- AI-assisted workplace decisions are already influencing disciplinary hearings, retrenchments, promotions and contract outcomes.
- Organisations should keep a contemporaneous record showing the task, sources, AI output, verification process, challenges raised and final human decision.
- A clear AI evidence protocol protects both affected employees and decision-makers by making reasoning traceable and accountable.
A recent Labour Court matter reported by Conviction should be read as a warning far beyond the legal profession. The court paused a municipal disciplinary hearing after finding initial evidence that the chairperson may have relied on nonexistent or misapplied legal authorities. Whether the errors came from an AI tool or from some other failure, the practical problem was the same: the reasoning chain could not be trusted.
South African employers and public bodies are already using generative AI to summarise policies, compare documents, draft allegations, search precedents and prepare recommendations. Those uses can save time. But when the output shapes a disciplinary process, a grievance, a retrenchment, a promotion or a contract decision, speed is not enough. The organisation must be able to show how it moved from source material to outcome.
That requires an AI evidence protocol: a short record created while the decision is being prepared, not a reconstruction assembled after lawyers become involved.
The six questions every organisation should answer
The protocol should answer six questions.
First, what decision was the system helping to prepare? “Used AI” is too vague. A record should identify the task: summarising witness statements, comparing an employee’s conduct with a policy, checking a contractual clause or drafting reasons for a recommendation. Defining the task makes it possible to judge whether AI was suitable for it.
Second, what source material did the system receive? The record should identify the policy version, contract, correspondence, witness material, legislation or cases used. If confidential personal information was entered, the organisation should also record who authorised that use and where the system stored or processed the data.
Third, what did the system produce? Keeping the relevant output matters because a final polished document can conceal a faulty intermediate step. A summary may omit an exculpatory sentence. A legal research response may cite a case that does not exist. A comparison may treat two different policy versions as identical.
Fourth, who verified the output, and how? A named person should reopen every cited authority, compare every material factual claim with the original record and note any correction. Verification cannot mean asking the same system whether its first answer was accurate. It requires independent checking against authoritative sources.
Fifth, what did the affected person challenge? A fair process needs a route for an employee, contractor or other affected party to identify a mistaken fact, missing document or misleading inference. The challenge and the response should become part of the same record. That is especially important when a system’s apparent confidence may discourage people from questioning it.
Sixth, who made the final decision? The human decision-maker should explain what was accepted, rejected or changed after review. This is not a ceremonial signature. It is evidence that someone exercised judgment and remained accountable for the result.
The point is not to turn every routine use of AI into litigation paperwork. A protocol can fit on one page or in a structured digital form. Its depth should match the stakes. A low-risk draft may need only a source check. A disciplinary finding, retrenchment recommendation or decision affecting someone’s livelihood deserves a fuller record.
This approach also protects managers and professionals. Without a contemporaneous record, a careful decision-maker may later struggle to prove that the machine’s output was checked. With one, the organisation can show the sources considered, the corrections made and the person who owned the conclusion.
Conviction.co.za recently reported Deputy Chief Justice Dunstan Mlambo’s view that technology should strengthen rather than replace human judgment. That principle becomes operational only when institutions design for traceability. Human judgment is not preserved merely because a person clicks “approve” at the end. It is preserved when the person can explain the evidence, identify the system’s limits and defend the reasoning.
Why the timing matters
The timing matters. Legal Indaba 2026 is set to focus on AI, ethics and accountability in South Africa’s legal profession. The same conversation belongs in HR departments, municipal offices, corporate legal teams and bargaining rooms. By the time an AI-assisted decision reaches court, the cost of a weak process has already been paid in delay, distrust and damaged relationships.
South African workplaces do not need to choose between useful technology and procedural fairness. They need a record that connects the two. An AI evidence protocol would make better decisions easier to defend and bad decisions easier to correct before a dispute hardens into litigation.
Get your news on the go. Click here to follow the Conviction WhatsApp channel.
The opinions expressed in this article are solely those of the author and do not necessarily reflect the views of Conviction.co.za
