Every public-sector AI proposal now includes a reassuring sentence: a human will remain in the loop. The Abu Dhabi announcements discussed in Part 1 use a stronger version: officials retain oversight and responsibility for final decisions. Both are the right starting point. Neither, on its own, describes a lawful or fair administrative process.
In South Africa, decisions by public bodies that affect people's rights are administrative action, and administrative action has to meet standards set by section 33 of the Constitution and the Promotion of Administrative Justice Act 3 of 2000 (PAJA). Those standards are about the decision and the process, not about whether a person was somewhere in the workflow. An official who approves forty AI-prepared assessments a day without the time or information to disagree is in the loop. Whether administrative justice is being done is a separate question.
This part is general information about system design, not legal advice. Public bodies should take legal advice on how PAJA applies to their specific decisions.
What PAJA expects, in brief
PAJA is detailed, and its application depends on the decision and the circumstances. In summary, four parts of it matter most for AI-assisted decisions.
Section 3 requires administrative action that materially and adversely affects a person's rights or legitimate expectations to be procedurally fair. In general that includes adequate notice of the nature and purpose of the proposed action, a reasonable opportunity to make representations, a clear statement of the action, adequate notice of any right of review or internal appeal, and adequate notice of the right to request reasons. What fairness requires depends on the circumstances of each case.
Section 5 gives a person whose rights have been materially and adversely affected, and who was not given reasons, the right to request them, generally within 90 days of becoming aware of the action. The administrator must then provide adequate written reasons within 90 days of the request. If adequate reasons are not given, the action is presumed, in judicial review proceedings and in the absence of proof to the contrary, to have been taken without good reason.
Section 6 sets out grounds on which a court can review administrative action. Several are directly relevant to AI assistance: taking irrelevant considerations into account or failing to consider relevant ones, and acting because of the unauthorised or unwarranted dictates of another person or body. The Department of Justice's PAJA pages give a plain-language summary.
The final ground is worth pausing on. A system is not a person or body, but if an official signs whatever it recommends, and especially if a supplier configured it, it becomes arguable that the official did not apply their own mind and that the decision was in substance dictated from outside rather than made by the person with the power to make it. How a court would treat that in an AI-assisted case has not, as far as we can find, been tested in South Africa, which is a reason to design against it rather than rely on it.
From legal duty to system capability
Officials meet these duties, not systems. But an AI-assisted process can make the duties easier or much harder to meet. The design question is what the system must support.
Using the fictional Thukela Provincial Licensing Authority, an invented body whose AI assistant prepares liquor-licence renewal assessments, the mapping looks like this.
Notice before a refusal. Where the assessment points towards refusal, the system drafts a notice to the applicant explaining the proposed action and the facts and rule it rests on, and the official sends it before deciding. The AI's preparation becomes the basis for a fair hearing, rather than a substitute for one.
Representations that reach the decision-maker. The applicant's response is attached to the case file and shown to the official alongside the original assessment, with any new facts re-checked against the rule. A representation that disappears into a mailbox does not count.
Notice of rights in every decision. Every decision notice states the internal appeal route and the right to request reasons. This is easy to template and easy to forget.
Reasons that are actually the reasons. Adequate reasons explain which rule was applied, which facts were found and how the facts led to the outcome. Because the system in Part 2 cites the rule version and clause, and shows the extracted facts separately, the reasons can be generated from the real basis of the decision. The danger is fluent, AI-written reasons that sound right but do not correspond to what the official actually weighed. Reasons must record the official's decision, including where the official departed from the recommendation.
A reviewable record. If a decision reaches court, the authority will need to produce the record. The case file should contain the application, the evidence, the rule version, the AI's recommendation and trace, the official's decision, any departure and its reasons, and every notice sent. The case file part of the Machine-Readable Policy series describes that record in detail.
Four conditions for real oversight
The hardest problem is not notice or reasons. It is making sure the official is genuinely deciding.
Evidence. The official must see more than a conclusion. They need the rule and version applied, the facts the system extracted and where it found them, the points the system was unsure about, and anything it could not decide. A recommendation with no visible basis invites either blind trust or a complete redo.
Authority. The official must be able to reject or change the recommendation without friction or penalty. If disagreeing takes five times as long as agreeing, or is noticed by a supervisor measured on throughput, the incentive is to agree.
Time. Caseloads must allow a real look at each case. If an AI-assisted process triples the number of cases per official, the authority has to show that the extra volume came from removing clerical work, not from removing scrutiny.
Recourse. The applicant must be able to appeal to someone other than the official who decided, and the appeal must examine the decision afresh rather than confirm the original assessment.
These conditions are a CloudNala framework, not a legal test, but each of them bears on whether a court would see a decision as the official's own. They also echo the earlier AI Before the Rulebook series, which set out how assistance quietly turns into authority, and the Agent Claims Decoded series, which treats human intervention as a design problem rather than a button.
Measure whether oversight is still real
Oversight decays without anyone deciding to remove it. The authority should watch a small number of signals.
| Signal | What it suggests |
|---|---|
| Agreement rate rising towards 100 percent | Officials may have stopped scrutinising, or the system is genuinely reliable; investigate which |
| Time per case falling below a credible minimum | Review has become a click |
| Departures concentrated in a few officials | Others may feel unable to disagree |
| Appeals succeeding on AI-prepared cases | The process is producing wrong or poorly reasoned decisions |
| Reasons that are identical across different facts | Reasons are templates, not explanations |
A high agreement rate is not automatically bad. It is a question that needs an answer. Thukela's design includes a monthly sample of agreed decisions that a second official reviews from scratch, without seeing the recommendation. If the second official disagrees often, the first official's agreement was not real oversight.
Appeals are part of the architecture
An internal appeal is often treated as an administrative afterthought. In an AI-assisted process, it is also the main feedback loop. Every successful appeal tells the authority something: the rule was encoded wrongly, a fact was extracted wrongly, the official did not have the information needed, or the reasons did not explain the outcome. Feeding those findings into the scenario library from Part 3 and into the rule lifecycle turns appeals into a correction mechanism.
The appeal process also needs its own design: the appeal official sees the full case file but decides independently; the applicant can add evidence; and the outcome is recorded against the original case. The appeals part of the Machine-Readable Policy series goes further into contestability and reassessment.
Where personal information is involved, POPIA adds its own requirements about decisions based solely on automated processing, which our earlier piece on a route to challenge AI-assisted decisions covers. PAJA and POPIA overlap but are not the same, and a design that satisfies one does not automatically satisfy the other.
Start this week
- For one AI-assisted or soon-to-be AI-assisted decision, write down what notice, representation, reasons and appeal steps exist today.
- Check whether the official can see the rule, the facts and the uncertainties, or only a conclusion.
- Time how long it takes to agree with a recommendation versus to depart from it.
- Pull agreement rates by official for the last quarter.
- Ask legal advisers which of your decisions are administrative action and what fairness requires for each.
Next: Part 5, Sovereign AI is more than data residency. All parts are on the series page.
How CloudNala can help
CloudNala works with public bodies, alongside their legal advisers, to design AI-assisted decision processes that officials can stand behind. We map the notice, representation, reasons and appeal steps, design the case view so officials see the evidence and not just a conclusion, define the oversight signals worth monitoring and build the record a reviewer or court would need. We do not give legal opinions; we make sure the systems support the duties your lawyers identify.
Work with CloudNala
CloudNala helps organisations move from technology ambition to practical execution across cloud, AI, data, platform engineering and digital services.
Whether you are exploring AI, modernising your cloud environment, building a public-sector digital service, or turning an idea into a working MVP, we can help you shape the roadmap and deliver the next step.
Start a Digital Discovery Sprint or write to us at consult@cloudnala.co.za