From View From The Top, Issue #44
The tech does the grunt work. The grunt work was the apprenticeship.
This past week I sat in a private session led by the head of a leading investment manager, on private credit and AI in private credit. The conversation was good — deployment, cost per deal, speed. A year ago these rooms debated whether. Now they discuss how fast it can do it.
But in ninety minutes around a dinner table talking about agentic AI doing the work of analysts, almost nobody asked what happens to the analysts.
This same week, the answer arrived anyway — in several press releases. Citi cut its analyst programme from three years to two. Goldman is putting new hires in charge of AI agents from day one — supervising the machine rather than doing the work. AI job postings at banks are up 49% in a year.
Three issues into this series, I’ve made tech’s case as strongly as I can: the cost structure falls, the spread compresses, the £3m loan becomes worth writing, the judgement concentrates. And I believe all of it.
But every cost structure that “falls away” is partly made of people — and I’ve spent forty years hiring them, promoting them, and sometimes letting them go. The least this series owes them is honesty about what’s coming next in my opinion.
WHOSE WORK GOES
Let’s be specific, because vagueness is how industries avoid this conversation.
Go back to the cost anatomy from Part 2 and the sixty to seventy per cent that is information work. Now put faces on it. The analyst who assembles the pack — the two weeks of collecting, checking, rekeying, formatting that precede every credit committee. That work is already going. We ran a nine-figure first pass in an afternoon.
The portfolio team’s reconciliation layer — the people whose week is spent making the contractor’s numbers, the QS’s numbers, and the lender’s numbers agree. Software work now.
The reporting function — the monthly pack that three people read, assembled by someone who spent two days on it. Gone to the machine, and honestly, few will mourn it.
The fund administration layer — the NAV calculations, the investor statements, the compliance checks that are rules applied to data. Rules applied to data is the exact definition of what agentic systems do.
What does not go: the origination relationships, the sponsor judgement, the month-fourteen uncomfortable call, the step-three check. The analog layer. But count the chairs. In a typical mid-market lending platform, the analog layer is maybe a fifth of the headcount. The agentic agents aren’t trimming the team. It’s inverting its shape.
The honest sentence: most of the jobs in investment management exist because information used to move slowly — and information doesn’t move slowly anymore.
I’VE SEEN THIS BEFORE. LET ME TELL YOU HOW IT ACTUALLY WENT
The trading floors I started on employed thousands. Clerks, runners, specialists, clearing staff. Within fifteen years, most of those jobs didn’t exist.
Here’s what the tidy version of that story leaves out.
The tidy version says: volumes exploded, the industry grew, everyone found new roles. Partly true. Trading volume went up a hundredfold. Whole new professions appeared — electronic market-making, quant research, systems. The industry employed more people a decade later than it ever had.
But they were not the same people. The clerk who matched tickets for twenty years did not, by and large, become a quant. Some made the jump — the ones who were young enough, adaptable enough, or lucky enough to have a boss who moved them early. Many didn’t. The industry’s transition was a triumph. Thousands of individual transitions were not.
Both things were true at once. That’s what an honest account of a technology transition looks like, and anyone who tells you only one half is selling something.
So when I say real estate credit in 2026 is the trading floor in 1985, I mean all of it. The volume explosion is coming — the £3m loans, the underserved middle, the thousands of SMEs that finally get funded. And so is the sorting. Both at once.
THE APPRENTICESHIP PARADOX
Now the part that worries me more than the job losses — because the job losses, at least, are visible.
Every credit judgement I trust was built the slow way. Years of assembling packs. Building models by hand. Chasing valuations. Sitting silently in credit committee watching someone greyer get it right — and occasionally, instructively, get it wrong. The grunt work was never just work. It was the apprenticeship. It’s where pattern recognition comes from: a thousand small exposures to how deals actually behave, accumulated before anyone trusts you with a decision.
The machine eats exactly that work.
And here’s the paradox this series has been building toward: Part 3 argued that judgement becomes more valuable as it concentrates — someone has to spot the mistake at step three, make the month-fourteen call, read the sponsor across the table. Experience becomes the scarcest asset in the industry at the precise moment we remove the way experience has always been built.
Citi’s shorter programme and Goldman’s day-one agent supervisors are answers by suggestion — nobody has really debated them out loud. A junior who supervises machine output from day one sees a hundred times more deals than I did at that age.
But seeing is not doing. Reviewing the machine’s model teaches you what right looks like. Building your own, wrongly, three times, teaches you why — and the difference between those two kinds of knowing is exactly what shows up at step three, twenty years later.
We are running an uncontrolled experiment on the next generation of judgement, and the results arrive in 2040.
AND WHAT WE’RE DOING
I don’t have a policy answer. I distrust people who claim one this early. But I run a platform going through this transition, so I don’t get to stay theoretical. Three principles we’re working to:
Honesty beats false comfort. Nobody at our firm is told the machine won’t change their job. It will. I think we are all aware of the efficiencies. The people who thrive in a transition are part of it, not the ones protected from the news until the news arrives.
The ladder has to be rebuilt deliberately. If the grunt work no longer builds judgement, something else must build it — and it doesn’t happen by accident. The juniors in the room for the month-fourteen calls, not just the clean approvals. Deliberate exposure to the workout files, the mistakes, the deals that died — the material the machine can’t teach from because, as I wrote in Part 3, the records are thinnest exactly where the lessons are richest. Apprenticeship by design, now that it no longer comes free with the job.
And the arithmetic of obligation. Firms will bank enormous savings from this transition. Some of that saving has a claim on it — retraining the people whose work went to the machine, moving them toward the analog layer where humans still compound in value. Not all will make that move. Pretending otherwise would be the false comfort. But a firm that pockets the full saving and shrugs at the people is making a statement about itself that borrowers, investors, and the next generation of talent will all hear.
HARD LANDING
Forty years ago, someone on a trading floor decided whether a clerk got moved onto the new systems desk or got a month’s notice. Nobody remembers that man’s name. Thousands of careers turned on which kind of boss he was.
That decision is now sitting on the desk of everyone who runs a lending platform, a credit fund, an investment firm. The machine is coming either way — the economics settled that argument, and this series has spent four issues showing how. What’s not settled is what kind of transition we run: the kind that moves people early, retrains honestly, and rebuilds the ladder — or the kind that discovers, in 2040, that it saved a generation of costs and lost a generation of judgement.
The machine underneath is cheap. The grey hair on top never will be.
