Last month a Durban physiotherapist asked me whether her new WhatsApp booking bot was POPIA compliant. The answer she wanted was yes or no. The honest answer: it depends on four questions I had to ask her first, and three she had to ask the vendor who sold her the bot. The vendor had told her "fully POPIA compliant" on the sales call. Of course they had. Nothing on their website explained what that meant.
This article is for the business owner in that position. Not a legal textbook. Not an invitation to pay a lawyer R12,000 for an opinion before you can accept a booking. Just the questions that actually matter when AI tools start touching your customers' information in a South African context, and what the Information Regulator seems to care about in practice.
The short frame: POPIA is not a technology law. It is a law about how you handle other people's information. AI tools make that handling faster and more opaque. The Act asks the same eight conditions of a Zoho form, a Make.com scenario, a Claude-powered WhatsApp bot, and a cardboard box of signed consent slips in your reception. The box is slower. It also leaks less.
What POPIA actually says, in one page
POPIA — the Protection of Personal Information Act, 4 of 2013 — applies to any "responsible party" that processes personal information in South Africa. If you have a customer database, you are a responsible party. If your business runs through Pastel and WhatsApp, you are a responsible party. Size does not get you out.
The Act asks you to meet eight conditions when you process personal information. In plain language:
- Collect it for a specific, lawful reason the person understands.
- Collect only what you need for that reason.
- Keep it only as long as you need it.
- Keep it accurate and let the person correct it.
- Keep it secure.
- Be transparent about what you are doing with it.
- Let the person see what you hold on them, or ask you to delete it.
- Only pass it to a third party under conditions that still respect the above.
That is the whole Act, stripped of jargon. Everything else (the regulations, the Information Regulator's guidance, the court cases) turns on how those eight conditions apply to a specific situation. When you plug AI into your business, the question is simply: do my new tools still let me honour all eight? Most SME setups quietly fail two or three of them without anyone noticing.
Where AI tools "touch" personal information in a typical SME
Four patterns cover most of what we see in the Cape Town and Joburg SME market.
The WhatsApp bot. A customer messages a hair salon on the Business API through a provider like 360dialog or Infobip. The message hits a Make.com or n8n flow. That flow sends the message content to OpenAI or Anthropic for a response. The response goes back to the customer. The customer's phone number, name, and the content of what they said have now been processed by at least three parties — the WhatsApp API provider, the automation platform, and the AI model vendor.
The AI meeting notes tool. A sales manager at a Sandton SaaS startup runs Fireflies or Otter on every call. The recordings and transcripts contain customers' names, titles, procurement details, and sometimes sensitive commercial information. Those recordings sit in the vendor's cloud, often in the United States, sometimes available for model training unless the account was configured otherwise.
The CRM enrichment layer. A B2B agency pulls in Clearbit or Apollo to enrich lead records in HubSpot. Personal information about people who never knowingly submitted it to the agency flows in automatically.
The AI inbox assistant. A property manager runs an AI inbox tool on Gmail or Outlook. The tool reads every email in the account, including from a tenant disputing a damages invoice who has no idea their complaint is being parsed by a third party for a draft response.
None of these are illegal under POPIA. All of them can be made compliant. Most of them, as sold off the shelf, are not — because consent, purpose, and transparency are missing somewhere in the chain.
The three questions you need the vendor to answer
Before you buy an AI tool that will touch personal information, make the vendor give you written answers to three questions. If they cannot, that is already your answer.
- Where is the data processed, and under whose law? You need the actual country and the actual legal regime. "The cloud" is not an answer. "AWS" is not an answer either — Amazon runs data centres in Cape Town and also in Virginia, and the two have different legal exposure for your customer.
- Do you use my data to train your models, by default or on request? If yes, that is a problem for sensitive data. If "only if you opt in", confirm it in writing and keep that confirmation.
- How do I get the data back and get it deleted? If a customer invokes their POPIA right to erasure, you need a documented path to delete their data from every system in your chain, including the vendor's logs and backups. "Email our support" is not a documented path.
In my experience, most SME-targeted automation vendors in SA cannot answer the first two questions precisely. Many will send a boilerplate "POPIA compliant" PDF that answers neither. Push once more. If the second answer is still vague, assume the worst and limit what you feed the tool.
Consent, and why most SME consent is quietly wrong
The POPIA regulations do not require signed, witnessed consent for every interaction. They require consent that is informed, specific, and voluntary for the processing of personal information for a defined purpose. The common SME mistakes:
- A tiny "I accept the privacy policy" checkbox on a contact form, which the Regulator does not consider specific consent for marketing.
- No written consent at all for WhatsApp broadcasts, on the assumption that the customer messaged you first.
- Pre-existing customer databases built before 2021 and never re-consented. If you cannot show when and how someone agreed to receive your communications, that database is a liability, not an asset.
The fix is boring and effective: when a customer first interacts with your business, make them answer two clear questions. One for the specific service they came for. One for optional future communication. Store the date, the IP or phone number, and the exact wording they saw. The Information Regulator has published guidance recommending exactly this; it is not a controversial requirement.
Cross-border data: what the Regulator actually cares about
Section 72 of POPIA says you can only send personal information outside the Republic if one of five conditions is met. In practice, for an SME using cloud AI tools, the usable conditions are the recipient being bound by law or binding corporate rules that give adequate protection, or the person consenting to the transfer.
The practical read: if your AI vendor's data processing agreement names a jurisdiction with meaningful data protection law (the EU, the UK, Canada), and your customer has consented to that transfer at onboarding, you are standing on reasonable ground. If your vendor's DPA is silent on where processing happens, or routes through a jurisdiction with no equivalent protection, you have a Section 72 exposure that will become visible if a customer ever complains.
Most Anthropic, OpenAI and Google enterprise agreements will give you a defensible answer. Most consumer-tier accounts will not. The difference is sometimes R400 a month per seat. It is money worth spending once your business is at the stage where a POPIA complaint would actually hurt.
What happens if someone complains
Realistic picture. The Information Regulator is understaffed and backlogged. A single POPIA complaint about an SME almost never results in a dawn raid. What it does result in, in the cases we have seen, is a letter asking for information, a response window of about 30 days, and a request to describe your lawful basis and your controls. If you can produce consent records, a privacy policy that matches reality, and a reasonable explanation of your data flows, the matter typically ends there.
If you cannot — if your consent is theoretical, your policy was copied off a template site, and nobody can describe where customer messages actually go after they hit your bot — the Regulator can issue an enforcement notice. Non-compliance with an enforcement notice is where the real penalties sit, up to R10 million or imprisonment under Section 107. That is a long way from a first complaint, but the distance is covered by whether you can answer the first letter well.
A practical POPIA posture for an SME using AI
You do not need an outside DPO. You do not need ISO 27001. You need five artefacts and the discipline to keep them current.
A short privacy policy on your website that honestly describes what you collect and why. A one-paragraph consent capture at every point someone's information enters your systems. A written list of every third party in your data chain, including AI vendors, with the country of processing and the DPA on file. A documented process for handling a Section 23 access request or a Section 24 correction request within 30 days. And, somewhere accessible, the name of the person in your business who gets the Regulator's letter if one arrives.
Those five things, maintained for a year, put an SA SME in a defensible position for almost any AI stack it chooses to run. They are also, not coincidentally, the hygiene that stops you from making the kind of operational mistake — a leaked database, a mis-sent broadcast, a bot that answers a POPIA question with a hallucination — that creates the complaint in the first place.
The physio asked whether her bot was compliant. By the end of our call she had answers for her three vendor questions and a plan to rewrite her booking consent. The bot stayed. The posture around it improved.
That is usually how this ends when it is done honestly.