Every official language, honestly

Four now. The rest where you trade.

South Africa has twelve official languages — Sign Language became the twelfth in July 2023 and has no written form, so a dropdown cannot serve it. Of the eleven written ones, translating all of them equally is the wrong shape. What matters is not the national ranking but which languages the people in front of this brokerage actually read, and whether an AI draft genuinely saves a sworn translator time or quietly wastes it.

The four in service

English · Afrikaans
isiZulu · isiXhosa

60% of South Africans in their first language — and about 95% in the Western Cape, KwaZulu-Natal and the Eastern Cape, where most bond volume sits. Decided: these four are the service. Everything else is commissioned per province when a brokerage there signs.

THEN →

Add by province, not by national rank

Whatever that
brokerage needs

A Limpopo brokerage gets 1.5% coverage from those four. It runs on Sepedi, Xitsonga and Tshivenḍa — which are exactly the ones the AI is weakest at. That tension is the whole decision, and it is on the By province tab.

The only screen that has to be understood

Document intake, in the four languages Loan Ranger offers. The three after the divider are drafts kept for sales conversations — they are not in service.

Why only the buyer’s screens

The commission ledger, bank packs and analytics are used by consultants and principals who work in English all day, submitting to banks whose systems are English. Translating those is real cost for no comprehension gained. The buyer’s screens are where a person makes a decision with legal consequences, and where “I did not understand what I was agreeing to” becomes a dispute. Translate where comprehension has consequence.

Sorted by whether the AI draft actually helps

Confidence is roughly how much of the draft a sworn translator would keep. The cuts are at 70 and 55.

The threshold is about the translator’s time, not the AI’s pride

A sworn translator signs every language regardless — that never goes away. What changes is what they are handed. Above roughly 70 they review and lightly correct, which is cheap. Between 55 and 70 they rework substantially. Below 55 the draft is worse than a blank page, because they must first read it and then distrust it. For those four, commission a human translation directly and let the AI keep it in sync afterwards — the opposite of what “AI does all eleven” would suggest.

Decided — and here is the only thing left before a buyer sees it

English is the source. That leaves three sworn translations to commission: Afrikaans, isiZulu and isiXhosa. Each translator receives the AI draft, not a blank page, and signs string by string; after that the AI keeps them in step and only re-routes what actually changes. Until those three signatures exist, every non-English screen in this prototype stays exactly that — a prototype.

And the downside is not symmetric. Good isiZulu reaching a quarter of the country is a genuine differentiator. Bad isiNdebele reaching 1.6% is worse than English, because it reads as contempt dressed up as inclusion. That asymmetry is another reason to be narrow and excellent rather than broad and approximate.

National ranking is the wrong instrument

Pick where a brokerage trades. Section 63 of the National Credit Act says the same thing — a language policy should reflect the population ordinarily served.

So the rollout writes itself

Cape Town, Durban and East London brokerages are essentially done on day one. Gauteng wants Sesotho and Sepedi next. Bloemfontein needs Sesotho before anything else and Mahikeng needs Setswana — neither is in the national top four. Limpopo and Mpumalanga run on the four hardest languages and should carry budget for human translation from the start. Commission a language when a brokerage in that province signs, and charge it to that deal rather than carrying eleven for markets you have not entered.

The reason translations rot, and the fix

Every multilingual product dies the same way: English moves and the other ten quietly go stale.

English changes

A consultant rewords one line in the document list. One string, committed like any other.

Eleven drafts, instantly

The AI retranslates that string alone into all eleven and marks the old versions superseded. Nothing else is touched, so nothing else needs re-reviewing.

It scores itself

Each draft carries a confidence. Legal phrasing, money, noun-class agreement and anything it has not seen before score low and get flagged.

A person signs

The sworn translator sees one changed string, not four hundred. They approve or correct it, and their name and date attach to that string in that language.

That is the whole trick

Human translation does not scale to eleven languages across a product that changes weekly, which is why almost nobody does it and why the ones who try end up with ten stale languages and an English one that works. AI makes the marginal cost of a string near zero; the signature is what you are actually buying, and you only buy it for what changed.

Nobody should have to pick from a dropdown

The AI already knows, from the first thing the buyer sends.

Where each language actually stands

Honest state, not a tick list. Share of the population is first-language speakers, Census 2022, approximate.

Not one of these is ready for a real buyer

Everything below English is an AI draft. Afrikaans I would back; isiZulu and isiXhosa are credible; Tshivenḓa, Xitsonga, siSwati and isiNdebele will contain errors a first-language speaker spots immediately. Shipping an AI draft to a buyer and calling it a translation is worse than shipping English, because it looks like you tried and tells them what you think of their language. Budget for a sworn translator per language. The AI makes that cheap and repeatable; it does not make it skippable.

The twelfth official language has no written form

So a language picker cannot deliver it, and claiming otherwise is a false claim.

South African Sign Language is a full natural language with its own grammar. It is not signed English and it is not a writing system. A deaf first-language SASL user reading English text is reading a second language, often one they were taught badly. Turning on an "SASL" option in a dropdown would deliver English text under a label that promises otherwise.

What actually serves a deaf buyer

Short signed video for each step. Eight to ten clips — what we need, why we need it, what happens next, what the offer means, what you are agreeing to. Recorded once with a qualified interpreter, re-recorded only when the process changes.

Never put information in audio alone. No voice-note-only instructions, no calls as the single channel. Everything reachable in text or video.

An interpreter on the consultation, not a family member. Using a relative to interpret a credit agreement is how consent becomes disputable.

Plain-language English as the written companion, written for a second-language reader — short sentences, no legal register where it is avoidable.

What the AI can and cannot do here. It can generate the plain-language scripts, keep them in step with the process, and caption everything. It cannot generate SASL — avatar signing is not accepted by the Deaf community and is not equivalent to a human interpreter. This is the one place in the platform where the answer is a person on camera, and that is worth saying out loud rather than quietly shipping a dropdown entry.

Prototype. Every non-English string here was drafted by AI for this demonstration and none has been reviewed by a sworn translator — they illustrate the mechanism and are not fit to put in front of a buyer. Confidence scores are my own estimate of draft quality, not a measured benchmark. Population shares are approximate first-language speakers, Census 2022, rounded. The National Credit Act position is a plain reading of section 63 and is not legal advice; confirm the filed language policy with your attorney and the NCR. SASL guidance follows generally accepted Deaf-access practice and should be checked with DeafSA before anything is recorded.