From View From The Top, Issue #42
AI answers. Agentic AI executes. But the decisions that make or lose the money are analog — and they always will be.
For two issues now I’ve been making the tech case. The sandwich that gets made rather than just described. The sixty to seventy per cent of a lending platform’s cost that is information work, about to become AI generated. The spread that compresses and the £3m loan that finally becomes worth doing.
More than one reader has written back with a fair question: if the machine is that good, what exactly are the humans for?
This issue is my attempt of an answer. And it starts with a distinction the industry keeps blurring — because there aren’t two categories in this conversation. There are three.
THREE LAYERS, NOT TWO
AI answers. Ask it how to make a sandwich and it gives you the recipe. It drafts, summarises, explains. Useful. A cheap assistant at the next desk.
Agentic AI executes. It opens the fridge, selects the ingredients, makes the sandwich, plates it, and hands it to you. It does the thirty steps — reads the pack, reconciles the costs, builds the model, flags what doesn’t fit the credit box. This is the layer that restructures the cost of running a lending book, and it’s the layer Parts 1 and 2 were about.
And then …… there is the third layer, which has no silicon in it at all: analog judgement. Then human side of things.
So do you take for certain the decision the tech hands up. Whether to do the deal at all. Or whether to trust this sponsor on this site. But what do you do when the thing the model assumed at step two or three turns out to be wrong — because the machine that made the sandwich beautifully cannot tell you whether it should have been made. It carries the bad ingredient confidently to the end, plates it, and presents it with total conviction.
Somebody has to be standing in the kitchen. And not just anybody.
NO TECH MAKES THESE DECISIONS. NO YOUNG HUMAN DOES EITHER.
Here’s the uncomfortable symmetry the technologists miss.
The decisions that sit in the third layer can’t be made by a machine — and they can’t be made by a human without years in the trenches either. They are not intelligence problems. They are experience problems. The raw material isn’t data. It’s cycles lived through, deals gone wrong, experience and the pattern recognition that only forms after you’ve been wrong at full price.
Take the workout I published last summer — the one readers still forward more than anything I’ve written since. Walk through its decisions and ask, honestly, which layer each one lived in:
Appointing an administrator early, on our own initiative, before a creditor forced worse terms — when every instinct says wait. Analog.
Keeping the delivery team in place when the textbook says clear the decks and re-tender. Analog.
Committing millions of new money to a scheme that had already gone wrong, because the honest maths said the other path was a guaranteed loss. Analog — and the kind of call that requires having conviction and capital, plus the scar tissue to know the difference between conviction and hope. Analog.
Re-underwriting the exit three times as the market spoke, rather than fixing it at default and defending it. Analog.
No model makes one of those calls. Not today’s models, and — not tomorrow’s either, because the cases that matter are the ones too rare, too particular, and too badly recorded to ever be in the training data. Workouts are where records are thinnest and most misleading. The file says what happened. It never says what it felt like the week before it happened, or what the alternative would have cost.
THE MONTH-FOURTEEN CALL
The clearest example of the third layer, and the one I keep coming back to.
Let’s say you wrote the loan for twelve months. And it’s now month fourteen. The scheme is ninety per cent finished, and the borrower needs more time and more money to get the roof on.
Is that a delay worth funding, or a failure worth enforcing?
On the file, the two look identical. The same overrun, the same request, the same spreadsheet. Every data point the machine can see is the same in both cases. The difference lives in things that were never written down: how the sponsor behaved the last time the ground shifted, whether the contractor’s programme was honest three months ago, what the tone in the site meeting was, what this particular market does to a half-finished building.
That decision is made by a person who has been wrong before. There is no other qualification for it.
THE MOST EXPENSIVE PART IS ALSO THE POINT
Now put the three layers into the cost anatomy from last week — and notice what happens to the value.
The first two layers are cost. Real cost, but cost — and a cost that’s falling fast.
The third layer is the revenue. The relationship that produces the deal in the first place — the broker who calls you rather than somebody else because of fifteen years, not because of your interface. The judgement that prices the risk. The call that turns a would-be loss into a recovery. Every pound this business has ever earned traces back to a decision made in the analog layer.
Which means the machine doesn’t diminish judgement. It does the opposite:
The judgement concentrates to data.
When tech does the thirty steps, the human contribution compresses into a handful of moments where everything rides on being right. Fewer decisions per person, each one carrying more of the outcome. Experience becomes more valuable per hour than it has ever been — because each hour of it is applied to the calls that actually move the money, not to assembling the pack.
And that raises a question the industry has not begun to grapple with. The traditional way of acquiring that judgement — years of doing the grunt work, building the models, chasing the valuations, sitting quietly in the credit committee — is exactly the work the machine is eating. We are about to make experience more valuable and its apprenticeship rarer at the same time.
That’s next week’s issue. It deserves its own.
HARD LANDING
Forty years ago the trading floor was full of people doing what machines do now. The ones who survived weren’t the fastest clerks. They were the ones whose judgement was worth keeping after the clerking disappeared.
The same sorting is coming to this industry, and the sorting question is brutally simple: when the machine does your thirty steps, what exactly do you bring to the kitchen?
For some people the honest answer will be nothing. For the ones with real scar tissue, the answer is: the only part that was ever worth paying for.
AI answers. Agentic AI executes. Humans decide.
The order of that sentence is the future of this business.
