POPIA restricts decisions with legal or substantial effects on a person that are based solely on automated processing which profiles them, for example on creditworthiness, reliability or conduct. An AI agent that approves or declines on its own falls inside. It is allowed under a contract or law only with safeguards: a chance to make representations, and information about the logic.
What does POPIA require on automated decisions?
The Protection of Personal Information Act contains a specific protection against automated decision-making. In substance:
- A person may not be subject to a decision that has legal consequences for them, or affects them to a substantial degree, where the decision is based solely on the automated processing of personal information intended to build a profile of them. The Act gives examples of what such a profile covers: performance at work, creditworthiness, reliability, location, health, personal preferences or conduct.
- The restriction does not apply where the decision is taken in connection with concluding or performing a contract and either the person's request has been met or appropriate measures protect their legitimate interests; or where the decision is governed by a law or code of conduct that specifies appropriate measures.
- Appropriate measures must give the person an opportunity to make representations about the decision, and require the responsible party to give sufficient information about the underlying logic of the automated processing.
The Information Regulator enforces POPIA. For POPIA's wider conditions, see whether POPIA applies to AI agents.
Who does it apply to?
Responsible parties processing personal information in South Africa, which includes banks, insurers, lenders, employers and service providers. Because POPIA covers juristic persons too, decisions about business customers, such as automated credit limits for small companies, also fall within it.
What does it mean when an agent makes the decision?
Three questions decide whether an agent's work triggers the protection.
Is it solely automated? If a person reviews the agent's recommendation and has the authority, information and time to change it, the decision is not solely automated. If the person approves batches without looking, it may still be. Regulators elsewhere take the same view; the ICO's approach to automated decisions by agents is a useful reference for what meaningful review looks like.
Does it profile the person? The protection targets decisions based on processing that evaluates personal aspects. An agent that checks whether a document is signed is not profiling. An agent that assesses whether a claimant's account of events is credible, or whether an applicant can repay, is.
Are the effects legal or substantial? Declining credit, cancelling cover, refusing a claim or flagging an employee for discipline clearly qualify. Routing a query to a queue usually does not.
Agents complicate this because one agent may do several things, only some of which are decisions in this sense. Assess each capability. An agent's authority to execute should be lowest for capabilities that both profile and produce substantial effects.
How should agents meet the exceptions and safeguards?
Where you rely on the contract or law exceptions, the appropriate measures must actually work for agent decisions.
Representations. Every decision communication should say how the person can respond, and the response should reach a person with authority to change the outcome, not the same agent.
Information about the logic. "An AI model decided" is not enough. Useful information names the main factors and the information relied on. An agent that records which facts it used, and where each came from, makes this possible to produce for every decision.
Take an illustrative case. A South African microlender uses an agent to assess small loan applications. The agent reads bank statements and affordability data, then recommends approval or decline. For approvals within a set amount, it acts on its own, under the contract exception, since the applicant's request is met. For declines, a credit officer reviews each recommendation with the agent's factors and sources in view. Decline letters explain the main factors and invite the applicant to make representations to a named team. Override rates on declines are tracked to show the review is real.
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Identify decisions in scope | Register of agent capabilities that profile people and have legal or substantial effects | Capability register with flags |
| Solely automated or not | Meaningful human review where claimed; override rates tracked | Review records; override statistics |
| Contract or law exception | Ground recorded per capability | Grounds register |
| Opportunity to make representations | Route to a person with authority to change the decision | Representation log and outcomes |
| Information about the logic | Decision record lists factors and sources used | Per-decision factor and source record |
| Fairness and accuracy | Outcome testing across groups and on past cases | Test and monitoring reports |
What will the Information Regulator or an auditor ask?
- Which agent decisions are based solely on automated processing?
- What makes your human review meaningful, and how often do reviewers change outcomes?
- What do you tell people about the logic of an agent's decision?
- How can a person make representations, and who considers them?
- Which exception do you rely on for each automated decision?
What does this protection not cover?
It does not ban automated decisions; it restricts them and attaches safeguards. It does not cover decisions without legal or substantial effects. It does not set credit or insurance conduct rules, which sit with financial regulators. Comparable provisions exist in other markets: see NDPA automated decision-making and agents for Nigeria and what automated decision-making under GDPR is for Europe. For the full map, see our guide to AI agent regulation mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where Pryme fits
In Pryme, a capability that decides about people can be held at depth 2, so the agent drafts and a named person decides, until Attest evidence on the firm's own cases supports more. Every verdict is filed with the facts and sources used, which is the material a decision explanation and a representation review need.
Questions people ask
Does a human signature make an agent's decision non-automated?
Not by itself. The person must actually consider the case and be able to change the outcome. A signature applied without review leaves the decision solely automated in substance.
Does the protection cover decisions about business customers?
It can. POPIA protects juristic persons, so automated decisions about a company's credit or cover that rest on profiling may fall within it.
How detailed must the explanation of logic be?
Detailed enough for the person to understand the main reasons and respond. Naming the key factors and the information relied on usually meets the need better than a technical description of the model.