AI Assurance Logo
AI Assurance aiassurance.co.za
Get In Touch

Keeping the Framework Alive

Regulation 4(a) does not end when a framework is written. The Information Officer must ensure it is implemented, monitored and maintained. That is the difference between a POPIA file and a working system.

AI makes the difference visible. A tool assessed in the procurement year may be a different product a year later. Vendors change models, add training uses, move regions and switch on new features. If the framework is not watched, the organisation is accountable for processing it no longer understands.

King V asks those charged with organisational oversight to exercise that oversight on information they can rely on. A framework that is never tested does not give them that information.

What monitoring means under POPIA

Monitoring is not a dashboard slogan. It is checking whether the conditions for lawful processing still hold in the processes that actually run.

For AI-supported processing that includes:

Purpose and further processing

Is the system still used only for the purpose that was specified and assessed? Has a chatbot become a performance scorer? Has a hiring tool started feeding unsuccessful CVs into a training set?

Minimisation and quality

Is more personal information being sent to the model than the PIIA described? Are outputs treated as facts when they are inferences?

Operators

Are the contracted security measures still in place? Have subprocessors changed? Has the supplier's data-use policy changed?

Rights

Can an access or correction request still be completed if the information sits in a vendor model, log or ticket store? Time spent finding out during a live request is evidence that monitoring was not done.

Security measures

Section 19 requires safeguards updated as risks change. An AI system introduces assets that older reviews never listed: prompt stores, embeddings, fine-tunes, vendor training pipelines. Verification has to include those.

Useful monitoring is sampled and scheduled. Pull a set of automated rejections and see whether a human could have intervened. Submit an internal test request for a data subject whose information is known to sit in an AI tool. Read the supplier's current terms against the operator contract. Record what was checked and what failed.

What maintenance means

Maintenance is what happens when monitoring finds a gap, when the estate changes, or when something goes wrong.

Change

A new model, a new purpose, a new operator or a new automated cut-off should update the inventory, reopen the PIIA where the change is material, and amend notices and contracts if needed. "We'll catch it at the annual review" is how unauthorised further processing becomes normal.

Non-compliance

If a team has adopted a tool that sends customer files to an unassessed model, the framework must provide for stopping or constraining that use, not only for writing a lesson-learned note.

Security compromises

Section 22 applies where there are reasonable grounds to believe that personal information has been accessed or acquired by an unauthorised person. AI-related events can meet that test: a prompt or output containing personal information sent to the wrong party, an unauthorised training use, a vendor breach of a dataset the organisation provided. The process must be able to receive those events from IT or from the supplier and then assess notification to the Regulator and to data subjects.

Regulator and data-subject pressure

Issues raised by the Information Regulator, and complaints that reveal a broken rights path or an unexplained automated outcome, should change the framework. If they only change a single case file, the same failure will recur.

Evidence that the work is being done

Assurance is the ability to show what was decided and what was checked. For an Information Officer and for a board pack, useful evidence looks like this:

These records should be produced as the work happens. Building them after a complaint is reconstruction, not maintenance.

Incidents that never reach the privacy function

A recurring failure is technical monitoring that stays in operations. Model-quality alerts, vendor status pages and data-loss tools may already show an event that is a POPIA problem. The living framework needs a defined path from those tools into the Information Officer's process: when to escalate, who assesses Section 22, who owns communication with the supplier.

Without that path, the organisation can be technically alert and legally late.

What the governing body should see

King V does not require directors to inspect models. It does require them to oversee technology and information risk with enough quality of information to do so.

A short, regular view is enough if it is honest:

That is oversight of organisational risk. It is not management of the systems. Management remains the work of running, monitoring and fixing the processing. The framework is how that work stays aligned with POPIA and visible to those who must account for it.

International practice in its place

The idea that control must continue after an AI system is released appears in the EU AI Act's post-market monitoring and in EN 18286's living QMS. ISO/IEC 42001 similarly requires the management system to be maintained and improved. Those sources are useful where they apply, and useful as method.

For a South African responsible party the binding loop is already in POPIA: identify processing, assess impact, implement measures, verify them, update them when risk changes, and handle compromises. AI does not invent that loop. It makes the cost of leaving it idle obvious.

What "alive" looks like

The framework is alive when:

That is AI assurance in this market: POPIA applied to processing that moves, King V applied to technology risk that is reported, and a compliance framework that still matches the systems in use.