This is the sixth instalment of AI Fluency Corner, a 16-part weekly series in Business Day building one connected mental model of artificial intelligence in plain language.You have been talking to machines your entire working life. You just did not know they were listening.Every complaint you phrased carefully and sent to a call centre trained a classifier to read tone. Every contract clause your legal team flagged as unusual taught a model what to surface next time. Every voice note you dictated became training data the moment it was transcribed. None of this required your awareness. It happened because language — your language — is the most abundant, most valuable and most underestimated data your organisation produces. For the past decade, a branch of artificial intelligence called natural language processing has been quietly learning how to use it.This week we move from neural networks — the architecture that lets machines see, hear and transcribe — to what they are predominantly deployed for: working with words.The problem that took 50 years to crackFor most of computing history, language broke machines. Not because machines were unintelligent, but because language is not logical. It is contextual, ambiguous, ironic and culturally loaded in ways that rules cannot capture.Consider four words: “I see your point.” They can mean agreement, polite resistance or outright sarcasm, depending on what preceded them, who said them and in what register. Early natural language processing (NLP) systems responded by looking for keywords. Customers learnt to phrase complaints without trigger vocabulary. Contracts buried risk in clauses that used none of the obvious terms. The systems responded to the surface. Language operates beneath it.What changed was a single insight: meaning does not live in words. It lives in the relationships between them. A system trained on billions of sentences learns “I have now been waiting” in a service e-mail is a dissatisfaction signal even when the message is polite. That “terminate without cause” is structurally different from “terminate after breach”. That “not bad” is rarely negative.This is the first lightbulb: NLP does not understand words. It has learnt the statistical neighbourhood of words. That turns out to be enough to do remarkable things.Where it is already workingYour inbox is already being read before you open it. Not by a person. By systems deciding, in milliseconds, which messages deserve attention, which are promotional and which is the phishing attempt dressed as an invoice.The spellchecker that silently fixed your e-mail — NLP. The autocomplete that finished your sentence — NLP. The chatbot that handled your 11pm insurance query — NLP. The meeting transcript with action items sorted by owner, delivered 10 minutes after the call ended — NLP. The CV-screening tool that ranked candidates before a recruiter opened a single application — NLP.NLP does not read. It predicts ... A system that predicts well enough, on familiar enough material, becomes functionally indistinguishable from one that actually reads. That is the magic. That is also the trap.Four tasks do most of the practical work: Sentiment analysis flags the customer whose language has shifted from patient to frustrated before the account escalates. Text classification sorts documents — claims by type, complaints by urgency, contracts by risk profile — without anyone reading each one. Named entity recognition extracts critical details buried in free text: the renewal date in clause 14.3, the penalty threshold on page six. Language generation produces the first draft — the summary, the suggested reply and the meeting notes your team once produced manually.These are not experimental. They are running today in tools most organisations already pay for. The question is not whether you are using NLP. It is whether you know what assumptions are baked into what you are using.The distinction that changes everythingHere is the second lightbulb, and it is the one most vendor pitches deliberately bury: NLP does not read. It predicts.Reading involves understanding, inference and the ability to notice what is missing. Prediction involves pattern matching — identifying what typically follows what, based on everything the system was trained on. A system that predicts well enough, on familiar enough material, becomes functionally indistinguishable from one that actually reads. That is the magic. That is also the trap.A language model summarising a supplier contract does not verify facts against your procurement policy. It predicts what a plausible summary looks like, based on summaries it has seen before. If the contract contains an unusual indemnity structure, the model may smooth over it — unusual structures do not appear frequently enough in training to register as significant. The summary sounds professional. The clause was missed.Fluent language can disguise weak reasoning. A system that writes beautifully is not necessarily thinking carefully. Learning to tell the difference is not a technical skill. It is the whole jobThis failure mode has a name: hallucination. In NLP, it means confident, fluent output that is factually wrong. The danger is not that it sounds wrong. It is that it sounds entirely right.South Africa adds a specific structural risk. The models underlying most commercial NLP tools were trained predominantly on American and European English. South African business communication is multilingual, code-switched and contextually specific in ways those models were not built to handle. A sentiment tool calibrated on US social media may misread the register of a formal Zulu-inflected complaint. The question that cuts through every pitchWhen a vendor presents an NLP solution, one question reveals more than the entire demonstration: What language was this trained on, and how does it perform on material that looks like ours?Performance data on inputs that match your documents, your registers, your languages. If that data does not exist, the product has not been validated for your environment. The confidence in the pitch is borrowed from someone else.Fluent language can disguise weak reasoning. A system that writes beautifully is not necessarily thinking carefully. Learning to tell the difference is not a technical skill. It is the whole job.One task this weekIdentify one text-based process in your organisation where people repeatedly read, sort, summarise or respond to language — complaint e-mails, supplier contracts, CV screening, board pack summaries, or call transcripts. Then ask four questions: What text comes in? What decision follows? What must never be missed? Who checks the output before it is acted on? The answers will tell you exactly where NLP can help and precisely where it still needs a human in the room.Next week: Large language models — how tools such as ChatGPT, Gemini and Copilot predict language, why they sometimes get it wrong, and how to use them with informed scepticism rather than blind trust.• Mafinyani is senior partner in financial engineering & AI at specialised finance, risk and applied technology firm Intellica Analytics. Business Day