A decision layer is not useful because the country is named in the prompt. It is useful when it fits the way an organisation actually operates.
South African organisations often serve multilingual customers through mobile channels, while running a mix of modern cloud services and older core systems. Connectivity is uneven, specialist capacity is scarce, regulatory duties are real, and the data that arrives is rarely tidy. A typed decision service such as Jev earns its place in that environment only when it takes a repeated, bounded judgement off a person's desk and leaves everything consequential where it already sits.
This part works through six candidate patterns. The first is the series' running example. Kloofvale Local Municipality is fictional, and its systems and figures are illustrative. The other five are patterns, not descriptions of any named organisation's systems. As throughout this series, this is an independent CloudNala analysis; we have no commercial relationship with Jev AI or TypeSafe.
In plain language:
In every one of these examples the model answers a narrow question, and the organisation still decides what happens next.
1. Kloofvale Fix-It: service-report triage
Kloofvale's residents report service problems through Fix-It, a WhatsApp and web channel. About 1,400 reports arrive in a normal week, in English, isiXhosa and Afrikaans, often code-switched and often missing a location. Two call-centre staff read every one and route it by hand to roads, water, electricity or other. The queue is first in, first out. On one storm weekend in March, a burst-main report waited 31 hours behind reports about uncollected refuse.
The decision questions are small and well defined:
- Choice: roads, water, electricity or other.
- Score: urgency, on a three-level rubric the service managers wrote: routine, priority, urgent.
- Noul: is there enough information to find the problem?
The application policy around those questions does the real work:
- An existing keyword rule flags emergency wording and sends it to the 24-hour control room before any decision call is made. That rule does not change, and Fix-It is never presented as an emergency line.
- About a third of reports lack a usable location. When the Noul says there is not enough information to find the problem, the resident gets an immediate reply asking for a street name or a location pin, rather than a guess.
- Only a high-confidence routine report, with the chosen route at a probability of 0.85 or more and the needs-human signal below 0.20, goes straight to the right department's queue.
- Everything else lands in the triage queue the call centre already works.
The measure of success is not "accuracy". It is how much faster an urgent water report reaches the water team, and whether anything gets lost on the way. Parts 7 and 9 return to how Kloofvale measures both.
2. Medical-scheme service routing
A member writes in about benefits, a pre-authorisation, a claim or finding a provider. A Choice question can pick the service queue, and a Noul can flag a message that needs a person, for example a complaint or a vulnerable member.
What it must not do is decide clinical eligibility because the message happened to be easy to classify. Routing a message to the authorisations team is a different act from authorising treatment.
Health information is special personal information under POPIA, which starts from a prohibition with exceptions. A routing decision rarely needs the clinical detail in the message. The state builder should strip diagnoses, medication and test results unless the routing question genuinely depends on them, and the scheme's privacy officer should agree that list before anything leaves the building. Part 8 covers the due diligence in more detail.
3. Banking-agent tool gate
An AI assistant inside a bank proposes an operation on a customer's behalf: read a balance, update a notification preference, initiate a transfer. Before the operation runs, the harness can ask a set of typed questions:
- Choice: read-only, reversible write, financial action or restricted.
- Score: potential customer impact.
- Noul: does this need the customer's explicit confirmation?
The bank's identity checks, consent records, transaction limits, fraud engines and payment rails stay authoritative. A decision score can send an operation for more scrutiny. It cannot authorise the movement of money. Our Agentic Banking series covers why the agent channel needs its own controls in the first place.
4. Tender and opportunity classification
A platform ingests tender notices from national, provincial and municipal sources, each with its own format and quality. A decision service can classify the sector, estimate relevance to a supplier's capabilities and flag notices whose documents are incomplete.
The safe objective is to prioritise review, not silently remove notices. A notice the model calls "non-ICT" should drop down a list, not out of it. The outcome measure is therefore not classification accuracy alone. It includes the relevant opportunities that were missed and the reviewer hours saved, and a single missed tender can outweigh a month of saved hours.
5. Retail support and payment-failure routing
Customer messages to an online retailer or payments business describe duplicate charges, failed payouts, delivery problems or account access issues, usually with more emotion than structure. A Choice question routes the case, a Score estimates urgency, and a Noul asks one narrow thing: is the customer explicitly asking for a refund?
That last signal is useful for prioritising. It is not a refund. Approval still follows the account, fraud and finance controls that existed before any model arrived.
6. Public-sector document intake
A department receives forms, letters and service requests through counters, email and online portals. After trusted extraction turns scans into text, a decision service can classify the document type and ask whether the required information is present.
This is the pattern most at risk of turning into a hidden administrative decision-maker. "Incomplete" can quietly become "rejected" if nobody designs against it. Decisions that refuse, grant or affect someone's rights need explicit legal authority, recorded reasons, an appeal path and human oversight. The decision service can sort the post. It cannot decide the application.
What the six have in common
Look across the six and the same shape repeats:
- A high volume of messy, human-written input.
- A small set of outcomes the organisation can name in advance.
- A safe place for uncertain cases to go.
- A consequential action that stays behind existing authority.
Where any of those is missing, the pattern weakens. A decision that happens a handful of times a month does not need a model. A decision with no safe fallback should not be automated. Part 10 turns this into a selection test.
Language and code-switching
South African production evaluation has to include the language people actually use. For Kloofvale that means:
- English as it is written locally, including local idiom.
- isiXhosa and Afrikaans, the two other languages Fix-It sees most.
- Code-switched messages that move between languages mid-sentence.
- Informal spelling, abbreviations and SMS-style shorthand.
- Local place names, ward numbers and the municipality's own service vocabulary.
A message such as "Molo, the pipe e-corner ye-Main Rd ivuza since yesterday, water everywhere" mixes isiXhosa and English, names a landmark rather than an address and never uses the word "burst". A routing question has to handle it, and the evaluation set has to contain it.
The Jev AI Agent Skill's configuration lists en-US as its default language and zh-CN (Simplified Chinese) as an option. That is a setting for the skill's guidance text. It says nothing about how well the underlying model handles isiXhosa, Afrikaans or code-switched messages, and nothing in the public documentation we reviewed on 13 October 2026 makes that claim either. The only way to know is to test each important cohort separately, with real examples, and keep a review route for the cohorts that do not perform.
Kloofvale's rule is simple: the weakest cohort sets the automation boundary. If code-switched reports are routed correctly far less often than English ones, automatic routing is limited to the cohorts that earn it, and the rest keep going to people. That is uncomfortable, because it can mean the residents who already find services hardest to reach get the slower path. The answer is to improve the cohort, with better examples and clearer criteria, not to hide it inside an average.
Key points
Local value comes from operating reality. Multilingual, mobile-first, mixed-estate organisations with messy input are where a narrow decision layer can help.
The model answers a narrow question. Routing, urgency, completeness and risk class are good candidates. Eligibility, refunds, payments and rights are not.
Uncertainty needs somewhere to go. Every pattern above has a review queue or a clarifying question, and every one keeps emergencies and consequential actions out of the model's reach.
Measure the outcome, not only the classification. Missed tenders, lost reports and slower urgent cases are the numbers that matter.
Test each language cohort. A configuration option is not evidence of accuracy. The weakest cohort sets the boundary.
Previous: Part 5, Jev and BMAD together. Next: Part 7, thresholds, evals and human review. All parts are on the series page.
How CloudNala can help
We help South African organisations find the decisions in their own operations that are repeated, bounded and costly enough to be worth a decision layer, and the ones that are not. We start from the real channel, the real languages and the real queues, map where authority has to stay, and build an evaluation set from the organisation's own cases before any model is chosen. That way the first pilot tests the idea against the work people actually do.
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.
Book an AI Decision Review or write to us at consult@cloudnala.co.za