An AI model that reads Zulu and Xhosa can now fit in about 20MB. South African start-up Lelapa AI released InkubaLM in 2024 at about 0.4-billion parameters. Last year Stefan Strydom cut it to 40-million by trimming its vocabulary and sharing internal components.
Now the World Bank has taken up the argument. Its chief knowledge officer, Paschal Donohoe, told the Guardian that developing countries should build this kind of “small AI”: specialised models trained on narrow data, running on phones with poor connectivity and in local languages.
Small models are the first, most visible of three issues now emerging in South African AI. The other two receive far less attention and carry more risk: the memory AI systems accumulate about customers, and the move to paying suppliers only when a task is completed. All three are arriving faster than the contracts, controls and policies meant to govern them.
AI moves onto phone
Size comes down to parameters, the numbers a model learns in training, which decide how much memory and computing power it needs. Quantisation, storing each parameter in 4 bits instead of 16, cuts that memory about fourfold. InkubaLM at full size would then need about 200MB; Strydom’s version is about 20MB.
That changes who can afford AI. In August the World Bank’s “World development report of 2026” noted that almost a third of rural schools in Sub-Saharan Africa have no electricity and more than two-thirds lack a consistent internet connection. Chief economist Indermit Gill said developing economies “do not need large models or big data centres” to benefit.
It also changes procurement. Classifying a query, extracting fields from a form or flagging an incomplete application rarely needs a frontier model billed in dollars. Before renewing an AI contract ask for three numbers measured on your own data: accuracy, cost per task and performance with the network off.
Small models fail confidently outside their narrow task and so do large ones without checks. The state learnt this in April when it withdrew its draft national AI policy after at least six of its 67 references were found to be fictitious, apparently AI-generated and unchecked.
Models invent sources when they lack the right material, and the usual fix ― giving AI access to an organisation’s own records ― creates the second issue.
AI memory outlives the data
That fix is called retrieval. The system searches an organisation’s records, places the most relevant passages into the model’s context window and answers from them. To search by meaning each passage is converted into an embedding ― a long string of numbers stored in a vector database.
That is where the risk begins. Deleting a client does not delete the embeddings, cached passages and AI-written summaries derived from their file. Section 14 of the Protection of Private Information Act says personal information should not be kept longer than its purpose requires, so every copy left behind is exposure. Section 72 limits sending personal information outside South Africa, which bites when the vector database sits on an offshore server.
The same copies set a second trap when an organisation changes supplier. Embeddings work only with the model that produced them; a competitor’s system generally cannot read them. An organisation that leaves its supplier with an export of vectors has, in practice, left its memory behind.
Both traps are best closed in the contract. Require the source text, metadata, conversation logs and AI-generated notes in an open format, a register of where each copy is stored and written confirmation that derived copies are deleted when the source is. Whoever holds the memory also gains leverage over the third issue: deciding when a job counts as done.
Pay only for results
Suppliers pay for computing on every conversation, successful or not. Outcome pricing moves that cost of failure to the supplier, and prices are falling fast. Intercom’s Fin charges $0.99 per resolution. In April HubSpot moved its customer agent from $1 per conversation to 50c per resolved conversation. At its stated 65% resolution rate the old price worked out at about $1.54 per resolution, so the effective price fell by two-thirds overnight.
Falling prices put all the weight on one word. Does “resolved” mean the customer confirmed it, or only that they stopped replying before the chat timed out? Contracts should state the test, the time window, who audits a sample and who pays when an automated answer causes harm.
They land hardest on South Africa’s business process outsourcing sector, which has created about 180,000 jobs in 10 years, about 90% filled by young people. Research by Caribou and Genesis Analytics found that more than 60% of South African business process outsourcing workers already use chatbots or copilots and that 40% of sector tasks could be automated by 2030. With youth unemployment at 47.4%, a query priced at 50 US cents now competes with a school-leaver’s first job.
What does it all mean?
Smaller applications, retained context and outcome pricing affect different parts of the same calculation: the cost of delivering a service, the information needed to deliver it accurately, and the basis on which suppliers earn revenue. Lower computing costs do not remove the expense of maintaining records, handling exceptions or correcting mistakes.
These developments connect AI adoption to existing concerns about service delivery, personal information and employment. A recorded resolution, a supplier’s payment and a customer’s experience remain three different measures of success. The differences between them explain why cheaper AI does not automatically mean a cheaper or better service.
• Mafinyani is senior partner in financial engineering & artificial intelligence at specialised finance, risk and applied technology firm Intellica Analytics.
Business Day
Would you like to comment on this article?
Sign up (it's quick and free) or sign in now.
Please read our Comment Policy before commenting.