AI for SA business
AI and POPIA: is your data safe with AI tools?
AI and POPIA are compatible, but only if you treat an AI tool like any other place you send personal information. The worry is reasonable: you want the speed of AI without handing customers’ details to a system you do not control. POPIA gives you a clear set of rules for that. Follow them and AI is safe to use for most business work.
Is your data safe with AI tools?
It depends far more on the plan you use than on the brand. Consumer AI apps are built for individuals, and several use conversations to improve their models by default. Business plans are built for organisations bound by laws like POPIA, so they switch training off and put a contract behind it.
| Tool | Personal or free plans | Business plans |
|---|---|---|
| ChatGPT | OpenAI says content is used to train its models, with an opt-out | No training on Business, Enterprise or API data by default; DPA available |
| Google Gemini | Chats train Google’s AI and some are read by human reviewers while Keep Activity is on | Workspace prompts are customer data under Google’s Cloud Data Processing Addendum and are not used for training without permission |
| Microsoft Copilot | Consumer conversations can train Microsoft’s models unless you opt out | Work accounts are covered by Microsoft’s Data Protection Addendum, and prompts are not used to train foundation models |
| Claude | You choose whether chats are used for training | Anthropic says it does not train on Claude for Work or API data by default |
These terms come from each vendor’s own pages in September 2026. They change, so check the plan you actually buy.
AI and POPIA: what the law requires
POPIA (the Protection of Personal Information Act 4 of 2013) does not mention AI, and it does not need to. Its conditions apply to any processing, and section 1 defines processing broadly, from collecting and storing information to using and transmitting it.
| POPIA rule | What it means | What to do with AI |
|---|---|---|
| Lawful basis and purpose | Process personal information only for a lawful, clear purpose | Only give AI data you are allowed to use for that job |
| Minimality | Use only what you need | Strip names, ID numbers and account numbers the AI does not need |
| Security (section 19) | Take reasonable technical and organisational measures | Control who may use which tool, and vet the vendor |
| Operators (sections 20 and 21) | Anyone processing for you must act on your authority, under a written contract | Sign the vendor’s data processing agreement before real data goes in |
| Cross-border transfers (section 72) | Personal information may leave South Africa only on a listed basis | Check where the vendor processes data and what contract protects it |
| Breach notification (section 22) | Tell the Regulator and affected people about a compromise | Report through the Regulator’s eServices Portal |
Where AI tools put you at risk
Personal accounts used for work. A staff member pastes a client email into their own chatbot to “just tidy it up”. That data now sits with a vendor under consumer terms, possibly training its model. Make a written rule against it.
Special personal information. Section 26 gives extra protection to information about religious or philosophical beliefs, race or ethnic origin, trade union membership, political persuasion, health or sex life, biometric information and criminal behaviour. Health records and similar data should not go near a consumer AI tool at all.
Data leaving the country. Most AI services process data outside South Africa. Section 72 allows that only if the recipient is bound by law, binding corporate rules or an agreement giving adequate protection, or if the person consents, or it is necessary for a contract with or for them. A vendor’s data processing agreement is usually how a business meets the first option.
Decisions with no human. Section 71 limits decisions with legal or substantial effects that are based solely on automated profiling. If AI screens job applicants or scores credit, keep a person in the decision.
The practical rules that keep you safe
- Choose a business plan for anything touching customer data, and sign the data processing agreement.
- Turn off training where a plan allows it, and use temporary or incognito chats for one-off questions.
- Anonymise where you can. See how to anonymise data before using AI.
- Decide who may use which tool for what, and write it down in a one-page policy.
- Keep a simple record of which AI tools process personal information, why, and under which contract.
- Plan for breaches. Know how you would notify the Regulator and affected customers.
For the free-versus-paid detail on specific tools, see ChatGPT free vs paid in South Africa, Microsoft Copilot in South Africa and DeepSeek in South Africa, whose privacy policy says it stores personal data in China.
Who carries the responsibility
Under POPIA you are the responsible party: the one who decides why and how personal information is processed. That accountability does not pass to the AI vendor. The vendor is your operator, and the written contract under section 21 is what binds it to keep the data secure. If you choose a tool with weak terms and it leaks customer data, the exposure is yours.
The Information Regulator can issue enforcement notices, and failing to comply with one is an offence that can carry a fine or up to 10 years’ imprisonment. The Regulator can also impose administrative fines of up to R10 million. As at September 2026 the Regulator’s POPIA resources page lists guidance notes, including one on direct marketing from December 2024, but no guidance note specific to AI. For the wider legal picture, see AI regulation in South Africa.
If your business wants to use AI on customer data and needs help choosing a safe setup, Zaiq’s AI consulting covers that kind of assessment. This guide is general information, not legal advice; for a high-risk decision, confirm with a POPIA practitioner.
Questions people ask
Does POPIA apply to AI tools?
Yes. POPIA defines processing as any operation on personal information, including collecting, storing, using or transmitting it, so typing a customer's details into an AI tool is processing. The Act applies to AI exactly as it applies to a spreadsheet or a CRM. Using AI does not create an exemption, and you remain the responsible party for what happens to the data.
Is it safe to put customer data into ChatGPT?
Not into a personal ChatGPT account, as a rule. OpenAI's pricing page says content on consumer plans is used to train its models unless you opt out. For customer data, use ChatGPT Business, Enterprise or the API, which OpenAI says it does not train on by default and for which it signs a data processing agreement, or remove the personal details before you ask.
What does POPIA require of a small business using AI?
The same conditions as any other processing: a lawful reason such as consent or a contract, a clear purpose you stick to, only the data you need, reasonable security safeguards under section 19, a written contract with any operator that processes data for you, and openness with the people concerned. POPIA also gives people rights to access and correct their information.
Can I use AI for marketing under POPIA?
Yes, within the rules for direct marketing. Section 69 says electronic direct marketing, such as SMS, email or WhatsApp messages, needs the person's consent, unless they are an existing customer and you are marketing your own similar products with a chance to opt out every time. AI that writes or sends marketing must only use contacts you are allowed to market to.
Does POPIA allow automated decisions made by AI?
With limits. Section 71 says a person may not be subject to a decision with legal consequences, or one that affects them substantially, based solely on automated profiling, such as a credit decision. There are exceptions, for example in a contract, but then you must let the person make representations and explain the logic behind the decision. Keeping a person in the loop is the simplest route.
Who is responsible if an AI tool leaks our customer data?
You are, as the responsible party. The AI vendor is your operator, and section 21 requires a written contract that makes it keep the security measures section 19 requires. If a breach happens, section 22 requires you to notify the Information Regulator and affected people as soon as reasonably possible; since 1 April 2025 the Regulator wants notifications through its eServices Portal.
Sources
- Protection of Personal Information Act 4 of 2013 (gov.za)
- Information Regulator, POPIA resources, guidance notes and security compromise reporting
- ChatGPT pricing, including the consumer training and opt-out note
- OpenAI, enterprise privacy commitments and DPA
- Microsoft Learn, enterprise data protection in Microsoft Copilot
- Google Workspace, generative AI privacy hub
- Gemini Apps Privacy Hub
- Anthropic, is my data used for model training (commercial products)
Checked September 2026. Prices and features change, so confirm on the official site before you buy.