Compliance is usually pitched as the thing that slows AI down. For agents, it's closer to the opposite.
This is an operator's perspective, not legal advice. Take advice on your own obligations from an attorney or your information officer.
Businesses adopting AI agents tend to ask two questions in the wrong order: how fast can it work, and is it compliant. In South Africa, the second question shapes the answer to the first.
The regulator's position
At a May 2026 webinar, the Information Regulator's chairperson, Advocate Pansy Tlakula, said AI may offer efficiency and competitiveness but must comply with existing law, and that POPIA should remain the legal backbone of AI regulation in South Africa, with any AI-specific rules complementing it rather than replacing it. (See the Bowmans overview of POPIA considerations for AI in the workplace.)
The practical reading is straightforward. You don't have to wait for an "AI Act" to be careful. The Protection of Personal Information Act already applies to every AI tool that touches the personal information of your clients, employees and suppliers.
Where AI and POPIA rub against each other
Commentators point to a handful of tensions:
- Purpose. POPIA expects personal information to be processed for a specific, lawful purpose. AI systems that find new patterns can drift beyond the purpose the data was collected for.
- Transparency. POPIA expects openness and accountability. Many AI systems can't explain how they reached an answer.
- Decisions about people. Section 71 deals with automated decision-making, and it's the one most relevant to agents.
The Regulator has also flagged risks such as excessive collection, profiling, continuous monitoring, poor data quality, discriminatory outcomes, opaque decision-making and weak security safeguards.
Section 71, in plain English
Section 71 protects a person from being subject to a decision that has legal consequences for them, or significantly affects them, where that decision is based solely on automated processing of their personal information. There are exceptions and conditions, and what applies to you depends on the facts.
The word that matters for agents is solely.
An AI agent that approves or declines a claim, a credit application or a job candidate with no person in the loop is on the wrong side of the line this section draws. An agent that prepares the recommendation, shows its reasoning and waits for a named, accountable person to make the decision is a very different thing.
That's the case for human approval, and it isn't only a compliance argument. A person who reviews and signs off can also catch the cases the model got wrong.
What good looks like for an AI agent
If you're evaluating or running an AI agent that touches personal information, these are the controls worth insisting on.
1. Mandatory human approval for anything consequential. Not optional, not a setting buried in a menu. The AI prepares; a named person decides.
2. A real, accountable sender. If an email goes to a client, it should go from an actual employee's address, so the client knows who they're dealing with and who is answerable.
3. A complete audit trail. Who accessed which record, what the AI drafted, who approved it, and when. If a data subject asks what happened with their information, or you have to investigate an incident, this is how you answer.
4. Role-based access. People and modules should see only what they need. Cost, margin and personal data shouldn't be visible to everyone by default.
5. Sources and honesty. An answer that cites where it came from, and says when it couldn't reach some data, can be checked. One that doesn't can't.
6. Security you can describe. Encryption at rest and in transit, and clear answers on where data is stored and who can access it.
7. Data minimisation. The platform should work from the information the task needs, not hoover up everything.
Compliance and usefulness point the same way
It's tempting to see these controls as friction. In practice they're what make an AI agent usable in a business that answers to clients and a regulator.
Think about it from the point of view of the person running the practice. You don't want an AI that acts alone, because you're the one who answers for it. You want one that does the typing, filing and chasing and brings you the decision. The controls the regulator would want are the same ones a sensible principal wants.
How Clara approaches it
Clara is built around this idea, and it was a design decision from the start rather than a feature added later.
- Approval is mandatory: Clara prepares work and a named person approves it before it goes to a client, supplier or system.
- Outbound email goes from the real employee's address, not a shared robot account.
- Every record, document and approval is logged with who, what and when.
- Permissions are role-based, by module and policy type.
- Data is encrypted at rest and in transit, in support of POPIA and GDPR compliance.
"In support of" is deliberate. No platform makes you compliant on its own. Your obligations as the responsible party remain yours, and the platform should make them easier to meet, not harder.
A short checklist before you deploy any agent
- What personal information will it touch, and for what purpose?
- Which decisions affect people significantly, and who signs those off?
- Can you produce a log of what it did and who approved it?
- Who has access, and is that limited by role?
- Has your information officer seen the plan?
If you can answer all five, you're ahead of most. For the broader picture of how to evaluate these platforms, see Agent Washing: How to Tell If Your AI Agent Is Just a Chatbot and What Is an AI OS Platform?.