Artificial Intelligence has stopped being a “future of work” conversation in South Africa. It is in your inbox, your finance team’s spreadsheets, your recruiters’ shortlists, and almost certainly in the browser tab of an employee right now pasting client data into a public chatbot to “save time”.
The problem is not the technology. The problem is that most South African SMEs have no written rules about how employees may use it, what data may go into it, and who is accountable when it produces nonsense. In a country governed by the Protection of Personal Information Act 4 of 2013 (POPIA), the Labour Relations Act 66 of 1995 (LRA), and the Employment Equity Act 55 of 1998 (EEA), that silence is an unaffordable liability.
This guide takes HR Directors, IT and Compliance Officers, and SME owners through drafting a workplace AI Use Policy that is legally defensible, POPIA-aligned, and genuinely workable. With template clauses you can adapt and a clear link back to your existing Information Officer obligations.
Why Every South African Employer Needs an AI Policy in 2026
Three pressures are converging in 2026, and any one of them justifies a written policy.
1. POPIA is already an AI law, whether you realise it or not. South Africa does not yet have a standalone AI Act. The Draft National Artificial Intelligence Policy was published on 10 April 2026 for public comment and then embarrassingly withdrawn just two weeks later, after it emerged that the document contained fictitious academic citations — almost certainly generated by AI itself. While the policy timeline has stretched into 2027 and beyond, this does not create a regulatory vacuum. POPIA already governs every workplace AI interaction that touches personal information. The Information Regulator confirmed 1,607 reported data breaches between April and September 2025 alone, a 60% jump on the previous year. Employee-introduced AI tools are an emerging contributor to that figure.
2. Intellectual property and confidential data are walking out the door. Samsung learned this in April 2023, when engineers pasted proprietary source code, internal meeting transcripts, and chip-testing sequences into ChatGPT within a single 20-day window. The data was unrecoverable, and Samsung banned generative AI tools company-wide. South African SMEs face the same exposure with no comparable resources to detect or recover from a leak. Once your pricing model, client list, or draft contract sits in a public Large Language Model (LLM), you have lost control of it.
3. Hallucination liability is now a documented business risk. In the 2025 Mavundla judgment and the more recent Northbound Processing case, South African courts referred legal practitioners to the Legal Practice Council after they cited AI-fabricated cases that simply did not exist. The same risk pattern affects HR: imagine submitting CCMA papers, a disciplinary outcome, or a Section 189 consultation pack containing AI-invented “facts”. Even outside the courtroom, an AI-generated employee reference letter containing fabricated qualifications could expose your company to defamation or negligent misrepresentation claims.
A documented AI Use Policy is your first line of defence against all three.
Step 1: Decide Your Stance — Permissive, Restricted, or Prohibited Use
There is no “one-size-fits-all” AI policy. Start by deciding which of three postures fits your risk appetite and operational reality:
- Permissive use allows employees to use approved AI tools across most tasks, with clearly defined data-handling rules. This works for businesses where AI productivity gains are essential and competitive – agencies, software firms, content businesses.
- Restricted use confines AI to specific approved tools and specific use cases (for example, internal brainstorming and grammar checks only — never client work). This is the most common posture for SMEs handling personal or financial information.
- Prohibited use bans all employee use of generative AI tools on company systems and devices. This is appropriate for businesses processing large volumes of Special Personal Information under POPIA, such as medical practices, legal firms, or financial advisers.
Most SMEs will land somewhere between “restricted” and a tiered model. One set of rules for marketing teams, a stricter set for HR and finance. Whichever posture you choose, write it down and make it the opening clause of your policy.
⭐ Key Takeaway: A vague “use AI responsibly” instruction is unenforceable. Procedural fairness under the LRA requires that an employee knew, or could reasonably have known, that the conduct was prohibited. Your policy must be explicit.
Step 2: Map Data Sensitivity to AI Tools
POPIA distinguishes between ordinary Personal Information and Special Personal Information (defined in Section 26). Special Personal Information includes religious beliefs, race, trade union membership, health, sex life, biometric data, and criminal behaviour. The bar for lawful processing is significantly higher, and consent requirements stricter.
Your policy must map data categories to permitted AI tool tiers. A practical framework:
Tier A — Public information (low risk): Marketing copy already on your website, public press releases, general industry research. May be processed in public LLMs (ChatGPT, Gemini, Claude) for drafting and editing.
Tier B — Internal confidential information (medium risk): Internal SOPs, draft policies, anonymised data, generic templates. Should only be processed in enterprise AI tools with contractual data protection (Microsoft Copilot for business, ChatGPT Enterprise, or equivalent with a signed data processing agreement).
Tier C — Personal Information (high risk): Employee CVs, performance reviews, payroll, client contact lists. Restricted to enterprise AI tools where the vendor has signed an operator agreement under Section 21 of POPIA, and where data is not used for model training.
Tier D — Special Personal Information and privileged data (prohibited): Medical certificates, disciplinary records, legal advice, BEE verification documents, biometric data. May not be entered into any AI tool, full stop.
The reason this matters: under POPIA, your organisation remains the Responsible Party even when data is processed by a foreign AI vendor. Cross-border transfers (Section 72) require either the data subject’s consent, adequate protection in the receiving country, or a binding agreement. Pasting a candidate’s ID number into a public US-hosted chatbot satisfies none of these conditions.
Step 3: Set Rules for Client and Personal Information
Your client and employee data are not yours to give away. Section 19 of POPIA imposes a security safeguard obligation on every Responsible Party, requiring “appropriate, reasonable technical and organisational measures” to prevent unauthorised access, loss, or damage.
Your policy must contain at least the following client-and-personal-data rules:
- No identifiable information may be entered into a public AI tool. This includes names, ID numbers, contact details, banking information, medical conditions, photographs, and contract numbers. If anonymisation is required to use the tool, the burden lies on the employee to anonymise before pasting.
- Client deliverables drafted with AI assistance must be reviewed before release. A clause in your service agreements may also be needed where clients have prohibited AI use in their work — increasingly common in legal, financial, and government contracts.
- Operator agreements must be in place with any enterprise AI vendor processing personal information on your behalf. This is a direct POPIA Section 21 requirement, not a “nice to have”.
- Conversation history must be disabled in any AI tool where employees handle internal information. Most enterprise versions allow this.
- Any suspected leak must be reported to the Information Officer within 24 hours. Under POPIA, security compromises affecting personal information must be reported to the Information Regulator and to affected data subjects “as soon as reasonably possible”.
Step 4: Build the Disclosure and Human-Review Obligation
This is where many AI policies fall short. Allowing AI use without requiring disclosure creates a hidden risk: nobody knows what has been generated, edited, or fabricated. When something goes wrong, you cannot trace it.
Build two non-negotiable obligations into your policy:
Disclosure. Any work product substantially generated or substantially edited by AI must be flagged internally. This could be as simple as a footer on a document (“Drafted with AI assistance; reviewed by [name]”) or a tag on a shared file. The purpose is traceability — if an AI-introduced error reaches a client, you can investigate.
Human review. No AI-generated output may be sent externally, relied upon for a decision, or filed in an official record without a named human reviewer who takes responsibility for accuracy. This is non-negotiable for any document going to the CCMA, the Department of Employment and Labour, SARS, a client, or a court. The Northbound Processing judgment was unambiguous: “Admonishment alone is unlikely to be a sufficient response” where false AI material reaches a court. Time pressure is not a defence.
⭐ Key Takeaway: Your AI policy is the document that converts a “rogue employee” defence into a defensible position. Without disclosure and review obligations on paper, the employer carries the full liability.
Step 5: Address AI in Recruitment and Performance Management
This is the most legally sensitive part of any AI policy, and it deserves its own section.
POPIA Section 71 prohibits decisions based solely on automated processing where they have legal or substantially affecting consequences for an individual — including their performance at work or their suitability for employment. There are narrow exceptions, but the practical effect for HR is clear: an AI tool may shortlist, but it may not decide.
The Employment Equity Act prohibits unfair discrimination on listed and analogous grounds. AI recruitment tools are notorious for absorbing historical bias from training data — they have been shown internationally to penalise CVs containing women’s universities, predominantly Black-sounding names, or career gaps consistent with maternity. A South African employer who relies on such a tool without auditing it carries the discrimination liability, not the vendor.
Your policy must therefore require:
- Human-in-the-loop hiring decisions. AI may rank or summarise, but every shortlist, rejection, and offer must be reviewed and signed off by a named person.
- Disclosure to candidates. Where AI is used in the recruitment process, candidates should be informed in your privacy notice, and given an opportunity to make representations. This is a direct Section 71(3) requirement.
- Bias auditing. At least annually, the outputs of any AI recruitment or performance tool should be reviewed against your Employment Equity reporting categories. If your shortlists systematically exclude designated groups, your tool is a discrimination engine.
- No AI-only performance management. A Performance Improvement Plan, written warning, or dismissal recommendation may not be generated by AI without substantive managerial input. Schedule 8 of the LRA requires that performance management be based on a fair, individualised assessment — an AI summary alone will not survive CCMA scrutiny.
Step 6: Disciplinary Consequences for Breach
A policy without consequences is a suggestion. Your AI policy must be incorporated into your existing Disciplinary Code, and the relationship between the two must be explicit.
Categorise potential breaches in line with your existing progressive discipline framework:
- Minor breaches (using a non-approved AI tool for low-risk tasks, failure to disclose AI assistance on internal documents): verbal or written warning.
- Serious breaches (entering client information into a public AI tool, failing to review AI output before client release): final written warning.
- Gross misconduct (entering Special Personal Information into a public AI tool, submitting AI-fabricated information in a formal proceeding, deliberate concealment of AI use to mislead the employer): potential summary dismissal after a fair hearing.
The Code of Good Practice: Dismissal in Schedule 8 of the LRA requires that employees know what conduct is prohibited and what the consequences are. Once the policy is written, you must train on it, get signed acknowledgement of receipt, and incorporate the rules into onboarding for new hires. Without this paper trail, a CCMA arbitrator will be unlikely to uphold a dismissal for AI misuse, regardless of the actual harm caused.
Template Clauses (Free Download)
To save you starting from a blank page, HRSpot has prepared a free downloadable AI Use Policy template containing adaptable clauses for each of the six steps above, including:
- Scope and definitions section
- Tiered data classification matrix
- Approved tools register
- Disclosure and human-review clauses
- Recruitment and performance management protections
- Disciplinary schedule cross-referencing your existing Code
- Employee acknowledgement form
The template is drafted in plain South African English and aligned to POPIA, the LRA, and the EEA. It is designed to be customised — a cookie-cutter policy is exactly what you do not want when the Information Regulator comes calling.
How This Links to Your Existing POPIA Information Officer Obligations
The piece that ties everything together is the role of the Information Officer. By default, this is the head of your organisation, the CEO, the sole proprietor, or the equivalent, and accountability cannot be delegated away, even where duties are assigned to a Deputy Information Officer.
Section 55 of POPIA, read with Regulation 4, places the following responsibilities on your Information Officer that map directly onto your AI policy:
- Encouraging organisational compliance with the lawful processing conditions
- Developing, implementing, and maintaining a compliance framework
- Conducting Personal Information Impact Assessments (PIIAs) to confirm adequate safeguards
- Developing internal measures and adequate systems to handle requests
- Ensuring internal awareness sessions for staff
Each of these is a hook for your AI policy. The compliance framework is incomplete without an AI policy. A PIIA must now assess AI-related risks. The internal awareness sessions must include AI training. If your Information Officer has not registered with the Information Regulator yet, that is your first priority — they cannot lawfully exercise their duties until they have done so.
If your business is too small to have a dedicated Information Officer, the head of the business holds the role by default. Outsourcing the operational duties to a consultant is allowed; outsourcing the accountability is not.
Final Word
The 2025–2026 wave of AI-related judgments, the Information Regulator’s increasing enforcement appetite, and the rapid arrival of enterprise AI tools in South African workplaces make an AI Use Policy a 2026 priority – not a 2027 nice-to-have. The cost of drafting and implementing one is a small fraction of the cost of a single POPIA enforcement notice, a CCMA award following a procedurally unfair AI-related dismissal, or the reputational damage of a leaked client database.
Start with stance, map your data, define your rules, build in disclosure and review, protect recruitment and performance management, and back it all with disciplinary consequences. Then train your people on it. Seek expert advice when in doubt.
Need a hand customising it to your environment? Get in touch with our team for a free consultation.
Disclaimer: This article provides general information and guidance on workplace AI policy drafting in South Africa. It is not intended as legal advice. Employers should consult a qualified labour law or data protection specialist for advice specific to their circumstances.
