This is the 15th instalment of AI Fluency Corner, a 16-part weekly series in Business Day building one connected mental model of artificial intelligence (AI), in plain language. On Monday Standard Bank named Amazon Bedrock as its enterprise AI platform. Some 39,000 employees — 72% of the group — actively use generative AI, and about a third of its technology staff use AI coding tools, with claimed productivity gains near 20%. It has not said which models it runs, or how the 20% was measured. The week before, a TechCentral survey of just over 100 local technology decisionmakers found only about a quarter of their organisations running AI in production. The largest group, 43%, was still piloting: working in a test, not yet trusted with real work. The gap is not computing power; only 7% said their infrastructure needed significant work. It is a map. Standard Bank’s chief operating officer, Margaret Nienaber, put it plainly: access to the newest models will not decide the winners, because most organisations will have the same access. This edition is that map. The landscape has four layers, and most shopping happens on two. In July Pick n Pay put Penny into its asap! app: a Gemini-powered assistant that builds a grocery basket from a typed request, a voice note or a photograph of a handwritten list. To a shopper it is one tool. Underneath, four. The first layer is the model: Claude, GPT, Gemini and the open-weight families from Meta, Alibaba and Moonshot AI. Capability is converging, prices keep falling, and Standard Bank’s answer is to run several and swap as they change. The second is the assistant: that intelligence wired to your things. Copilot inside Microsoft 365, Gemini inside Google Workspace, ChatGPT and Claude in a browser tab. What separates them is not cleverness but reach: a modest model that can see the right spreadsheet beats a brilliant one facing an empty box. The third layer is the one everyone forgets: the AI already inside software you pay for monthly — Sage, Xero, Salesforce, Workday, your contact-centre platform. Licensed, invoiced and, in most organisations, never switched on. Standard Bank’s cloud costs rose 37% while total technology costs rose 2%, partly absorbed by rationalising licences. Someone counted what was paid for and not used. The fourth is plumbing, and it keeps this year’s choice from becoming next year’s cage. In December 2025 the Model Context Protocol — one standard for connecting any AI tool to any system — went to the Linux Foundation, with OpenAI, Microsoft, Google and Amazon signed on. In March Alibaba released OpenSandbox under an Apache 2.0 licence: a free, self-hostable room where software runs code without being handed the machine. Penny sits here too, reaching Pick n Pay’s catalogue and loyalty data through restricted interfaces, never the systems themselves. Where your data lives is now a configuration, not a refusal. The market renames itself constantly; Google renamed NotebookLM to Gemini Notebook in July, the product’s third name in three years. Work does not churn like that. So ask not what a tool is called, but what it does to information. There are four answers. Extract turns long into short and messy into structured: contract clauses in legal, invoices in finance, claim documents in insurance, minutes anywhere. Copilot and Gemini do the everyday version; your accounting platform probably already does invoices. It is the cheapest return on the map, because checking takes a minute Draft turns a brief into a first version: proposals, job specifications, campaign copy. ChatGPT, Claude and Copilot all do it, Canva the visual equivalent. The value is the blank page, never the final page. A tool saving 40 minutes and creating 20 minutes of correction has won. Retrieve turns scattered knowledge into an answer: internal helpdesks, policy lookup, product questions in a call centre. Gemini Notebook and Perplexity work against sources you choose; Glean, Zendesk and Intercom against your own systems. It is the hardest to do well, because your information must be connected, current and correct — and where the largest gains sit, because it removes the searching, not the typing. Act performs the step instead of describing it: GitHub Copilot, Claude Code and Cursor in software teams; Power Automate, Zapier or n8n across systems; Agentforce inside Salesforce. Capitec runs agentic AI in business banking credit processing and, in the same report, raised model risk to tier one. Edition 13’s arithmetic explains that caution: a 10-step process 95% right at each step is only 60% right end to end. Three questions that shrink any shortlist:Is this already inside something I pay for? Answer that before the demonstration, not after. Much AI spending here is a second licence for a capability lying dormant in the first. What must it reach, and through what? A tool that cannot see your data is a toy. A tool that can see everything is a liability. The answer should name a connection, a permission and an owner. What does verification cost? If checking the output takes as long as producing it, you have bought a hobby. The measure that survives a finance committee is cycle time and error rate on one named process. For most people the honest shortlist is three tools, not 30: the assistant embedded in software your organisation already runs, one general model you have learnt to prompt properly, and one specialist tool for whatever eats most hours of your week. The right model will not be the differentiator. The best stack is not the one carrying the most logos; it is the smallest one that reliably produces useful work without creating invisible risk. Our task this week Pick one task that eats at least 30 minutes of your week. Name its verb — extract, draft, retrieve or act — and check whether software you already pay for performs it. Then test no more than two tools on the same redacted example, scoring accuracy, time saved and rework. Keep the winner only if it improves the finished work, not merely the first draft. • Mafinyani is senior partner in financial engineering & artificial intelligence at specialised finance, risk and applied technology firm Intellica Analytics.Next week, the final instalment: a practical framework for planning and designing AI initiatives around real business needs — and measuring their return on investment. Business Day