An AI policy for employees in South Africa is a short set of written rules that tells staff which AI tools they may use and on which accounts. It also says what information may never go into those tools, who checks the output and who staff tell when something goes wrong.
No South African law asks for one by name, and the country has no AI Act. POPIA and labour law still make the policy the document you'll rely on, both to show you took reasonable care of personal information and to discipline someone fairly if they ignore the rules.
A policy works only when staff have an approved tool to use in place of the ones they use now, so the rollout matters as much as the wording.
Let's say your sales team has used ChatGPT on personal accounts for a year to write proposals, and your finance team uses it to tidy up debtor emails. One morning a PDF lands in everyone's inbox saying AI is now restricted, and nothing else changes. Within a week the proposals are being written on personal phones instead of work laptops, and the company knows less about its data than it did before the policy existed.
I'm not a lawyer and this article isn't legal advice. For your own policy, speak to your information officer, an employment lawyer or a POPIA attorney.
Does South Africa have an AI policy or AI law?
South Africa has no AI law, and it no longer has a draft national AI policy either. The Department of Communications and Digital Technologies put out a framework for comment in 2024, and then gazetted a draft National AI Policy on 10 April 2026.
The Minister withdrew it 16 days later, on 26 April, after fictitious sources were found in its reference list. He said the most plausible explanation was AI-generated citations included without proper verification. Cabinet later approved the withdrawal so the policy could be reworked in full.
This leaves companies setting their own rules. The Information Regulator hasn't published a guidance note on AI, and its chair, Pansy Tlakula, said in September that without a national AI policy "it's a free for all". Much of what shows up in search for "South Africa AI policy" was written before the withdrawal and still describes the draft as on its way.
Is a company legally required to have an AI policy?
No law names an AI policy as a requirement, but two existing duties come close to demanding one. The first comes from POPIA, where section 19 requires the company to secure personal information with "appropriate, reasonable technical and organisational measures". A written rule about AI tools is one of the simplest organisational measures a company can take.
Section 20 of the same Act says staff may process personal information "only with the knowledge or authorisation of the responsible party", which is the employer. A policy is how the employer gives, and limits, that authorisation.
The second duty comes from labour law. Under the Code of Good Practice: Dismissal, in force since 4 September 2025, an employer disciplining someone for misconduct has to show that a rule existed and that the employee knew it or could reasonably have been expected to know it. Without a written AI rule that staff were told about, that test is hard to pass.
My earlier article on whether ChatGPT is safe for your company's data goes through those POPIA sections, and the cross-border rules in section 72, in more detail.
What should a company AI policy include?
A company AI policy should cover ten parts, and most of them fit on a page each. A short policy is more likely to be read, so write each part as a rule a staff member can follow without asking anyone.
- Scope. Who it covers (staff, contractors, temporary workers) and which tools count as AI, including features built into software the company pays for, such as an email client that drafts replies.
- Approved tools and accounts. A named list of tools, each on a company account. Personal accounts are out for company work.
- Data classes. What may go into an approved tool, what may go in only with care, and what never goes in (the table in the next section).
- Human review. Who checks AI output before it reaches a client, a regulator, the board or the public.
- Decisions about people. AI may help with decisions about staff, applicants or customers, but a person makes the decision (more on section 71 below).
- Ownership. Who owns what staff create with AI tools, and what may never be uploaded because the company doesn't own it.
- Disclosure. When staff must say that AI helped produce something, and who decides that for client work.
- Incidents. Who staff tell, and how fast, when something goes into the wrong tool.
- Training and acknowledgement. How staff are trained, and a signed record that each person has read the policy.
- Review. Who owns the policy and when it's next reviewed, because the tools change faster than most company documents.
The government's own draft shows why the fourth part belongs in the list. The draft National AI Policy was withdrawn over citations that weren't properly checked, and a human review rule is there to stop the same thing happening in your board pack or your client proposal.
Which data should staff never put into an AI tool?
A data rule is easier to follow when it fits in a table staff can read in a minute, and three classes are usually enough to start with.
| Class | Examples | Rule |
|---|---|---|
| Never | ID numbers, bank and card details, passwords and API keys, payroll, special personal information under POPIA section 26 (for example health, sex life, religion or philosophical beliefs, race, trade union membership, political views, biometrics and criminal behaviour) | Never goes into any AI tool unless a specific, approved system has been built for it |
| Company tools only | Client names and contact details, contracts, internal financials, unreleased plans, source code | Approved tools on company accounts only, never a personal account |
| Open | Public information, generic drafting, your own published content | Any approved tool |
Section 26 is the reason the first row is so strict. POPIA generally prohibits processing special personal information, and a sick note or a medical report is the kind of document someone might paste into a chatbot to "make it sound more professional".
Can AI make decisions about your staff or customers?
AI can help, but under POPIA it shouldn't make the final call on its own. Section 71 says a person may not be subject to a decision with legal consequences, or one that affects them to a substantial degree, if it's based solely on automated processing that profiles them, including their "performance at work" or their credit worthiness.
For your policy, that means a person has to make the decision when AI screens CVs, scores performance or assesses a customer's credit. There are exceptions in section 71(2), and where they apply the person still has to be able to make representations and get enough information about the logic behind the decision. The simplest rule for staff is that AI can prepare the work, and a named person decides and signs.
Who owns work that staff create with AI?
The policy should say that work created with company AI tools belongs to the company, but the law behind that is less settled than the rule suggests. The Copyright Act treats the author of a computer-generated work as the person who made the necessary arrangements for it to be created, and according to Werksmans, South African courts have yet to deal directly with AI-generated works.
That uncertainty is a reason to keep the ownership part short and practical. It should cover two things: what the company claims, and what staff may never upload, such as a client's confidential documents or material under someone else's licence. Anything more detailed is a question for your lawyer.
Why does an AI ban not work on its own?
A ban without an approved alternative moves AI use onto personal phones, where nobody can see it. Staff use AI tools because they save real time, and a ban asks them to choose between the policy and their workload.
Unapproved AI use is common in South Africa. World Wide Worx's SA AI Index 2026 surveyed 100 decision-makers at South African organisations with more than 200 staff and found that 75% report some shadow AI (tools staff use without the company's approval).
The same survey found that organisations running approved and unofficial tools side by side rose from 20% to 54% in a year. Approving one tool doesn't end the use of the others, which is why the policy and the tool have to arrive together.
For many companies the approved tool is a company plan from a major vendor. On those plans the employer gets visibility too: on ChatGPT Business, according to OpenAI, admins "can view, access, export, and delete end user conversations", so the policy should tell staff that.
Other companies want an assistant that works on their own documents and keeps the data in the country. CBOS builds private AI systems for this, with the data stored and processed in South Africa and no outside AI model training on it. Each one has enterprise guardrails and permissions set per user, so staff get answers only from documents they're allowed to open.
CBOS's page on a private AI assistant built on your own data explains how permissions, traceability and human sign-off are set for each use case.
How do you roll out an AI policy so staff follow it?
A policy that staff follow is rolled out over a few weeks, with the approved tool in place before the rules take effect. These seven steps work in roughly this order:
- Find out what staff use now. Ask openly, and make it clear that answers won't be held against anyone, because you'll only hear the truth if it's safe to tell it.
- Choose the approved tools. Pick tools that cover the work staff were already doing, and move people onto company accounts.
- Write the policy short. Keep it to the ten parts in plain language, with the data table on its own page.
- Train with your own examples. Show staff the debtor email, the proposal and the HR letter, and which class each falls into.
- Get a signed acknowledgement. This is your record that each person knew the rule, which the Code of Good Practice asks you to show.
- Apply it the same way to everyone. Consistency is part of the fairness test, so a rule that's enforced for junior staff and ignored for managers won't hold up.
- Review it on a fixed date. Every six months is a sensible starting point while the tools and the law keep changing.
What you learn in the first step shapes the rest of the rollout. A company that learns its finance team relies on AI for month-end emails will choose a different tool, and write a different data table, from one that assumed nobody was using it.
Can staff be disciplined for breaking an AI policy?
They can, if the policy meets the same test as any other workplace rule. Under the Code of Good Practice: Dismissal, the employer has to show that the rule existed and that the employee knew it or could reasonably have been expected to know it. It also has to show that the rule was applied consistently and that the sanction was fair.
A written policy, a training session and a signed acknowledgement cover the first two. The third depends on how you enforce it, and the fourth depends on what happened: a first, honest mistake reported straight away is a very different case from someone uploading a client database to a personal account after being told not to. The same test protects staff, because without a clear rule, discipline for using AI is much harder to justify.
Can you use a foreign AI policy template?
You can start from one, but most templates on page one of search are written for the United States or Europe and refer to laws that don't apply here. Before using one, check these points:
- Replace GDPR or HIPAA references with POPIA, and name your information officer as the person staff report to.
- Use POPIA's section 26 list for special personal information, which differs from the GDPR's.
- Add section 71 for automated decisions about people.
- Point the disciplinary part at your own disciplinary code and the Code of Good Practice: Dismissal.
- Remove references to an AI law that South Africa doesn't have.
A template also can't tell you which tools your staff are using today, which is why the rollout starts with asking them. Once you know, the policy and the approved tool can be set up together, and a discovery call with CBOS is a good place to work out what that tool should be.
Frequently asked questions
Is it okay to use ChatGPT at work in South Africa?
It is if your employer allows it and you use it the way the company's policy says. If the work involves personal information about clients or colleagues, using a personal account without permission puts you outside the authority section 20 of POPIA gives staff.
Can a company ban ChatGPT?
An employer can usually set which tools staff use for work, as long as the instruction is lawful and reasonable. A ban works better when staff are given an approved tool for the same job.
Do we have to tell clients we use AI?
That depends on your contracts with each client and, in regulated professions, your professional body's rules. If client personal information goes into the tool, POPIA applies as well. The policy should name who decides this for client work, so staff don't have to guess.
Can an employee refuse to use AI at work?
No South African rule deals with AI specifically, so the general rule applies: an employer can decide how work is done if the instruction is lawful and reasonable. An employer who requires AI should train staff to use it first, and an individual case is a question for HR or an employment lawyer.
Who should own the AI policy?
One named person, usually with input from IT, HR and the information officer. Ownership matters most at review time, because a policy nobody owns stops being updated.
How often should an AI policy be reviewed?
Put a fixed review date in the policy itself, then bring the review forward after any incident, when you add a new tool or if South Africa publishes a new national AI policy.
Does the policy cover AI built into Microsoft 365 or Google Workspace?
It should. AI features built into software the company pays for count as AI tools, so list them under approved tools with the accounts staff may use them on.