What SARS's responsible AI strategy actually asks of your tax team

Sarah Blignaut, deputy CTO at SARS, recalls a time when the volume of paper returns threatened the structural integrity of the buildings storing them. The number worth reading twice in her account, though, is not the paper. It is 300 million third-party data records processed a year to auto-assess 6.8 million taxpayers, more than 99% of whom accept the result without lifting a finger.

Most heads of tax read a story like that and jump straight to the enforcement question: when does the anomaly detection turn on us, and how do we defend a position against a model? This is reasonable, but it skips the part that arrives first and costs more. Auto-assessment does not work because SARS bought clever software. It works because employers, banks, medical schemes and retirement funds submit prescribed data, in prescribed formats, on a prescribed schedule, so the matching has something consistent to match.

That is the real consequence of a responsible AI strategy at a revenue authority: an authority that commits to explainable, governed AI cannot run it on inconsistent inputs, so the cost of its governance moves upstream to the taxpayer. Expect demands for standardised formats, comparable entity-level data and cleaner submissions long before you meet anything resembling an AI-driven assessment.

Tracy Potgieter, Co-founder and Director of vanbain Tax, sees the same shift from the tax function’s perspective.

“The direction of travel is clear: tax authorities are becoming more data-centric, and that means tax teams need to become more data-ready. The challenge is not simply producing more data. It is creating data that is structured, traceable and defensible.”

You can already see the pattern in the corporate lane. IT10B, CbCR schemas and Pillar Two returns each move in the same direction: the authority specifies the shape of the answer, and the work of getting fragmented local data into that shape sits with you. Blignaut is explicit that the transformation was organisational before it was technical, that friction became visible through technology but usually lived in process. The same is true on your side of the submission.

For Chris Cochrane, CEO of fluid, this is where tax technology needs to move beyond automation for automation’s sake. The objective is to create a structured data layer that can support multiple compliance obligations rather than repeatedly transforming the same underlying information for each submission. That means mapping and reconciling data once, retaining the logic behind it and reusing it across workflows.

“The real opportunity is not to automate individual tax returns. It is to create a trusted data foundation that can serve multiple tax obligations, with the rules and judgments applied on top,” he says.

“The advice from an AI perspective is don't try and eat the elephant whole. Start slicing off little pieces.”

Her second point matters even more for tax leaders. SARS staff are expected to keep the human in the loop, checking AI outputs rather than accepting them, because ungoverned AI can hallucinate. That is the identical boundary in-house tax operators and external tax lawyers keep landing on: AI can normalise variable inputs, map trial balances, extract financial statement data and reconcile sources, while a tax professional selects the treatment, makes the call and approves it.

Cochrane argues that this distinction is critical to responsible automation. “AI and automation can take on the repetitive work of extracting, classifying, mapping and reconciling data, but they should not remove the professional from the decision. The goal is to make the evidence behind that decision easier to assemble, review and defend,” he says. “Automation should not replace tax judgment. It should make the evidence behind that judgment easier to access, easier to validate and easier to defend.”

Judgment cannot be standardised. The inputs underneath it can, and that is the part now being asked of you.

When a request lands with roughly 21 business days to answer on events seven or eight years old, after the local contact has left and the drive has been replaced, the question is not whether your reasoning was sound. It is whether you can produce the facts considered, the rule applied, the reviewer and the linked support without rebuilding them.

This is where the value of structured tax data becomes tangible, says Potgieter. A defensible position is not just the conclusion reached at the time; it is the ability to reconstruct how that conclusion was reached, what information was considered and who reviewed it.

“Audit readiness is not something you create when the request arrives. It is the result of how you structure, reconcile and retain your tax data every day.”

So, take Blignaut's advice literally and slice something off. Pick the regime where the input pressure is closest, usually CFC data collection, validation and review or Pillar Two data readiness, and fix the collection, mapping and reconciliation once rather than per deadline. The same governed dataset then serves CbCR, transfer pricing and finance analysis, with only the rules and adjustments changing.

For Cochrane and Potgieter, the implication is broader than any single SARS process. As revenue authorities become more sophisticated in how they collect, match and analyse information, multinational tax teams need to think about data as part of their tax operating model, not simply as an input to individual compliance exercises.