On 2 November 2026, Abu Dhabi's government announced a Machine-Readable Government initiative with OpenAI. A defined set of policies and regulations is being converted into machine-readable form so that AI systems can help officials with routine compliance assessments for permits and certificates. Each assessment is described as traceable to the governing rule, and officials are described as retaining oversight and accountability for final decisions. The first pilot, concerning food imports, is reported to be in production with the emirate's agriculture and food safety authority (Abu Dhabi Media Office).
These are announcements, not measured outcomes, and the first part of this series, What an AI-native government actually means, set out what was and was not claimed. But they make a South African question unavoidable. Could a South African department do something similar? And should it?
This closing part argues that the answer to both is a qualified yes. South Africa should not copy a centrally directed programme from a very different state. It can build machine-readable government the way most durable public infrastructure is built: one bounded service at a time, on shared foundations, with evidence at every gate.
What South Africa already has
South Africa is not starting from nothing. In 2025 the government launched a Roadmap for the Digital Transformation of Government, organised around four pillars of digital public infrastructure: digital identity, digital payments, data exchange and digital services. National Treasury launched a pilot of the data-exchange pillar, MzansiXchange, in October 2025, designed to let government bodies share data without creating a central store (Engineering News; SAnews). The other pillars, including the MyMzansi digital services platform, are still being developed, so departments should check their current status before planning around them.
These foundations matter for machine-readable government. A rule engine that checks a permit application is only useful if it can confirm who the applicant is, fetch the records the state already holds and record a payment. Without them, machine-readable rules sit on top of the same manual coordination steps that Digitising bureaucracy is not transformation argued should be removed.
What South Africa does not yet have, as far as we can find, is shared infrastructure for the rules themselves: authoritative, versioned, machine-readable statements of what a regulation requires, maintained by the bodies that own the regulation. That is the missing layer, and it is the one this series has been about.
The constraints are real
A realistic plan starts with the constraints rather than the ambition.
- Fragmented authority. Powers are spread across national, provincial and local government, and the same activity can need permissions from several of them. Rules that interact are owned by different bodies with different priorities.
- Rules that were never written to be executed. Regulations, by-laws, circulars and delegations are written in prose, often with discretion deliberately built in. Some are inconsistent. A policy document is not an executable rule explains why translation is a policy act, not a technical one.
- Uneven capability. Some departments have strong digital and legal teams. Many municipalities have neither.
- Procurement and audit exposure. As Procurement is the first test of AI-native government argued, a poorly specified AI contract becomes an audit finding.
- Trust. Citizens who have experienced unreliable public services will not assume an AI-assisted one is better. It has to be visibly accountable.
None of these makes machine-readable government impossible. Together they rule out a national moonshot.
Choose the first service carefully
The first service sets the pattern for everything after it, so it should be chosen for learning value and low harm, not for headline impact. Two questions sort candidates: how clear are the rules today, and how much harm does a wrong answer cause?
Good first candidates share six properties:
- High volume, so the improvement is measurable.
- Rule-bound, with conditions that are already applied consistently.
- Low discretion, so the official's judgement is focused on exceptions.
- Reversible, so an error can be corrected without lasting harm.
- Data available, ideally from records the state already holds.
- Owned, by a body that controls both the rule and the service.
Completeness checks for event permits, routine licence renewals for holders with clean records, and verification of certificates the state itself issued are typical examples. Discretionary social benefits, enforcement decisions and anything difficult to reverse are not, and should wait until the foundations and the oversight record are mature.
A three-horizon path
The following path is illustrative. It is not a plan any South African government has adopted, and the timelines will vary with the service and the institution. Its value is in the gates between stages.
Months 0 to 6: one service, shadow mode. A single body translates the rules for one service into machine-readable form, with legal and policy staff owning the translation. It builds a scenario library from past cases, as described in The government AI evaluation harness, and runs the system beside officials without affecting decisions. Every recommendation is compared with what the official decided. The case-file record from Who audits the algorithmic state? is in place from day one.
Gate: the system's recommendations agree with correct official decisions at an agreed level on the scenario library and in shadow mode; disagreements have been analysed; the rule owner has signed off the translation; and independent reviewers who did not build it have tested it.
Months 6 to 18: live with review, and a shared registry. The service goes live with officials deciding and the system assisting. The rules move into a registry that other services can reuse, with versioning, effective dates and owners, as described in The rule must be authoritative before the agent can help. Two or three more services, ideally in other bodies, adopt the same pattern. The first external audit takes place.
Gate: appeal and override data show decisions are at least as sound as before; the recourse loop has been exercised; the audit raised no unresolved material findings; and a second body has successfully reused the registry.
Months 18 to 36: shared rule infrastructure. The registry becomes cross-departmental infrastructure, connected to the data-exchange and identity foundations. A public register of AI-assisted services is published. Capability building moves from pilot teams to the wider public service, along the lines described in AI-native government needs AI-confident officials.
At no stage does the model become the authority. The model should not be the law explains why the rule registry, not the model, is the source of truth, and why officials remain the decision-makers.
What should not be copied
Abu Dhabi's approach combines central direction, concentrated investment and partnerships with frontier AI companies. Some of that will not transfer to South Africa, and some of it should not.
- Do not start with a single strategic AI partner. South Africa's procurement framework, competition concerns and sovereignty interests all favour open rule formats and supplier-neutral infrastructure. Sovereign AI is more than data residency sets out the control model.
- Do not measure success by announcements. Measure time to a lawful decision, appeal outcomes and the number of services reusing shared rules.
- Do not skip the rule work. Machine-readable government is mainly a legal and policy discipline that uses technology, not the other way round.
What is worth adapting is the core design principle in the announcement: assessments traceable to the governing rule, with officials accountable for final decisions. That principle is consistent with South African administrative law, and it is achievable.
The series in one paragraph
Across ten parts, this series has argued that AI-native government is not a model governing. It is public services redesigned around authoritative rules, complete evidence, officials with real authority and the capability to use it, procurement that buys control and exit, and oversight institutions that can reconstruct any decision. South Africa can build that, starting with one well-chosen service and the discipline to pass each gate before the next.
Start again at Part 1, What an AI-native government actually means, or see all parts on the series page. The technical layer underneath this series is covered in Machine-Readable Policy and Accountable Decision Systems.
How CloudNala can help
CloudNala helps public bodies take the first bounded step towards machine-readable government. We help select a suitable first service, work with your policy and legal staff to translate its rules into a versioned, testable form, build the scenario library and case-file record, and run a shadow-mode pilot with clear evidence gates. Our aim is a pattern your institution owns and can reuse, not a dependency on us or on any single supplier.
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