Your automation programme and your labour relations function have never met
The first time they are in the same room, it will be at the CCMA, with a commissioner, a deadline and a case neither of them prepared.
Sign up for my Substack daily AI newsletter https://johanosteyn.substack.com
Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/
Deputy President Paul Mashatile told the eighth national congress of the Federation of Unions of South Africa in October 2026 that artificial intelligence presents organised labour with a historic challenge, with the official unemployment rate at 33.6%, approximately 8.5 million people without jobs, and a 0.2% economic contraction in the second quarter of 2026. He said retrenchments should not be treated as an inevitable consequence of corporate restructuring, and called for engagement between government, employers and unions before retrenchment notices are issued. Most coverage read this as a politician worrying about jobs. Something more consequential was happening. The Deputy President was describing where South Africa's first real constraint on artificial intelligence deployment will come from, and it is not an AI regulator. It is a retrenchment consultation, under a statute written in 1995.
CONTEXT AND BACKGROUND
No AI statute is on the way. The Department of Communications and Digital Technologies has told Parliament the national AI policy will not be standalone legislation, will be layered onto existing laws, will be overseen by several regulators rather than a dedicated one, and will not be finalised before the 2026-2027 financial year. The automated decision-making provision in the Protection of Personal Information Act is the closest thing to a binding South African rule aimed at artificial intelligence.
The Labour Relations Act, by contrast, defines operational requirements in section 213 as the economic, technological, structural or similar needs of an employer, so technology has been a lawful ground for dismissal here since 1995. Where an employer of more than 50 people contemplates dismissing at least 10, facilitation by the Commission for Conciliation, Mediation and Arbitration must be requested within 15 days of the section 189(3) notice, and a commissioner then has 60 days before termination notices may be issued.
INSIGHT AND ANALYSIS
Mashatile's proposal is more precise than it first appears. He asked for engagement before notices are issued, which is the one window the Act leaves largely empty. By the time that notice is drafted, the business case is approved, the software bought, the implementation begun and the headcount number already in a board pack. What follows is real in law and settled in fact.
The greater difficulty is evidentiary. An employer must disclose its reasons and consider alternatives. When the reason is a supplier's claim about what a system will do to a workflow, nobody in the room can test it. The union cannot audit the model, and the employer frequently cannot either, because the claim arrived on a vendor's slide. A process built to interrogate falling demand is being asked to assess a procurement decision.
I have previously written about this in Tasks, Not Jobs, the Most Convenient Reframe in the History of AI, where a shift in language let the industry acknowledge AI's effect on work while avoiding the people affected. A restructuring case built from task-level efficiency claims does the same work inside a consultation.
The strongest counter-argument is that the premise is overstated. A firm survey from the National Bureau of Economic Research finds roughly 78% of firms using AI, yet executives report essentially no net employment effect over the past three years. Job losses here are more plausibly explained by electricity, logistics and weak demand than by automation, and a rate of 33.6% was not built by machine learning. A sharper objection is that raising the friction around formal restructuring pushes work offshore, where no consultation is owed to anyone.
IMPLICATIONS
A study by the Friedrich-Ebert-Stiftung and UNI Europa reviewed 108 European collective agreements carrying clauses on AI and algorithmic management. South Africa already has sectoral bargaining councils, statutory consultation and a facilitation body. The comparison is imperfect, because European works councils carry co-determination rights with no local equivalent, yet the forums exist here and sit underused.
University of Johannesburg researchers found exclusion written into the Act's definition of an employee, leaving platform drivers only partly covered by transport regulation and remote online workers covered by nothing at all. The workers Mashatile wants represented sit outside the machinery he proposes to use.
For boards, three things follow. First, report net headcount movement, junior hiring included, beside any claimed productivity gain. Second, decide in advance what evidence the organisation will produce when a union asks what the system actually does, because improvising under a 60-day clock is the worst version of this conversation. Third, get the automation programme and the labour relations function into one room before a commissioner convenes one.
CLOSING TAKEAWAY
Governance does not wait for the instrument designed for it. It arrives through whatever instrument is nearest, and here the nearest one is a retrenchment consultation that has never been tested on an algorithm. No published South African bargaining council agreement appears to carry an algorithmic management clause, and no reported case names an AI system as the reason for a dismissal. Both are coming. The organisations that handle them well will treat a labour consultation as an AI governance forum while there is still time to prepare.
Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence's impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan's work at https://www.aiforbusiness.net
Comments