Every AI-native government strategy includes a skills line. Usually it is a number: so many officials trained, so many licences issued. Abu Dhabi's Department of Government Enablement, for example, announced in October 2026 a training programme with a stated aim of upskilling 20,000 government employees by the end of the year, and said it followed an earlier rollout of an AI productivity assistant to government staff (Abu Dhabi Media Office). Targets like that are reasonable. They measure reach, and reach matters.
They do not measure the thing an administrative state depends on: whether the official who signs a decision can tell when the AI-assisted recommendation in front of them is wrong. A department can train every employee and still have decision-makers who approve whatever the system proposes, because they lack the time, the confidence or the authority to disagree.
This part is about that second kind of capability. It is not a prompting curriculum. It is the judgement, role design and working conditions that let officials use AI without surrendering their responsibility to it.
Two failure modes: shelfware and shadow AI
Departments that buy tools without building capability tend to fail in one of two directions.
Shelfware. Officials do not trust the tool, do not understand what it is for, or fear being blamed for its mistakes. Usage stays low, the licences renew, and the procurement looks wasteful at the next audit.
Shadow AI. Officials find the approved tool too limited, or there is no approved tool, and they use personal accounts on public services to draft letters, summarise files or answer queries. Personal information leaves the department's control, nobody records what the AI contributed, and the department discovers the practice only when something goes wrong.
Both come from the same gap. The organisation provided tools or policies but not the shared understanding of what AI is for in each role, what good use looks like, and who is accountable. We described the general version in AI literacy is not a prompting workshop. Government adds a constraint: the official is exercising public power, and the law expects them to have actually applied their mind.
What confidence means in each role
The Thukela Provincial Licensing Authority, the fictional licensing authority used in this series, defines AI expectations by role rather than by tool. The question for each role is not "can this person use the assistant?" but "what does this person need to be able to do when AI is part of the work?"
Front-line officers use approved tools, check outputs against the source record, keep personal information within approved systems, and explain to a citizen what the system did and did not decide.
Decision-makers must be able to reject a recommendation and write their own reasons without the system's help. This is the critical role. If a decision-maker cannot do the work unaided, the review is ceremonial, and the decision is effectively the system's. Human-in-the-loop is not administrative justice sets out why that matters legally as well as practically.
Rule owners keep the machine-readable version of each rule faithful to the legal instrument it comes from, approve changes and know which services use each rule. This is a new role in most departments, and it usually belongs with policy or legal staff rather than IT. The rule must be authoritative before the agent can help covers what they maintain.
Service managers read quality signals: override rates, appeal outcomes, cases escalated and time to decision. Usage counts tell them almost nothing.
Accounting officers must be able to say which services use AI, for which decisions, under whose authority, and who answers when something goes wrong. That requires a register, not a vendor briefing.
A ladder, not a course
Capability builds in stages, and not everyone needs every stage.
Every official needs the first two rungs: knowing what the approved tools are for and using them safely. Decision-makers need the third: being able to spot a misapplied rule, weak evidence or poor reasons in a recommendation. A smaller group needs to design scenarios, rules and review steps. Each AI-assisted service needs at least one named owner at the top.
The middle rung is where most programmes underinvest. Training an official to use a tool takes hours. Training a decision-maker to review AI-assisted recommendations critically takes practice on real cases, including cases where the system is deliberately wrong. In the illustrative design, Thukela runs monthly review sessions in which decision-makers work through a set of past cases, some with seeded errors, and compare their conclusions. The aim is not a pass mark but a habit of checking.
Working conditions decide whether judgement survives
Capability is not only skill. An official who has the skill but no time to use it will approve the recommendation. Three conditions decide whether judgement survives contact with a production service:
- Time. If case targets are raised because "the AI does the work now", review time disappears. Targets should be set after measuring how long genuine review takes, not before.
- Authority. Officials must know they may override, and overriding must be easy: one action, with a reasons field, no extra approval chain.
- Protection. An official who overrides the system in good faith, with reasons, should not carry more personal risk than one who follows it. If following the system is the safe career choice, the system is making the decisions.
These are management decisions, not training decisions. They belong in the service design and the delegation framework, and they need the support of the accounting officer.
How to tell whether it is working
Licence counts and training completions are easy to report and nearly meaningless on their own. Better evidence of AI confidence comes from the work itself, as argued in Measure AI capability through work products:
- Override rate with reasons. Zero overrides over months is a warning sign, not a success. It usually means nobody is checking.
- Quality of reasons. Sample decision notices. Can the official's reasons stand without the system's draft?
- Seeded-error detection. In review sessions, what share of deliberately wrong recommendations do decision-makers catch?
- Appeal outcomes. Are AI-assisted decisions overturned more or less often than before?
- Shadow AI signals. Are officials still using unapproved tools? If so, find out what the approved tool cannot do.
The human in the loop must show their work describes the record that makes these measures possible.
The public-service dimension
South Africa's public service has its own training institutions, professional bodies and collective agreements, and any change to roles, targets or performance measures will involve them. AI capability cannot be bolted on through a vendor's training portal. It has to fit job descriptions, performance agreements and delegations, and it has to be negotiated where those are negotiated.
That is slower than a licence rollout. It is also the only route to officials who use AI confidently and remain accountable for their decisions, which is what an AI-native government needs.
Next: Part 9, Who audits the algorithmic state?. All parts are on the series page.
How CloudNala can help
CloudNala helps public bodies design AI capability around roles rather than tools. We map which decisions in a service are AI-assisted, define what each role needs to be able to do, build case-based review exercises with seeded errors for decision-makers, and set up the work-product measures that show whether judgement is holding. We work alongside your human resources, labour relations and training functions, because role and performance changes belong with them.
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