This is the third entry in AI Fluency Corner — a 16-part weekly series building one connected mental model of AI in plain language. You open a maps app. Two routes appear. One is shorter. The app pushes you toward the longer one — traffic is already building near the offramp. You do not see the calculation. You see the instruction: take this road. A fraud analyst sees another: hold this transaction. A recruiter sees another: review this candidate before that one. A loan system sees another: decline. Last week we looked at data — the recorded evidence AI is allowed to learn from. This week comes the next layer. Once the data exists, something must decide what to do with it. That something is the algorithm. What an algorithm actually is An algorithm is a step-by-step method for turning an input into an output. A recipe is an algorithm: ingredients in, meal out. A loan affordability calculation is an algorithm: income, expenses, interest rate and term in; repayment capacity out. A call centre routing rule is an algorithm: complaint type, language, customer value and agent availability in; queue position out. AI did not invent algorithms. It made them more powerful, less visible and harder to interrogate. In ordinary software the steps are written by people: if an invoice is 30 days overdue, flag it; if stock falls below the threshold, reorder. Every foreseeable case coded in advance. Reliable for known scenarios. Brittle when the world changes. AI works differently. Instead of a human writing every rule, the system is shown many examples — the data we discussed last week — and figures out its own internal rules. The instructions emerge from the data rather than being hand-crafted. This is why a navigation app improves without a developer rewriting code every week. Traditional software: a human writes the full instructions first, then the computer executes. AI: the system discovers useful instructions from examples, then applies them to new situations. That reversal is worth holding on to. Three algorithms already deciding things about you The sorting algorithm decides what you see first. Gmail’s priority inbox was not programmed with a list of important people. It learned from your behaviour — what you open quickly, ignore or reply to. Importance is not a property of the email. It is an output of the algorithm that studied you. The scoring algorithm decides your access. At Absa, Nedbank and Standard Bank, loan applications are processed by models weighing payment history, utilisation, account age and dozens of other variables. You never negotiate that weighting. The score is not a judgment. It is a calculation — constrained by the data it was fed and the priorities encoded into the model. The recommendation algorithm decides what you encounter next. YouTube, Spotify and TikTok identify people who behave like you and surface what they engaged with. Your taste is approximated by your history, filtered through everyone else’s. What the result cannot do is introduce you to something genuinely outside your pattern. That limit is worth knowing. Algorithms carry priorities Every algorithm optimises for something. A maps app may optimise for the fastest route, fewest tolls or least fuel. A bank may optimise for fraud prevention, convenience, compliance or loss reduction. Those are not technical details. They are business choices — made before you arrived. If a fraud algorithm is tuned too aggressively, it protects the bank while frustrating legitimate customers. If a recruitment algorithm optimises for people who resemble past successful hires, it may quietly repeat the old definition of promising. The algorithm will not pause to ask whether the objective was fair, complete or current. It pursues the target it was given — or inferred from history. Algorithms do not merely calculate. They express priorities. When to question the logic Three situations deserve scrutiny: When the stakes are high. Algorithms informing credit decisions, recruitment shortlisting, medical triage or legal risk are not value-neutral. Before accepting the output, ask what the system was optimising for — and whose outcomes were represented in the training data. When the population is different. A model trained on salaried urban customers may misbehave when applied to gig workers or informal earners. Fraud models built on US transaction data behave differently in Joburg. This is a limit to check before deployment, not a flaw in the concept. When the explanation is absent. Many modern algorithms cannot produce a plain-language account of an individual decision. If a loan is declined or a profile downranked, the system may not say why in terms a human can verify. Where decisions must be explainable — by regulation, by fairness or by audit requirements — the choice of algorithm matters as much as the choice of data. What this means for your work Algorithms are embedded in more business processes than most managers realise: invoice approvals, staff rostering, customer segmentation, demand forecasting and compliance monitoring. Most run invisibly until something goes wrong. Fluency does not require understanding the maths. It requires four questions before trusting any algorithmic output: what went in; what was being optimised; what happens when it is wrong; and who is accountable? An algorithm that misprices a product loses margin. One that misclassifies a customer loses trust. One that makes an unsupported lending decision may violate the National Credit Act. The consequences are not technical. They are financial, reputational and legal. Automated confidence is not the same as accuracy. Our task this week Choose one automated decision you encounter this week — a route, a score, a recommendation, or an approval. Ask: what went in, what objective was it serving, what happens when it is wrong, and how would you challenge the result? • Mafinyani is a senior partner in financial engineering & artificial intelligence at the specialised finance, risk and applied technology firm Intellica Analytics. Next week: machine learning — how systems learn from data. If algorithms are the instructions, machine learning is the process by which AI writes those instructions for itself.