioner stressed that technology is not a substitute for people, but a tool to improve outcomes for taxpayers, traders and SARS employees.
“The future of tax administration is not people versus technology, but it is people and technology working together to make compliance easier, services faster, and enforcement more precise. Technology must strengthen human capability, not replace it.”
How AI could help SARS
Vanmali said AI could help SARS prioritise its debt book, distinguish recoverable debt from debt that is disputed or prescribed, identify amounts owed by entities that no longer exist, and time its interventions more effectively.
However, he said AI cannot collect from taxpayers who cannot pay. A realistic use of the write-off and compromise provisions in the Act would do more for the quality of the debt book than any collection model.
Is SARS equipped to use AI?
Asked whether SARS is ideally equipped for this task, Vanmali said it is not — and that it would be surprising if it were.
He said SARS has rebuilt a great deal since the Nugent Commission and has received additional funding in recent budgets. However, it still competes with the private sector and offshore employers for data scientists and engineers, operates some legacy systems that were not designed for AI, and faces ordinary public-sector procurement constraints.
He said the commissioner’s identification of sustainable funding as one of four strategic priorities illustrates the pressure.
“Two capacity points are easy to overlook. First, if AI raises more assessments, the dispute machinery has to absorb more objections and appeals, and that side of SARS has to be funded too.
“Second, the more SARS depends on digital identity and third-party data, the more attractive a target it becomes for fraudsters and hackers, and cybersecurity funding has to be treated as part of the same AI investment.”
According to Vanmali, what the commissioner described at the Indaba — AI applied to routine processes, case management and service delivery — does not require a legislative amendment.
He said an assessment, penalty or refund decision must meet the same legal requirements, whether it was produced by a person or a computer. Existing law therefore already applies to the output of an AI system.
However, if SARS intends to move towards more extensive automated decision-making, an amendment would be prudent, even if it is not strictly necessary.
Vanmali said international experience illustrates why. He referred to a case in the United Kingdom in which taxpayers argued successfully in the Upper Tribunal that computer-generated notices had not been issued by a revenue officer as required by statute. Parliament subsequently legislated retrospectively in 2020 to validate them.
How South Africa can avoid litigation
Vanmali said South Africa can avoid similar litigation by amending the Tax Administration Act to expressly recognise automated processes, specify which decisions may and may not be taken without human review, and include the section 71 safeguards in the Protection of Personal Information Act (POPIA) in tax legislation.
“The draft Tax Administration Laws Amendment Bill, published in July 2026, does not address any of this. It does, though, oblige a bank that suspects a SARS refund paid into a client's account is linked to a tax offence to report it to SARS and freeze the money for up to two business days.
“It also gives a taxpayer less time after SARS refuses to remit an administrative penalty before SARS may start collecting it, namely 10 business days instead of 21. Both changes shorten the time between a system flagging a problem and SARS acting on it. However, neither change says anything about how that flag is generated, whether a person has to review it, or what the taxpayer is entitled to know about it. The Bill makes enforcement faster but says nothing about the automated systems behind it. That is what the legislation should address.”
Vanmali said SARS could also take several steps without legislative changes.
These could include:
- Publishing an AI governance framework explaining where automation is used and what its objectives are.
- Committing to having a human review every adverse decision before it is issued to the taxpayer.
- Disclosing in the letter to the taxpayer when a selection or decision was automated.
- Keeping an audit trail so that, when a dispute reaches objection or the Tax Court, the reasoning can be reconstructed.
- Testing its models for accuracy and bias against particular classes of taxpayer, and publishing the results.
Riana de Lange, digital communications lead, and Jashwin Baijoo, partner and head of strategic engagement and compliance at Tax Consulting SA, said in a joint statement that the effectiveness of AI depended on the quality of the underlying data, the systems with which it interacted, the quality of the models, the skills of the people using them and the governance surrounding their use.
They said feedback from tax practitioners at the Tax Indaba pointed to issues such as AI-generated responses being too general or failing to adequately address the particular circumstances of a taxpayer’s query.
They said AI could handle large volumes of information and identify patterns at a scale that humans could not, but complex tax matters still required context, judgement and the ability to engage with a taxpayer’s particular case.
They added that legislative and regulatory changes would likely be necessary in some areas, particularly where the modernisation of tax administration changed the information that taxpayers or businesses were required to provide, or the legal basis on which SARS could use and exchange data.
Tax Consulting SA said AI could be a powerful tool for identifying anomalies, assessing risk, detecting potential fraud and improving the efficiency of routine and repetitive processes.
“Internationally, we see that revenue authorities around the world are increasingly using AI in precisely these areas. But the fact that a decision or intervention is technology-enabled does not remove the legal obligations attached to that decision, for instance the obligation to provide reasons to the taxpayer or keeping them informed about their cases. The way taxpayer data is collected and used is also subject to legal requirements and must not be breached.”
given.majola@nationalmg.co.za