From View From The Top, Issue #39
The Sandwich Problem
What agentic AI actually changes about deal execution — and the mistake at step three
After almost three decades underwriting deals, I can tell you the technology around this space is now accelerating at light speed.
I’ve seen this movie before. The trading floors of New York and Chicago in the early eighties — where I started — were filled with thousands of people. Clerks. Runners. Independent clearing firms. An entire ecosystem of human machinery moving information from one place to another. Technology changed that world abruptly, completely, and few inside it saw what was coming until it had already arrived.
I’ve written about the clearing firms in this newsletter before. What I haven’t told you is that I’ve already tried — and failed — to build what’s coming next.
Several years ago we set out to build a proprietary platform for ourselves. We called it Credit Stream. The idea was right. I knew exactly what I wanted it to do — take the long sequence of information work that sits inside every credit transaction and run it as a system rather than a scramble. But the LLMs weren’t there. Matching the complexity of integrating multiple tasks — the reconciliation, the underwriting logic, the monitoring, the handoffs between them — was impossible with the technology of the moment.
Fast forward three years, and the world has already changed.
What was impossible when we designed Credit Stream is now not just possible — it’s arriving faster than most of this industry understands. And the cleanest way I’ve found to explain the shift is a sandwich.
Ask AI how to make a sandwich, and it tells you how. It gives you the recipe.
Ask agentic AI — and it opens the fridge, finds the ingredients, makes it, slices it, plates it, and hands it to you.
One of these is a productivity tool. The other is a restructuring of how investment work gets done.
This is the first issue in a series I’ll run through the autumn — the question I promised in the August stocktake: how does this industry actually improve its delivery? Over five issues: what the machines change, what the cost structure does, what only humans can still do, what happens to the people, and who wins.
Today: the machine itself, and its most dangerous habit.
INSTRUCTION VERSUS EXECUTION
Most of what our industry currently calls AI is the first kind of sandwich. It answers questions. It drafts the memo. It summarises the data room. Useful — like a cheap assistant at the next desk — but the work is still yours.
Agentic AI does the work. Reads the pack. Pulls the market evidence. Reconciles the cost schedule against the QS report. Builds the first-pass model. Checks the deal against your credit box and flags what does not fit. Thirty steps, end to end, at machine speed, without a human touching the middle.
This is not theoretical for us. We ran one last week.
A first pass on a nine-figure development loan — from broker pack to screening deck — the same afternoon it arrived. Work that used to take the team days. Not a better assistant for the analyst: the first draft of the analyst’s work, done before anyone had opened the pack.
Now map that against what a real estate credit transaction actually is.
Let’s strip away the mystique, and a development loan is a long sequence of information work. Collect. Verify. Reconcile. Model. Document. Monitor. Report. I described the friction in the spring — the £75,000 to £100,000 of process cost on a £5 million loan, the twelve weeks that should be four, the five parties rebuilding the same picture from different materials. In the summer I told you about our workout, where the hardest enemy wasn’t the market. It was stale information moving slowly between parties.
Every one of those frictions is a sequence of steps. And sequences of steps are exactly what agentic systems eat.
The underwriting pack that takes an analyst two weeks — assembled in hours. The monitoring cycle that runs quarterly because that’s all human capacity allows — running continuously. The daily completion tracker that saved our workout, maintained by hand at real cost — running automatically, on every loan in the book, without anyone lifting a finger.
THE MISTAKE AT STEP THREE
Now the part the demos skip. And the part that, forty years of watching markets adopt new machinery tells me, will separate the winners from the casualties.
It makes mistakes. And a mistake at step three of thirty is carried, confidently, beautifully, all the way to step thirty.
The wrong comparable slips into the valuation logic at step three. Every subsequent calculation inherits it. The model builds, the sensitivities run, the screening deck lands on your desk — polished, internally consistent, professionally formatted, and wrong. The sandwich gets made. Beautifully plated. With the one ingredient that should not be in it.
Humans make mistakes too — constantly. But human error is usually loud: the analyst hesitates, a colleague queries, the number “looks off” to someone with a feel for the market. Agentic error is quiet. It arrives wearing the formatting of competence. There is no hesitation in the output, because the machine doesn’t know it’s wrong.
So you stay in the kitchen.
Two disciplines follow from that.
-
The first is about the systems: the ones worth using are the ones that show their work — every number traced to its source, so you find the bad ingredient in seconds instead of remaking the sandwich. Auditability isn’t a feature. It’s the difference between a tool you can supervise and a black box you have to trust.
-
The second is about the people. The judgement doesn’t leave the room — it moves up a level. From doing the thirty steps, to knowing which three to check. From building the model, to interrogating it the way a credit committee interrogates a sponsor: where did this number come from, what’s it resting on, what happens if it’s wrong?
In an agentic world, the scarcest skill in an investment firm won’t be building the analysis. It will be knowing what wrong looks like — the pattern recognition that comes only from cycles, workouts, and mistakes paid for at full price.
The machine does the work. The grey hair checks step three.
WHAT ACTUALLY MOVES — THE CAPITAL ARITHMETIC
Here’s where this stops being a technology story and becomes a capital markets story. And I want to be precise, because the lazy version of this argument is wrong.
The cost of running a lending book is about to fall. Not by a little. Fewer hours per deal. Less overhead per pound lent. The expensive human middle of every transaction — the reconciliation, the document assembly, the monitoring headcount — compresses toward the cost of computation.
The lazy conclusion is that return expectations fall with it. They don’t — not directly. Investors will not lower the return they need because the manager got cheaper. Risk sets required returns, and the risk of a development loan hasn’t changed because the underwriting got faster.
But two other things move. And they matter more.
First, the spread. The gap between what a borrower pays and what the investor receives is partly process cost — the machinery I priced for you in the spring at 1.5 to 2% of a facility before drawdown, and 3 to 4% of returns annually across the fund stack. Cut the process cost, and a lender can accept a lower coupon from the borrower while delivering the same net return to the investor. The saving doesn’t vanish. It moves — to the borrower’s cost of capital.
Second, the minimum ticket. The fixed cost of underwriting has always dictated the smallest loan worth doing. When it costs broadly the same to underwrite a £3 million loan as a £30 million one, the £3 million loan never gets looked at — which is precisely why the small regional developer can’t find funding, whatever the headline appetite for the sector. Collapse the fixed cost, and the £3 million loan becomes worth doing. An entire tier of the market that was never economic to serve becomes servable.
Now connect both to the viability numbers I gave you two weeks ago — completions at a decade low, approvals at a record low, a market full of schemes that “don’t pencil.”
A cheaper coupon changes the arithmetic on every marginal scheme. A lower minimum ticket opens funding to the small schemes — the 20-to-100-unit projects where, as I’ve argued all year, the missing homes actually live. Neither fixes a scheme that does not work. Nothing fixes those.
But for the deals that fail on friction rather than fundamentals — the friction is about to get cheaper.
It won’t undo the 1947 settlement. But it attacks the other half of the viability problem — the cost of capital’s machinery — in a way no policy lever has managed in decades.
WHAT I’M WATCHING
Three things, as this unfolds — and I’ll return to each later in the series.
Who adopts first: the incumbents with the most to protect, or the platforms with the least to unlearn? History suggests the second. The clearing firms didn’t invent electronic trading.
Where the failures happen: because there will be step-three failures at scale, some public, and the firms that pair machine execution with senior human judgement will be separated — visibly — from the firms that fired the judgement along with the process.
And what happens to pricing: because the first movers will bank the cost advantage as margin, until competition passes it to borrowers. That window — where the cost structure has fallen but the pricing hasn’t — is where the next few years’ outperformance lives.
HARD LANDING
Every technology wave in finance has followed the same arc. The machinery gets cheaper. The intermediation compresses. The people who thought their work was the value discover the work was the cost — and the judgement was the value.
I watched it happen to the clearing firms from inside the world they served. I wrote in the spring that real estate was next, and that the last arbitrage was speed. I now think I understated it. The machine doesn’t just make execution faster. It reprices what borrowers pay, and what gets underwritten at all — and with it, what this market can build.
The sandwich matters more than it looks.
Next week: the cost structure of an investment firm, taken apart line by line — what the machine eats first, and where the saving actually goes.
