This is the 12th instalment of AI Fluency Corner, a 16-part weekly series in Business Day building one connected mental model of artificial intelligence, in plain language. A manager pastes a confidential 40-page report into a free chatbot and asks for a summary and talking points. Ten seconds later it appears, well organised and useful. The deadline is saved, and nothing on the screen suggests anything has gone wrong. But something has. The report has left the company — sent to a server elsewhere, logged and processed under free-account terms that let the provider use what people type to train future models. Last week we found that “online” means real machines in real countries under real contracts. This week: once your work enters an AI system, what happens to it? Among employees who use AI at work, roughly three quarters have pasted company information into it, mostly on unapproved personal accounts. Enterprise AI is not a different technology, just the same models sold under different terms. It comes in four forms: Business versions of familiar chatbots such as ChatGPT Enterprise or Gemini for Workspace. AI built into software you already use, such as Microsoft 365 Copilot in Word, Excel and Teams. Cloud platforms like Azure OpenAI and AWS Bedrock, where a firm runs models against its own data. Models a firm downloads and runs on its own servers. What actually happens to your data Start with the mechanics. A language model does not keep your report the way a filing cabinet keeps a file. Training nudges billions of statistical settings across a vast body of text; one document is a rounding error, not a stored record. So the common worry — that a rival will one day pull your report out of a chatbot — is almost certainly unfounded. And that is far less comforting than it sounds. The real problem is duller and harder to undo. Your document is now a copy on equipment you do not own, under terms you never read. It is kept for a period you did not choose, may be read by the provider’s staff, swept into future training or reached by a court order or a breach. You cannot inspect it or take it back. The danger was never that the model would talk, but that you no longer know where your own information sits. A business account inverts this: what you type is excluded from training by default, data sits in a separate space with retention you set, permissions follow your access rights, interactions are logged and the provider signs an agreement making Protection of Personal Information Act duties contractual. The intelligence is identical; the rules around it are not, and those rules are what you are paying for. The model is rented. The knowledge is owned For three years the argument has been which model is best, while capable models multiplied. Stanford’s 2026 AI index puts the gap between the strongest American and Chinese systems at under 3%; three years ago it was a chasm. Intelligence is becoming abundant and interchangeable. A bank’s lending history cannot be downloaded; a retailer’s customer patterns cannot be bought from a rival. The model supplies intelligence; your organisation supplies relevance. The question is not which model to buy, but how to apply intelligence to your own knowledge without giving it away. Open weights and the sovereignty question Take that fourth form. Models called “open source” are, more accurately, open weight: you may download, adapt and host the finished model, but the data it learnt from stays undisclosed. What changes is direction — with a commercial service you send your data to the model; with open weights you bring the model to your data. This is now practical: Moonshot’s Kimi K2.6 matches leading closed systems on common coding tests at a fraction of the cost, alongside DeepSeek, Qwen and GLM. Two cautions. Calling a foreign provider’s service is not the same as running its open weights yourself: jurisdiction follows the endpoint, not the flag on the model. A model self-hosted in Johannesburg falls under South African law; calling an overseas service is a cross-border transfer. And open is not free: you inherit the hardware, the patching and the skills. Build a stack, then teach it your business One employee does not draft contracts, spot fraud and write software; one model shouldn’t either. Serious setups mix these forms into a stack: a small in-house model sorting documents cheaply, a specialist model reading code, a larger commercial model for hard reasoning once sensitive details are removed, and a routing layer sending each job where it belongs. Beneath it sits your own knowledge, connected in one of two ways. Fine-tuning trains the model on your examples until it adopts your style and terms, changing its behaviour permanently and baking your data into it. Retrieval connects it to your approved documents and looks up the correct passage when asked, so your policies stay in your own store, under your permissions, updated or withdrawn at will. Fine-tune the format; retrieve the facts. Train changing information into a model and you get confident answers from a truth that expired months ago. A year from now, today’s leading model may no longer be leading. Your customer histories, your claims records and your institutional memory will still be yours — unless you have already given them away.Enterprise AI isn’t really a product but a set of deliberate choices about stewardship. The advantage won’t go to whoever finds the cleverest model, but to whoever can change models without ever losing control of what they know. Our task this weekChoose one sensitive business workflow — such as contracts, customer complaints or financial reporting — and redesign how AI should support it. Decide which data must remain internal, which approved enterprise tool may access it, who should have permission, and what must be logged. Then identify one task that could use an external model only after confidential details are removed. That boundary is your first enterprise AI architecture. • Mafinyani is senior partner in financial engineering & artificial intelligence at specialised finance, risk and applied technology firm Intellica Analytics. Next week: AI agents — from assistants to autonomous actors. Business Day