POPIA and AI: a checklist before you switch anything on

AI tools are arriving in transformation and HR work quickly. Some are built in-house, some are added to software you already use and some are opened in a browser by an employee trying to save an afternoon. Each may process personal information about your employees, learners, suppliers and beneficiaries, and much of that information is sensitive.

The Protection of Personal Information Act (POPIA) has no separate chapter on AI. It does not need one: its conditions apply to any processing of personal information, whatever the technology. What follows is the checklist we work through before any AI system touches transformation data. It is not legal advice, and some questions will need your information officer or legal adviser. But if you cannot answer them, you are not ready to switch anything on.

1. What personal information will it touch?

List what the system will read, store or produce. Be specific: names, ID numbers, payroll, EE declarations, disability records, bursary applications, details of supplier directors. Include what the AI itself produces, such as summaries, classifications or flags about a person. Those outputs are personal information too.

2. Is any of it special personal information?

Section 26 lists special personal information: religious or philosophical beliefs, race or ethnic origin, trade union membership, political persuasion, health or sex life, biometric information and criminal behaviour. Disability information will usually be treated as health information. Transformation data almost always includes race, and often disability.

Processing special personal information is prohibited unless an authorisation in sections 27 to 33 applies. For race, section 29 allows processing to comply with laws and other measures designed to protect or advance people disadvantaged by unfair discrimination, which is the basis for much B-BBEE and EE processing. That authorisation is tied to its purpose. Ask whether each new AI use stays within it.

3. What is the lawful basis?

For ordinary personal information, section 11 sets out the grounds for processing, among them consent, the performance of a contract, a legal obligation and the legitimate interests of your company or a third party. Legitimate interest is not a blanket permission. You should be able to explain why the processing is necessary, and people can object to processing on that ground. Write the basis down for each use.

4. Is it limited to what the purpose needs?

Section 10 requires information to be adequate, relevant and not excessive. Section 13 requires a specific purpose, and section 15 restricts using information for a new purpose that is not compatible with the original one. Feeding the whole HR database into a tool to answer a question about skills spend fails this test. Give each AI task only the fields it needs, and use totals or masked identities wherever individual detail is not required.

5. Who is the operator, and is there a contract?

If a vendor processes personal information on your behalf, it is an operator. Sections 20 and 21 require it to process only with your knowledge or authorisation, to treat the information as confidential and to be bound by a written contract to maintain security measures. It must tell you immediately if it believes the information has been compromised.

For AI services, read the contract for two things in particular: whether your documents and prompts are used to train or improve the vendor's models, and how long they are kept. If the answer is unclear, assume the worst until it is settled in writing.

6. Will the data leave South Africa?

Many AI services process data on servers abroad. Section 72 permits transfers of personal information to a third party in a foreign country only on specific grounds, such as the recipient being bound by a law, binding corporate rules or a binding agreement that provides an adequate level of protection, or the person's consent. Know where processing happens. Where special personal information would go to a country without adequate protection, also check whether section 57 requires prior authorisation from the Information Regulator.

7. Are the security safeguards in place?

Section 19 requires appropriate, reasonable technical and organisational measures, based on an assessment of foreseeable risks, with regular checks that they work. For an AI system that means access by role, encryption, a log of who used the system and what it read, and a tested plan for a breach. If a compromise does happen, section 22 requires you to notify the Information Regulator and the people affected as soon as reasonably possible.

8. Can people exercise their rights?

People may ask whether you hold information about them and for a record or description of it (section 23), and may ask for it to be corrected or deleted (section 24). Your AI system must make that possible. If a summary or flag about an employee is stored, you must be able to find it, show it and correct it.

9. Does it decide anything about a person?

Section 71 restricts decisions with legal consequences, or substantial effects, that rest solely on automated processing intended to profile a person, including their performance at work. The exceptions are narrow, and where they rely on protective measures, those must include a chance to make representations and enough information about the underlying logic to do so. In transformation work, keep a person responsible for every decision about an individual.

10. Have you told people, and recorded it?

Section 18 requires you to tell people what you collect and why. If AI is a new way of processing their information, update your privacy notices and internal policies. Then record the assessment itself. The regulations made under POPIA require the information officer to ensure that a personal information impact assessment is done, and a new AI system is exactly the kind of change it exists for.

One rule for every desk

Add one plain rule for staff: no personal information goes into an AI tool the company has not approved. Most leaks of this kind are not malicious. They are someone pasting a spreadsheet into a chat window to finish a task before lunch.

We build transformation and AI systems with these questions answered in the design: data kept on the client's own systems, access by role, a full audit trail and a person confirming every figure. If you would like to work through this checklist for a system you are considering, write to us at sales@bsvision.co.za.