Skip to content
Risk X Group

InsightsCyber Risk

The Fiscal Scissors: How Agentic AI Could Hollow Out the Tax Base

Johan van Zyl on how agentic AI could cut tax revenue while raising demand for social support, why South Africa is exposed, and what policymakers and organisations can do.

By Johan van Zyl · Chief Operating Officer, Risk X GroupPublished 6 min read

Most conversations about agentic AI focus on productivity, cost savings and which jobs will change. Far fewer ask a harder question: who pays for the state when the taxpayer is a machine?

After nearly 30 years in cybersecurity, audit and compliance, I have learned that the risks that hurt most are rarely the ones on the agenda. They are structural and build quietly until they become expensive to fix. I believe the fiscal impact of agentic AI is one of them.

The fiscal scissors

Modern tax systems were built around one assumption: that work is done by people who earn salaries, pay income tax and spend their wages.

Agentic AI breaks that assumption. When an AI agent replaces an employee, two things happen at once:

  • Revenue falls. The salary disappears, and with it, income tax, payroll levies and part of the VAT that salary would have generated.
  • Spending rises. The displaced worker may now need unemployment benefits, social grants or retraining.

One blade cuts revenue while the other raises demand. I call this the fiscal scissors, and in countries with narrow tax bases and high unemployment, the blades close fastest.

Five risks, not one

As a risk practitioner, I break this into five connected domains.

  1. Revenue base erosion. Personal income tax, payroll levies and consumption taxes shrink as salaried roles disappear. Corporate tax rarely fills the gap, because profit is taxed at lower rates than salaries, and often elsewhere.
  2. Expenditure pressure. Demand for grants and unemployment benefits grows. Young people are hit hardest, as their entry-level roles are exactly the ones agents absorb most easily.
  3. Offshore value capture. When a local company replaces staff with an AI service from a foreign provider, the value created locally is booked and taxed abroad. The country loses the salary tax and gains a foreign-currency bill.
  4. Policy lag. Tax law moves in years, adoption moves in months. Official statistics count jobs, not agent output, so erosion may only be visible after it has happened. Equally, a rushed “robot tax” could drive investment away.
  5. Stability and sovereign risk. A persistent gap between revenue and need becomes social strain, rising debt, and pressure on credit ratings.

The first three feed the gap. The fourth decides how wide it gets. The fifth is where it lands.

Why South Africa should pay attention

This is a global issue, but South Africa starts from an unusually exposed position:

  • A narrow tax base. Personal income tax is the state’s largest revenue source, estimated at R844.8 billion in 2026/27, or about 40% of gross tax revenue. The top 13% of individual taxpayers pay over 60% of it.
  • A heavy social commitment. Around 26.5 million people receive social grants, compared with 8.3 million people who pay income tax.
  • Deep unemployment. Official unemployment reached 33.6% in the second quarter of 2026, and 47.4% for young people aged 15 to 34.
  • Exposed sectors. Financial services, business process outsourcing and contact centres are major formal employers, and among the most exposed to AI agents.
  • Imported AI. Most AI services used here are provided by foreign companies.

The tools meant to capture offshore digital value are also weaker than they look. Since April 2025, foreign suppliers selling digital services only to VAT-registered South African businesses fall outside the foreign e-services VAT regime. And under the OECD’s January 2026 side-by-side agreement, US-parented groups are largely exempt from key Pillar Two global minimum tax rules. An AI provider with no local entity remains largely beyond reach.

At the same time, Africa’s population is still growing fast. The UN projects that sub-Saharan Africa will nearly triple to 3.4 billion people by 2100. A young, growing workforce meeting a shrinking pool of entry-level jobs is not a comfortable combination.

Why the usual reassurances fall short

Three arguments are often used to dismiss this concern. Each weakens in a South African context.

  • “New jobs will replace the old ones.” That has been true of past technology waves, but it is a pattern, not a guarantee. The new roles and the value that AI creates are likely to sit where the providers are based, while we pay for the services.
  • “AI will augment workers, not replace them.” That may keep a job, but it does not increase taxable income unless wages rise. Productivity gains flow mainly to employers.
  • “Higher profits will offset lost income tax.” Profits are taxed at a 27% corporate rate, compared with a top personal rate of 45%, and some of those profits are booked offshore. More importantly, higher profits do nothing for unemployment or grant demand.

We do not know how fast agentic AI will be adopted. But uncertainty about timing is a reason to prepare, not to wait. Good risk management does not require proof that a risk has materialised before it is raised.

What to watch

Revenue figures will show the damage years after it starts. Leading indicators give earlier warning:

  • Entry-level job postings in finance, outsourcing and contact centres
  • Headcount in exposed sectors falling while their output holds or grows
  • The labour share of GDP
  • Personal income tax as a share of total revenue
  • UIF claims compared with contributions
  • Companies citing AI as a reason for hiring freezes

What can be done

There is no single fix, and every option has a trade-off.

For policymakers:

  • Rebalance the tax mix towards capital, consumption or wealth, accepting that VAT is regressive and capital is mobile.
  • Tax AI use directly, through an automation, compute or usage levy. This is measurable, but hard to define and could slow adoption.
  • Close the offshore gap through stronger digital nexus and withholding rules for cross-border AI services.
  • Invest in transition, through funded reskilling and youth pathways, recognising the long payback.
  • Measure what matters, by building statistics that track AI output, not just jobs.

For organisations deploying AI:

  • Assess societal impact before deploying agents. ISO/IEC 42001 already requires an AI system impact assessment covering consequences for individuals, groups and society (clauses 6.1.4 and 8.4).
  • Plan workforce transitions, redeploying and reskilling where feasible rather than simply releasing staff.
  • Disclose AI-related headcount effects in ESG and integrated reporting.
  • Put macro AI displacement on the board’s emerging risk register.

The conversation we need to have

Agentic AI will bring real benefits. But a tax system designed for human labour will not adapt on its own, and the countries least able to absorb a fiscal shock may feel it first.

The question is not whether we adopt AI. It is whether our fiscal and governance frameworks evolve quickly enough to share its gains, rather than concentrate them.

I would welcome views from policymakers, economists, tax practitioners and fellow risk professionals. What indicators are you watching, and what responses do you think could work?

Sources

Johan van Zyl is Chief Operating Officer of Risk X Group, Managing Director of Risk X Data Assurance, and South Africa Chapter Leader of the VigiTrust Global Board of Advisors. Views are his own. This article was first published on LinkedIn.

Keep exploring

Related Resources