AI literacy is the new digital literacy

Twenty years ago you could run a branch without touching a computer. Someone else did the typing. That stopped being true so gradually that most institutions never made a decision about it — they simply woke up one day unable to hire anyone who couldn't use a spreadsheet.

The same thing is happening now with AI, and it is happening faster. The gap is not between organisations that have bought AI software and organisations that haven't. It is between teams who know what these tools are good at and teams who are guessing.

We see the guessing constantly. A credit officer pastes a customer's full national ID and bank details into a public chatbot to summarise a file. A manager rejects a genuinely useful automation because a demo hallucinated once in 2024. A team spends four months building a model to solve something a database query answers exactly. All three are literacy failures, not technology failures, and no procurement decision fixes any of them.

Working fluency is a lower bar than people think. It is not prompt engineering and it is certainly not machine learning theory. It is four things.

First, knowing what the tool actually does. A language model predicts plausible text. That makes it excellent at drafting, summarising, reformatting, explaining and translating — and structurally unreliable for arithmetic, for citations, and for any fact it wasn't given. Someone who understands that one sentence will make better decisions than someone who has watched ten hours of vendor webinars.

Second, knowing what must never go in. Customer names, account numbers, national IDs, anything covered by your data protection obligations. This needs to be a written rule with a named approved tool, not a vague sense that people should be careful. Vagueness is how the credit officer ended up pasting the ID.

Third, knowing that output is a draft. Every AI-produced artefact that touches a customer, a regulator or a ledger gets a human check. Not because the model is usually wrong, but because when it is wrong it is wrong confidently and in the same tone as when it is right. That is precisely what makes an unchecked draft dangerous.

Fourth, knowing when not to reach for it. The most expensive AI projects we have been asked to rescue were ones where a deterministic rule would have worked, been auditable, and cost a fraction as much. Judgment problems want AI. Plumbing problems want plumbing.

You can teach all four in an afternoon. What you cannot do in an afternoon is build the habit, and this is where most training programmes fail. A single workshop produces enthusiasm that decays in three weeks. What works better is smaller and more boring: pick one real, low-risk task per team — drafting customer letters, summarising branch reports, translating notices into local languages — and have people do it with the tool for a month, in the open, with mistakes discussed rather than hidden.

The mistakes matter more than the successes. A team that has watched a model confidently invent a policy clause never needs to be told again to verify output. A team that has only seen the demos does.

There is an uncomfortable equity dimension too. In most institutions the people with the least access to this kind of learning are the ones closest to the customer — tellers, agents, field officers — and they are exactly the people whose work AI will reshape first. If literacy only reaches head office, you get executives with opinions about AI and frontline staff who cannot use it, which is the worst of both.

Start at the counter, not the boardroom. Give people one useful thing, let them break it safely, and be honest about the limits. That is the whole programme.

Digital literacy was never really about computers. It was about not being locked out of your own work. This is the same problem, arriving again, on a shorter timeline.