Who signs for the agent?

There is a question that ends most AI governance discussions in this country, and it takes six words.

Who signs for the agent?

Not which team owns it. Not which platform it runs on. Which human being would be asked to explain its decision to a regulator, an auditor or a court, and what would they produce when asked.

If you cannot name that person, you do not have governance. You have a diagram.

No single AI Act is coming

A great deal of local commentary has been waiting for South Africa's equivalent of the EU AI Act, on the assumption that a single statute would eventually arrive and tell everyone what to do.

That is not the model government has chosen. The Draft National AI Policy, gazetted on 10 April 2026 with public comments closing on 10 June, sets out a risk-based, sector-specific, multi-regulator approach. Oversight is to be embedded within the mandates of the regulators that already exist rather than concentrated in a new one.

For financial services that means the obligation lands inside supervision you are already subject to: the Financial Sector Conduct Authority, the Prudential Authority, the Reserve Bank and the Financial Intelligence Centre. For anyone expecting a compliance checklist, this is worse news than a statute, not better. A checklist can be worked through. A supervisory expectation arrives as a question in a review, and it arrives without notice.

The policy is risk-based and modelled in part on international frameworks, with stricter requirements anticipated for high-risk applications in sectors including healthcare, financial services and law enforcement. Finalisation is targeted for the 2026/27 financial year, with sector-specific instruments expected to follow from 2027/28.

Read that last date again, then look at your AI roadmap.

The collision nobody has diarised

Most enterprise AI roadmaps I have seen promise something like an agent ecosystem, cross-organisational automation or autonomous operations somewhere in the 2027 to 2029 band.

South Africa's sector-specific AI regulatory instruments are expected from 2027/28.

Those are the same window. One of them is a commitment your organisation made to itself and controls entirely. The other is a constraint arriving from outside, on a timetable set by people who have never seen your programme plan.

I am not predicting that regulation will stop anything. I am pointing out that the two dates were set independently, that almost nobody has put them on the same page, and that the organisations which do so this year will be the ones shaping the sector instruments rather than absorbing them. The consultation process is how that happens, and it is open to institutions that show up.

POPIA section 71 is already in force

Here is what I find most striking about the local conversation: a great deal of it is forward-looking, and the binding constraint is behind us.

Section 71 of the Protection of Personal Information Act restricts decisions that affect a person being taken solely on the basis of automated processing. It has been in force for years. It predates every agent in every organisation reading this.

If you are running credit decisioning, claims adjudication, benefit eligibility, fraud flagging or anything else where a machine's output materially affects an individual, the question of meaningful human involvement is not a 2028 problem. It is a today problem, and "we are waiting for the AI policy to be finalised" is not a defence, because the relevant law was never about AI in the first place.

The draft policy explicitly aligns with POPIA and references automated decision-making. The direction of travel is toward reinforcement of an existing obligation, not the creation of a new one.

The public sector version

For departments and state entities, the governing logic is different in form and identical in substance.

Public accountability turns on documented process. If you cannot produce the decision trail, it is an audit finding regardless of whether the outcome was correct. An agent that reached the right answer through a route nobody can reconstruct has failed the test before anyone gets to the question of whether it was right.

That is a higher bar than most private-sector governance imposes, and it is worth saying plainly because it changes the design. In a department, explainability is not a nice-to-have that improves trust. It is the thing being audited.

Three artefacts, and none of them need budget

When I am asked what good governance looks like in practice, I give the same three answers, because they are the ones that produce an answer to the question in the title.

A registry. Every agent running in production, with a named human owner and a risk tier. Not a team, a name. This can start as a spreadsheet this month, and the act of compiling it usually surfaces two or three agents nobody knew were live.

A decision record per action. What the agent did, on whose authority, using which data, at what time. If your logging captures the API call but not the intent and the authorisation, you have telemetry rather than an audit trail, and the difference only becomes visible when someone asks.

Validation evidence an auditor can read. Not a notebook. Not a Slack thread. A document showing what the system was tested against, what the acceptable performance was, who accepted it, and when it was last re-validated. Model risk management has demanded this of banks for years, and the discipline transfers directly.

None of these three requires new budget. All three require somebody senior to decide that they are being done, which is a different and harder kind of scarcity.

The sentence to avoid

Every organisation deploying agents at scale will eventually be asked why a particular decision was made. There is one answer that has never worked in any regulated context and will not start now.

"The model decided."

It was not an answer when it was a credit scorecard, it was not an answer when it was a rules engine, and it is not an answer when it is an agent. Somebody signed. The only question is whether your organisation knows who, and whether that person can produce the evidence without a three-week project to assemble it.