Most of what gets written about artificial intelligence sits somewhere between fear and hype, neither much use to someone deciding, on an ordinary Tuesday, whether to actually use the thing. The plainer truth: it is a good tool, and by now an unavoidable one, already part of how serious work gets done here, the way spreadsheets and email did decades ago. The question that matters now is how well to use it, and that depends on one thing more than anything else: what the person typing the question already knows.
We’ve been watching banking technology land in this market since the years when core systems, cards and ATMs were the new thing and most people hadn’t yet worked out what to make of them. The pattern with AI is not new here, just faster.
What the tool actually is
Strip away the mystique and a language model is doing something fairly ordinary. It has read an enormous amount of what people have written and predicts, word by word, what plausibly comes next. Ask it something general and the answer comes back general: fluent and confident, built from the average of everything the model has seen on the subject, right about as often as an average is right, and wrong in a way that reads just as confidently as correct. It has no idea what sits in a particular bank’s loan book, or why the head of the branch network is worried about one covenant this quarter. Whatever it is going to understand about the situation, somebody has to tell it.
Context is the condition
This is why the same tool produces such different results for two people asking about the same subject. A vague prompt gets a vague, generic answer, because there was nothing specific for the model to work with. Give it the real constraints instead: the numbers, the thing that matters this time and the thing that doesn’t, and the same model returns something genuinely useful, sometimes better than a junior colleague would produce in ten minutes. Swapping to a newer or supposedly smarter model changes surprisingly little; a sharper description of the problem itself changes almost everything. That’s a lesson we keep relearning with every new piece of banking technology to arrive here.
Where it does the most
Put those facts together and you can predict, fairly reliably, who gets the most out of it. A credit officer who has read covenants for fifteen years gives the model rich context almost without noticing: which clause matters most, what the borrower will argue, what “market standard” means for this size of deal. They can tell inside a couple of seconds when an answer is off, because they already know what a right one looks like. For that person AI is a genuine accelerator, clearing the drafting and the summarising and leaving more of the day for the judgment call only they can make. We see this pattern across finance and risk teams here: adoption runs highest where someone already knows the subject cold, and thins out the further a person drifts from their own ground, because the expertise is what supplies the context.
The one honest limit
Which brings us to the limit worth being honest about: it’s tempting to hope the tool covers the areas where confidence runs thin, the unfamiliar regulation, the skill never quite built. It mostly won’t, for the same reason it works so well on familiar ground. Good context can only come from someone who understands the ground, and without that, questions come out vague and a wrong answer looks as polished as a right one. A branch operations manager asked to draft a cross-border security opinion can’t tell a plausible invention from a correct citation, because the thing that would let them tell the difference is the expertise they were hoping to borrow. Closing that gap takes a person who already knows the subject checking the work, or the model reaching someone who does before a decision gets made on the strength of it.
AI multiplies what’s already in the room: where there’s real knowledge, it gives back a great deal, faster than most people expect; where there isn’t, it produces fluent text with not much underneath, and in banking the difference between those two is usually the whole job.