Banks are handing lending decisions to machines and calling it progress. The machines are genuinely good at it. But lending was never a problem with a right answer waiting to be calculated, and treating it as one quietly breaks something important. This is the long version of why.
Computer says no. Most of us made that joke, more than once. A bank clerk, a screen, and a verdict nobody could explain. Little Britain first aired the sketch in 2004, and it was funny because it was absurd. No real institution, surely, would hand a person a decision that mattered and then be unable to say why the answer was no.
That has stopped being absurd. Banks are now building, at scale and with some pride, exactly the thing the sketch was mocking. And they are calling it progress.
I want to argue that the pride is misplaced. Not because machines are bad at lending decisions. In many ways they are remarkably good at them. The problem is subtler, and it is about what a lending decision actually is. We have started treating it as a science, a sum with a correct answer that a clever enough model will find. It has never been that. It is a craft: a judgement made under real uncertainty, deliberately wrong a small share of the time, on purpose. Miss that, and everything clever you build on top sits on a crack.
First, how a bank actually makes its money
It is worth being plain about this, because the whole argument rests on it, and the usual way of describing it hides the point.
A bank pays savers a rate of interest for the use of their money. It charges borrowers a higher rate to use that same money. The difference between the two is called the margin, and living on that difference is, in essence, the whole business. But here is the part people skip: the margin is not profit. Not yet, and not by a long way. Out of that margin the bank has to cover its running costs, hold expensive safety capital that regulators require, and, crucially, absorb the loans that never get paid back. Only what survives all of that is profit.
This changes how you should think about a lender that never makes a bad loan. If a bank lends only to people who are certain to repay, it will lend to almost no one, because very few of us are certain. Lend too cautiously and the losses do disappear, but so does the volume of lending, and the margin shrinks until the business starves. A perfectly clean loan book is not a sign of prudence. It is a sign of a bank that has stopped doing the thing banks exist to do.
So a certain amount of bad debt is not a failure. It is a designed cost of doing business, the price of reaching the customers who make the margin worth having. The skill of lending is not avoiding losses. It is calibrating them: finding the level of expected default that leaves the most profit once everything the margin must cover has been paid. That number is a judgement about an uncertain future, and it moves as the world moves. You cannot simply compute it. That is what makes lending a craft.
Hold on to that idea, because almost everything that goes wrong when banks automate lending goes wrong by forgetting it.
Why more data pushes a machine toward “no”
Now point a modern model at the problem, and watch what it does when you feed it more and more information.
You might expect that more data makes a lender more generous. With a richer picture of someone, surely the machine can find reasons to say yes that a cruder system would have missed. Sometimes it does. But the stronger pull runs the other way. Give a model enough detail about a person and it does not, on the whole, get better at approving them. It gets better at finding reasons to decline. Look closely enough at anyone and there is always something: a thin patch of credit history, an irregular income, a faint resemblance to a group that defaulted years ago. The more data you pour in, the more of these reasons surface, and the more the system drifts toward refusal.
Left unchecked, you end up with what is only half-jokingly called “analysis paralysis”: a system that, knowing everything, says no to almost everyone, and cannot give a reason a human would actually recognise as one. The UK regulator has seen the shape of this from a related angle. When the Financial Conduct Authority tested how people cope with credit information in 2026, it found that simply giving them more did not reliably help them choose better. People fell back on a rough rule of thumb, and applied it even where it did not hold. The lesson was not that people are careless. It was that piling on information is no substitute for designing it well. That is a point about customer leaflets, but it rhymes with a deeper truth: past a point, more data clouds judgement rather than sharpening it. The urge to resolve doubt by adding information is exactly the urge that, in lending, quietly manufactures rejection.
What underwriting actually is
Let me be plain about the thing itself, because the whole argument turns on it. (“Underwriting” is simply the trade’s word for the act of deciding whether, and on what terms, to lend to someone.)
It is not a science, and it never was. It is a craft: a blend of science, judgement and hard-won experience. The science does the easy part. It clears the obviously sound applicants and screens out the obviously unsound, faster and more consistently than any human panel could. But there is always a remainder the science cannot settle, and reading that remainder is the art. It is the ability to weigh a life that does not fit the template, to tell the difference between someone who is genuinely risky and someone the standard scoring simply cannot see clearly.
That is why a pure science cannot lend. Lending is not a question with a right answer hidden in the data. It is a judgement about an uncertain future, one that will be wrong a calculated share of the time on purpose, and whose quality is measured not by whether any single decision was right but by whether the book as a whole was well calibrated. You do not calculate that. You learn it, the way any craft is learned: by doing it, getting it wrong, and being answerable for the outcome.
None of this means turn the machines off
It would be easy to hear all this as nostalgia, a plea to keep humans in the room for their own sake. It is not.
The machine should own the obvious. The clear approvals and the clear rejections, the two ends of the range where the science genuinely does settle the matter, are exactly where automation belongs, and it handles them at a speed and volume no human team could match. Insisting a person personally review every one of those would be indulgent, not careful. Most decisions are of this kind, and giving them to the machine is simply the right use of it.
The danger has never been at the easy ends. It is in the middle: the borderline case the rules cannot settle, the applicant the scoring cannot read, the decision where the craft actually lives. And here the industry has been making a quiet, costly mistake. The instinct, when a case is borderline, is to send it back to a human to judge. That instinct feels responsible. It does not scale, and it never will, because you cannot hand-review borderline cases one at a time at the volume a modern bank runs. There are too many, arriving too fast.
So the craft cannot stay where we have always pictured it, in the head of an underwriter deciding one file at a time. If it is to survive at scale, it has to move. It has to be built into the design of the system itself: the rules, the thresholds, the definition of what a good outcome looks like, the points where a case is escalated. The judgement does not vanish. It relocates, from the moment of the decision to the design of the thing that makes the decisions. From the desk to the design.
A real example, and the line that matters
It helps to make this concrete, because it is easy to say automation is fine “with the right safeguards” and never say where the safeguard actually sits.
Take a company like Zest AI, whose technology some US credit unions use to automate a large share of their lending decisions. At one such credit union, by the company’s own account, between 70 and 83 per cent of consumer loan decisions are made automatically: the model assesses the application, reaches a yes or no, and no human underwriter is involved in those cases.
It is worth pausing on what that figure does and does not prove, because the setting flatters it. A US credit union is a members’ organisation, not a bank open to the public. Its borrowers are members, usually people who have saved with it first, often for years, and who have therefore already shown they can manage money before they ever ask to borrow. So a great deal of the hard judgement has already been done, long before the model sees anything, by the simple fact of who is allowed to be a member. Automating four in five decisions from a pool that has already been filtered for reliability is a far less impressive feat than automating four in five from the general public. The craft has not been removed here. It has moved upstream: deciding who may join, and who may borrow, is itself an act of lending judgement, made by people, in the design of the institution. The model is automating what is left after that judgement has already done most of the work.
Now, the reflex, once you take the craft seriously, is to recoil at the automation: no human saw the decision, so surely it has been abandoned. That reflex is wrong, and seeing why is the heart of the whole argument. Automating the middle is not the sin. A well-built system may make fairer, steadier borderline decisions than a tired human at four on a Friday. The sin is not that no human saw the decision. The sin is if no one can reconstruct it.
That is the line that matters. The question is never simply whether a person laid eyes on a given decision. It is whether the bank can later rebuild the account of it, the information it rested on, the reasoning it followed, the policy that applied at the time, and stand behind that account when it is challenged, months or years later. If it can, the automation is legitimate, human eyes or not. If it cannot, the decision has not been automated. It has been abandoned, and the rejection letter is the only evidence that anyone ever chose at all.
And there is a subtlety here that is easy to miss and expensive to get wrong. When a decision is challenged, through a complaint, an ombudsman, or a court, the question is about a particular decision made on a particular day. It is not enough to show what the system would decide today. The model will have been retrained since, the policy updated, the data moved on. The accountable person has to be able to reconstruct the system exactly as it stood at the moment that decision was made: that version of the model, that version of the policy, the information as it was on the day. Explaining the case against today’s machine answers a question nobody asked. The honest account is always the one frozen at the point of the decision, and building a system that can reproduce that state, on demand, long after the fact, is a far harder thing than building one that simply decides.
On the hook, not in the loop
Which brings us to the point the industry finds least comfortable.
If the craft has moved into the design, then responsibility has to move with it. The comfortable answer, the one that satisfies a committee, is to put a human “in the loop”: a person sitting beside the machine, reviewing its decisions. But we have already seen that this does not scale. Someone nominally reviewing thousands of decisions a second is reviewing none of them in any real sense. That kind of presence is theatre. It produces the appearance of oversight and the substance of a rubber stamp with good attendance.
The honest position is harder. It is not a human in the loop, checking what cannot be checked at that volume, but a human on the hook: a named individual answerable for how the system is designed, able to explain why it lends and why it declines, and accountable when it gets things wrong. Not answerable for each decision, which is impossible, but answerable for the thing that makes the decisions, and for whether it can give an account of itself. The craft moves from the desk to the design. And the person moves with it, from making the call to owning the machine that makes it.
This is not a vague, feel-good kind of accountability. In UK financial services it has a name and an owner. Two rules already assume that behind any decision that affects a customer stands a person who can be asked to justify it. One, the Consumer Duty, requires firms to be able to show they are delivering good outcomes for customers, not merely that they followed a process. The other, the Senior Managers and Certification Regime, requires a specific, named senior person to stand personally behind their area, so that when something goes wrong there is an individual who has to answer, not just a company to fine. Both were written on the assumption that a decision can be explained after the event by a human who owns it. That assumption is exactly what heavily automated lending strains. The rules do not need rewriting so much as they need firms whose systems can actually answer to them. That is a design problem before it is a compliance one. If nobody owns the design, you have not automated the decision. You have abandoned it, and dressed the abandonment up as efficiency.
The regulator is asking the right question, one layer too shallow
To its credit, the regulator is circling this. The FCA’s Mills Review (an independent look at how artificial intelligence could reshape everyday financial services over the coming decade, with lending decisions squarely in its scope) published on 6 July 2026. That is welcome, and it is the right question to be asking. My only quarrel is with where the debate around it tends to stop.
It has settled on explainability: can the decision be explained? That matters, but it is not quite the binding question, and a decision can clear it and still leave no one truly answerable. A model can produce a tidy list of the factors behind a decision and there can still be no person who owns the outcome. The deeper question is not whether the decision can be explained, but who is on the hook for it, and whether the system was built so that person can actually answer. Explainability is a property of the model. Accountability is a property of the firm and the people in it. The second is the harder thing, and it is what will separate the lenders who endure from the ones who merely automated.
What winning actually looks like
The banks that come through the next decade well will not be the ones with the cleverest models. Clever models will become ordinary; before long everyone will have much the same capability, and cleverness at that level will stop being an advantage. The winners will be the ones who remembered that lending is a craft, and who built the machine to serve that craft rather than to replace it.
In practice that means letting the machine own the volume, without apology and without a human bottleneck pretending to add a safety it cannot add. It means the judgement, the hard-won sense of where to draw the line and how much loss to accept, is deliberately and legibly built into the design. And it means that behind the whole apparatus stands a named person who can be asked why it lends as it lends, and who can answer. A decision made in a moment; an account of that decision that lasts for as long as anyone might reasonably ask for it.
The bank clerk in the sketch could not say why. That was the joke. We are now within reach of building it for real, at a speed and scale the sketch never imagined, and calling it a breakthrough. The craft is the thing that stops us. Let us not build the unknowing bank clerk.
Francis Hellawell
I solve hard problems in banking.