Tax authorities have always differentiated their response to noncompliance. What is changing is the speed, precision and breadth with which they can do so. As e-invoicing, electronic receipts and other digital evidence make economic activity more visible, AI is allowing authorities to move away from controls applied broadly after filing and toward treatments shaped by the taxpayer, the transaction and the reason a risk has emerged.
The strategic question is not how to identify more risk
The conventional case for AI in tax administration begins with detection. Better models promise to find more anomalies, rank more cases and expose more fraud. Those capabilities matter, but they start halfway through the problem. Before deciding how to detect risk, an authority must decide what it is trying to change in the behavior of taxpayers and in its own response.
Most taxpayers do not encounter the tax authority as an enforcement institution. They encounter it when they register, issue an invoice, interpret a rule, file a return, claim a refund or correct a mistake. At each point, the authority influences whether compliance is easy, uncertain or needlessly expensive. A system designed only to detect failure after these interactions has missed the opportunity to prevent it.
This is why adaptive compliance is more than risk-based audit selection. It is an operating model in which the authority changes its treatment according to the evidence, behavior and consequence surrounding a decision. A taxpayer making a correctable error should receive help. A return consistent with known transactions should pass with little friction. A material inconsistency should receive targeted verification. A coordinated fraud network should receive concentrated investigation. The strategic capability is the ability to tell these situations apart and respond proportionately.
AI makes that differentiation possible at a scale that rules and manual judgment cannot sustain alone. Yet the goal cannot be to personalize suspicion. It must be to improve the fit between the problem and the treatment. An authority that detects more anomalies but responds to all of them in the same way has automated selection without modernizing compliance.
Compliance is moving closer to the economic event
Return-based administration concentrates attention at filing and afterward. The taxpayer aggregates activity, the authority receives the declaration and controls are applied to the result. Digital transaction evidence changes this sequence. An invalid identifier, inconsistent tax treatment or unexplained counterparty mismatch can become visible while the underlying event is still current and before it has propagated into a return, refund or assessment.
The significance is not merely that the authority learns sooner. It can act differently. A prompt delivered when an invoice is created may prevent an error at far lower cost than an inquiry months later. A prefilled position may remove ambiguity for a taxpayer trying to comply. A discrepancy that persists across multiple transactions may justify verification before it becomes embedded in a large refund claim. The point of control moves upstream, where assistance and prevention are still available.
The OECD has described this direction as getting tax right from the start and, more recently, as embedding taxation into the natural systems used by taxpayers. The Australian Taxation Office similarly describes a compliance model that ranges from making compliance easy to applying the full force of the law, with the response shaped by taxpayer behavior. SARS has made ease of compliance, data-driven risk management and appropriate responses recurring elements of its strategy. These ideas predate the current generation of AI. What AI changes is the authority’s ability to apply them with greater contextual precision.
That creates a different relationship with the taxpayer. The authority is no longer present only at the boundary of filing and enforcement. It can become part of the compliance environment through software, digital records and timely guidance. This proximity can reduce uncertainty, but it also increases institutional responsibility. If the authority influences a transaction earlier, its information must be reliable, its intervention must be proportionate and the taxpayer must be able to understand what is being asked.
The value lies in preventing avoidable friction
The economic prize is often measured through audit yield, collections and staff productivity. Those measures capture only the visible return to the authority. They do not capture the cost transferred to taxpayers through unnecessary reviews, delayed refunds, repeated evidence requests and uncertainty about how a rule will be applied. Nor do they measure the revenue protected when an error is prevented before it enters the system.
Adaptive compliance changes the economics by matching the cost of the treatment to the expected value of the risk. A low-value clerical inconsistency may warrant an automated prompt. A credible but incomplete claim may require one specific document. A pattern that crosses connected businesses and produces material revenue exposure may justify specialist investigation. Applying the most expensive treatment to every anomaly is not rigorous administration. It is a failure to discriminate.
The benefits are particularly visible in VAT and GST refunds. Legitimate claimants should not finance the authority’s uncertainty through avoidable delays. At the same time, false invoices and coordinated chains can turn refunds into a direct extraction of public revenue. Better evidence and network analysis allow the authority to release low-risk claims faster while concentrating expertise on claims whose transactions, counterparties and behavior do not form a credible economic story.
The economic case therefore depends on a balance. Revenue protected must be considered alongside taxpayer effort, refund speed, intervention cost and the rate at which automated concerns prove unfounded. An authority can improve one measure by damaging another. The stronger ambition is not maximum enforcement or minimum friction in isolation. It is the lowest combined cost of achieving sustainable compliance.
The treatment must follow the cause
The same visible discrepancy can have very different causes. A mismatch may reflect a misunderstanding of the rule, a timing difference, poor records, deliberate concealment or organized fraud. If the authority responds only to the symptom, it will apply the wrong treatment to many taxpayers. More aggressive enforcement will not correct a system design that repeatedly produces errors, just as more education will not change behavior built around deliberate abuse.
This is where AI can reshape compliance more profoundly than conventional risk scoring. It can combine transaction history, counterparty behavior, filing patterns, prior interactions and the materiality of the issue to form a view of what is likely happening and what response is most appropriate. The output should not be a universal score that compresses every concern into one number. It should be a recommendation tied to a decision, an explanation and a treatment.
For a taxpayer trying to comply, that treatment may be guidance, prefill or an opportunity to correct without escalation. For an uncertain case, it may be a narrowly framed request that resolves the specific inconsistency. For persistent behavior, the authority may increase monitoring or move to audit. Where connected evidence indicates deliberate and coordinated abuse, the response may need to shift from taxpayer-by-taxpayer examination to network investigation.
The decision system surrounding the model is therefore more important than the model alone. It includes the legal authority, evidence threshold, communication, accountable officer and route for review. It also includes the consequences of being wrong. A false positive that delays a small exporter’s refund can create immediate working-capital harm. A false negative in an organized fraud chain can create material revenue loss. Adaptive compliance requires those asymmetries to be explicit rather than hidden inside an accuracy measure.
Organize intelligence around the taxpayer journey
Tax administrations are commonly organized by function. Service teams answer questions, processing teams handle returns and refunds, compliance teams select cases, and investigators pursue serious abuse. AI initiatives tend to reproduce those boundaries. Each team builds its own data view and model, even though the taxpayer and the underlying transactions are the same.
An adaptive operating model connects intelligence across the journey. A problem identified during invoicing should inform prefill and should not reappear as an unexplained mismatch at filing. Evidence gathered to verify a refund should be available if the same pattern later appears across connected entities. An accepted explanation should reduce future friction rather than becoming a permanent marker of suspicion. Intelligence must follow the issue through its lifecycle, not remain trapped in the function that first observed it.
This requires a shared treatment architecture. Document intelligence can validate evidence. Entity resolution can connect businesses and intermediaries. Graph analytics can reveal coordinated behavior. Predictive models can estimate likelihood and materiality. Generative AI can summarize the issue and produce clearer communications. But these capabilities become useful only when they support a defined choice among assistance, acceptance, verification, monitoring and enforcement.
Human judgment remains essential where consequences are material or the evidence is contestable. The objective is not to insert an officer into every low-risk decision. It is to reserve human attention for cases in which discretion, context or coercive authority matters. Routine compliance can become simpler and more automated precisely because the operating model directs expertise toward exceptions that deserve it.
The feedback loop is equally important. Corrections, accepted explanations, confirmed noncompliance, appeals and overturned decisions reveal whether both the signal and the treatment were appropriate. Models should learn from those outcomes, but they must not treat historical enforcement as unquestioned truth. If past selection was biased or inconsistent, reproducing it more efficiently will make the system less legitimate, not more intelligent.
Differentiation must not become arbitrary treatment
Adaptive compliance creates a legitimacy challenge. Two taxpayers with apparently similar transactions may receive different treatments because the surrounding evidence differs. That may be rational, but it must also be explainable. The authority needs to distinguish lawful risk differentiation from inconsistent administration and to show that decisions are tied to relevant evidence rather than opaque profiling.
Data quality is the first constraint. Fragmented identities, missing cancellations, late credit notes and incomplete payment information can make legitimate activity appear suspicious. The more quickly the authority intervenes, the less time the commercial record has had to settle. Near-real-time control therefore requires an explicit understanding of timing, confidence and reversibility. Not every early signal should create an immediate adverse action.
Behavioral segmentation also needs discipline. A history of compliance can justify lower friction, but it should not create permanent immunity. Previous errors can justify attention, but they should not trap a taxpayer in a high-risk category after the underlying problem has been corrected. Models and treatment rules need routes for taxpayers to recover trust, for new evidence to change the assessment and for officers to record why an automated recommendation was accepted or overridden.
Smaller businesses require particular care. Large taxpayers may have sophisticated systems, professional advisers and established relationships with the authority. Small businesses are more likely to depend on commercial software and to experience a digital mandate as a direct operating cost. The most effective intervention may therefore sit in the accounting, invoicing or payment system they already use. Compliance by design can reduce burden, but only if standards are interoperable and the cost of participation is not shifted onto the least capable taxpayers.
The practical path is to begin with one decision journey rather than a general AI platform. Refund release is a strong candidate because the value, taxpayer consequence and fraud exposure are visible. The authority can define which evidence supports rapid release, which uncertainty can be resolved through a targeted request and which network patterns justify escalation. It can then measure revenue protected, cycle time, false positives, taxpayer effort and appeal outcomes together. That is a more credible foundation for scale than deploying a risk model and searching for processes willing to use it.
The next advantage will be proportionate intelligence
The future of VAT and GST administration will not be a choice between service and enforcement. Digital evidence and AI make it possible to improve both by changing the treatment according to the situation. Taxpayers trying to comply can receive earlier guidance and less friction. Uncertain cases can be resolved with narrower interventions. Deliberate and coordinated abuse can receive the concentration of expertise it warrants.
The strongest authorities will not be those that identify the greatest number of anomalies. They will be those that understand why a risk has emerged, select the least burdensome treatment capable of changing the outcome and learn when that treatment succeeds or fails. They will recognize that the authority’s response is itself part of the compliance environment and that unnecessary intervention can be as revealing of a weak operating model as missed fraud.
That is the transition from universal scrutiny to adaptive compliance. It is a shift from treating risk as a queue of cases to treating compliance as a sequence of decisions. AI supplies the capacity to differentiate at scale. Trust will depend on whether tax authorities use that capacity to become not only more effective, but more proportionate, transparent and fair.
References
- OECD, Compliance Risk Management, Managing and Improving Tax Compliance
- OECD, Right from the Start
- OECD, Tax Administration 3.0, The Digital Transformation of Tax Administration
- OECD, Tax Administration 2017
- OECD, Tax Administration 2023
- Australian Taxation Office, How we help and influence taxpayers
- Australian Taxation Office, How we assess risk
- SARS, Strategic Plan 2020 to 2025
- SARS, Annual Performance Plan 2024 to 2025
- SARS, Large Business and International
