This is the ninth instalment of AI Fluency Corner, a 16-part weekly series in Business Day building one connected mental model of AI, in plain language.For most of the digital age, the working professional has been, quietly, a translator. A meeting becomes minutes, a spreadsheet becomes a paragraph, the paragraph becomes slides, the slides become an email. The insight barely moves; only the container changes. Yet container-shifting has swallowed a startling share of the week. Multimodal AI matters because it attacks that waste at the source. It is not here merely to write cleaner emails or decorate slides. Its deeper job is to collapse the walls between the formats organisations have treated as separate worlds. That is the least glamorous description of the technology, and the most consequential.The end of information living in separate countries (Rufaro Mafinyani) Start with the definition. A “modality” is a form of information: text, image, audio, video or structured data. Multimodal AI is one system that takes these in combination — a photo, a voice note, a scanned contract, a spreadsheet — and reasons across them as one context. Earlier tools each owned a narrow lane: natural language processing (NLP) for text, optical character recognition (OCR) for scanned pages, speech recognition for audio, and computer vision for images. We have all seen the older version: a bank agent places your ID document on a scanner and the system reads your name, number and address into a form; your phone predicts the next word in a message. Useful, yes — but narrow. AI-enhanced multimodality does not merely grab a field or complete a sentence. It connects the ID, form, email, voice note and decision that must follow. The breakthrough is not that a machine can read a page or recognise a picture. It is that one system can understand the relationship between the picture, paragraph and recording. The complaint email, the angry voice note and the photo of the damaged product are no longer three files. They are one problem.This is no fringe capability. In 2024, GPT-4o, Gemini and Claude all became natively multimodal — vision and language trained together, not bolted on. Gartner projects that 40% of generative-AI solutions will be multimodal by 2027, up from 1% in 2023. That is not a feature release. It is a shift in what AI is becoming.The breakthrough is not that AI can read, see or listen. It is that it can connect the evidence across all threeIt is already inside the tools on your deskHere is the shift many professionals underplay: the model itself is becoming less of the advantage. Most leading AI systems are trained on the same broad public material — largely the internet — so their baseline capability is starting to look more like a commodity. Multimodal AI is embedded in the tools you use – bringing context together and reducing handovers (Rufaro Mafinyani) The real value comes from context: the email thread, the spreadsheet, the meeting transcript, the policy, the customer record, the slide deck. That is why connectors into Microsoft 365, Google Workspace, Slack, Teams and other everyday systems matter. They turn AI from a clever outside assistant into something more useful: a system that can work with the evidence of the business, not just the knowledge of the web.Why the value is not where you are lookingThis explains a common disappointment: why organisations invest in AI and struggle to see the return. They look for value in the visible output — the image, the summary, the slide — while the real gain sits in the friction between stages. A meeting becomes a transcript, the transcript becomes a summary, the summary becomes a proposal, the proposal becomes a decision. Call this the translation tax: the hidden time spent converting one format into another before anyone can act. Multimodal AI attacks that tax directly. When spreadsheets arrived, arithmetic mattered less than interpretation. When search arrived, recall mattered less than judgment. Now, as formats collapse, judgment becomes even more valuable.Context is the advantage (Rufaro Mafinyani) For South African businesses, this is where the technology becomes real. Banking, insurance, legal services, logistics and public administration still depend on paper, scanned forms, photographs, call recordings and system extracts. Optical character recognition could already lift text from a document; transcription could already turn speech into words. The step change is that AI can now connect those pieces: read the handwritten delivery note, understand the bilingual customer call, compare both to the system record and flag the mismatch. The bottleneck was never only paper. It was fragmented context. Multimodal AI starts to close that gap.Seeing is not understandingThat is also the catch. The richer the material a system handles, the more confidently it can be wrong. A model describing an image is not seeing as you see; it is predicting what the image most likely contains. OCR can misread a smudged number; a summary can skip the clause just outside the frame; a meeting recap can give certainty to what was only a maybe. The visual and audio versions of hallucination are fluent — and harder to spot, because a polished deck feels like proof.Where the output touches money, law, customers or reputation, a human must verify the source, logic and consequence. A multimodal model can read the room; it cannot carry what happens next. Under the Protection of Personal Information Act, neither can “the system told me so”.So resist the tempting misreading. Multimodal AI is not mainly about machines becoming creative. It is about machines becoming contextual. The creativity is a by-product; the context is the innovation. For decades, software was organised around file types: Word for documents, Excel for numbers, PowerPoint for slides. The new systems are organised around meaning. The professionals who pull ahead will not be those who can recite every model name. They will look at a messy, mixed-format pile of work and ask: what can AI perceive here, what can it produce and what must remain human judgment?ONE TASK THIS WEEKTake one task you completed this week and trace its formats: meeting, email, spreadsheet, slide, screenshot or voice note. Ask: where was judgment added, and where was information merely moved around? Then give one mixed-format item to Copilot, Gemini, Claude or Gamma: “Use only this material. Draft the output. Flag uncertainty. Invent nothing.” That gap is your translation tax.Next week: What is an application programming interface (API) – the invisible connector that lets these tools plug into the systems you already run, and what that means for control, cost and being locked in to a vendor.