POPIA Compliance When Your Business Uses AI: A Practical Checklist

POPIA compliance for AI means applying the Protection of Personal Information Act's eight conditions for lawful processing, and its rules on direct marketing, automated decisions and cross-border transfers, to every AI tool that touches customer data. POPIA has no AI-specific chapter, so the existing sections apply in full.

On 11 October 2026, the Information Regulator's POPIA page listed guidance notes on direct marketing, information officers, children and special personal information, but none on AI. The Regulator is watching, though. In September 2026 its POPIA executive, Advocate Tshepo Boikanyo, told ITWeb that AI models often keep personal information beyond its intended purpose: "With AI models, we don't know what then happens with this personal information." This checklist explains what the Act says. It is not legal advice, so get your own legal advice for your specific case.

In short

• Putting customer data into an AI tool is "processing" under POPIA, so all eight conditions apply.
• Your AI vendor is usually an operator. Section 21 requires a written contract that binds it to security measures.
• Most AI providers process data outside South Africa, which brings in the cross-border rules in section 72.
• AI-sent marketing by WhatsApp, SMS, email or phone needs prior consent from non-customers (section 69). Since April 2025, opt-out does not count as consent.
• Decisions with legal or substantial effect, based solely on automated profiling, are restricted by section 71.

Does POPIA apply to AI tools?

Yes. The Act defines personal information widely in section 1. It includes phone numbers, email addresses, location information, online identifiers, personal opinions and private correspondence. "Processing" covers collection, storage, "consultation or use" and transmission. Pasting a customer's email into a chatbot, or letting an AI agent read a WhatsApp thread, is processing.

Two roles matter. The responsible party decides the purpose and means of processing: that is your business. The operator processes personal information for you under a contract, without coming under your direct authority: that is usually your AI vendor. Chapter 3 then sets eight conditions: accountability, processing limitation, purpose specification, further processing limitation, information quality, openness, security safeguards and data subject participation.

One exclusion helps. Section 6(1)(b) says the Act does not apply to information that has been de-identified to the extent that it cannot be re-identified again.

The POPIA compliance checklist for businesses using AI

Work through these 12 checks for each AI tool or agent that handles personal information.

1. Register your Information Officer. Section 55(2) says information officers may only take up their duties once registered with the Regulator, which is done on its eServices portal. Regulation 4, as amended in 2025, adds that the officer must make sure a personal information impact assessment is done and that internal awareness sessions are held.
2. Map each AI tool and minimise what goes in. List the tools, the data each one receives and why. Section 10 says personal information must be "adequate, relevant and not excessive" for the purpose. Strip ID numbers, account numbers and anything the task does not need.
3. Confirm a lawful basis. Section 11(1) allows processing with consent, to perform a contract, to meet a legal obligation, or for a legitimate interest, among other grounds. If you rely on consent, you carry the burden of proof (section 11(2)(a)) and people can withdraw it at any time.
4. Check the purpose. Section 13 requires a specific, explicitly defined purpose. Section 15 says further processing must be compatible with it. Using old support chats to train or tune an AI model is a new use you must assess against the section 15 factors.
5. Tell people. Section 18 requires reasonably practicable steps to make sure people know what you collect, why, and whether it will go to another country and with what level of protection. Update your privacy notice to name your AI tools and transfers.
6. Sign an operator agreement. Section 20 limits operators to processing with your authorisation, in confidence. Section 21 requires a written contract and immediate notice of any breach. Check data-use terms too: Anthropic says it does not, by default, train on inputs or outputs from its commercial products, and OpenAI says data sent to its API is not used for training unless you opt in, with abuse-monitoring logs kept for up to 30 days.
7. Settle cross-border transfers. Section 72 allows a transfer abroad only if, for example, the recipient is bound by a law, binding corporate rules or agreement giving "substantially similar" protection, the person consents, or the transfer is needed for a contract with them. On 11 October 2026, OpenAI's API data residency list did not include South Africa.
8. Secure it. Section 19 requires reasonable technical and organisational measures: identify foreseeable risks, put safeguards in place, check them regularly and update them. For AI, that includes staff using personal chatbot accounts for work data.
9. Keep special personal information out by default. Section 26 prohibits processing information about religious or philosophical beliefs, race or ethnic origin, trade union membership, political persuasion, health, sex life, biometrics and alleged criminal behaviour unless a section 27 authorisation such as consent applies. Children's information has its own prohibition and exceptions in sections 34 and 35.
10. Keep a human in decisions that matter. Section 71 restricts decisions with legal consequences or a substantial effect that are based solely on automated profiling of things like creditworthiness, work performance or conduct. The Regulator's direct marketing guidance says all three elements must be present for the section to bite.
11. Get marketing consent right. See the next section. This is where AI sales agents most often go wrong.
12. Plan retention, rights and breaches. Section 14 says records may not be kept longer than necessary. Sections 23 and 24 give people rights of access and correction, so you must be able to find a person's data inside AI transcripts. Section 22 requires notifying the Regulator and affected people as soon as reasonably possible after a compromise, through the eServices portal.

What does POPIA say about AI marketing and consent forms?

Section 69 prohibits direct marketing "by means of any form of electronic communication, including automatic calling machines, facsimile machines, SMSs or e-mail" unless the person has consented or is an existing customer. The Regulator's Guidance Note on Direct Marketing of 3 December 2024 takes the view that telephone calls are electronic communication. It also lists direct messages on Instagram or LinkedIn. That matters for AI receptionists and voice agents as much as for WhatsApp bots.

Situation	What POPIA requires
Not yet a customer	Ask for consent first, and only once (section 69(2))
Existing customer	Market only your own similar products, with a free chance to object at collection and in every message (section 69(3))
Every marketing message	Identify the sender and give contact details for opting out (section 69(4))
Phone or automated call consent	Record the request and provide the recording or a transcript free on request (Regulations 6.2 and 6.3)
Opt-out lists	Opt-out does not count as consent (Regulation 6.4)

The "POPIA consent form" people search for is Form 4. Under the amended Regulations published on 17 April 2025, you may use Form 4, a form substantially similar to it, or any expedient manner, including email, telephone, SMS or WhatsApp. The guidance note explains that Form 4 asks you to name the goods or services and lets the person choose the communication method. An electronic version may end with "I give my consent" and "I do not give my consent". Keep a list of everyone who refused or objected.

The guidance note also treats bought, rented or shared lead lists as further processing that must meet sections 15 and 18. That is why AI lead qualification works best on inbound enquiries. For WhatsApp specifics, see Conversio's POPIA guide to WhatsApp marketing, and our WhatsApp marketing playbook.

What are the penalties for getting it wrong?

Section 109 lets the Regulator issue an administrative fine of up to R10 million. In setting it, the Regulator must weigh factors including the number of people affected, whether the breach could have been prevented, and "any failure to carry out a risk assessment". Section 107 sets prison terms of up to 10 years for certain offences under the Act.

Where should you start?

Start with checks 1, 2 and 6: a registered Information Officer, a map of your AI tools, and signed operator agreements. They expose most of the other gaps. Then fix marketing consent before you switch on any outbound AI messaging.

AmbitX builds AI agents for South African businesses, such as the Conversio WhatsApp sales agent. Run these checks on any AI vendor you consider, us included. If you are still choosing which processes to hand to AI, read our guide to AI automation for South African businesses.