Goldman Sachs and the Claim That AI Replaced Its Bond Traders

There's a trading floor at Goldman Sachs in New York that used to hold six hundred people. Phones, shouting, screens, the whole cliché. By twenty seventeen, two people were sitting in it.

Two. Out of six hundred.

Two. And the man who ran the technology behind it, Marty Chavez, described it to MIT Technology Review like this... those six hundred traders, there's a lot of space where they used to sit.

That is a bleak sentence. Okay — so this is the story I keep seeing. Goldman wiped out its bond desk, six hundred-odd traders, November twenty twenty-four, replaced by an AI. Is that real?

I went looking for it. And I'll be blunt, because this is the whole episode. I can't find it. No documented mass elimination of Goldman's bond trading desk in twenty twenty-four. No filing. No wire story. No collapse in fixed income headcount.

Then where does six hundred come from? Somebody didn't invent that out of thin air.

No, six hundred is real. It's just a different desk and a different decade. Cash equities. Stocks. Year two thousand. That's the room. The claim you heard is that story, wearing a bond trader's suit with a twenty twenty-four date stamp on it.

A fact that got laundered.

And here's the part that annoys me. Everyone repeats the empty room. Almost nobody repeats what was standing next to it — which is the bit that actually tells you what's coming for your job.

Go on then.

Those two surviving equity traders weren't alone. According to that same reporting, they were supported by about two hundred computer engineers.

So the desk didn't shrink to two. It changed species.

Changed species. And not in one corner of the firm — by twenty seventeen, roughly a third of Goldman's entire staff were engineers. Chavez's point was that it wasn't finished. Currencies next. Then parts of investment banking. He said they'd mapped a hundred and forty-six separate steps in taking a company public, and a lot of them were, his word, begging to be automated.

Okay, but hang on. Bonds aren't stocks. A share of Apple is a share of Apple. A bond is a specific loan, specific maturity — one company's got dozens of the things, right?

That's exactly the right objection, and it's why bonds took twenty years longer. Dozens outstanding, most of them not trading on any given day. Try teaching a machine to price something that last changed hands in March.

So how did they crack it?

Boring work, at volume. Goldman built an algorithm for corporate bonds that scrapes publicly available quotes and spits back a firm, tradeable price. Amy Hong, who runs market structure strategy for global credit there, told the Financial Times the desk gets thousands of inquiries a day — and the algorithm handles those systematically, so the humans go work the hard ones.

So it's not replacing the trader. It's eating the trader's inbox.

The small, annoying, low-margin tickets. And the clients are automating from their side too. MarketAxess — one of the big electronic bond platforms — says in its own annual report that in twenty twenty-three, nearly a quarter of its trades came through an automated request-for-quote tool. Forty-one percent of those were what they call no touch.

No touch meaning nobody...

Nobody on the buy side pressed anything. The fund's own system fired the order off against pre-set rules. Up from thirty-three percent the year before. Both ends of the trade are now machines being polite to each other.

And somewhere a guy with twenty years of relationships is wondering what happened to lunch.

He still gets lunch. He just has it with the clients who need something weird. That's the pattern in every market that's been automated — the simpler and more liquid the product, the faster the code shows up.

So the messy corners survive. Distressed debt, the structured stuff.

Still voice. A phone call, a judgement, a favour owed. An algorithm needs history to price something, and some bonds just don't have any.

But I can't shake the viral version. If the direction is that clear... isn't it just early rather than wrong?

Maybe. But early and wrong look identical until they don't — and the honest answer today is that the mass firing didn't happen. What happened is slower and much harder to tweet. Fewer people per dollar of trading. Engineers doing the work that used to be shouting.

And the ones who survive get paid more, I assume.

That's what Tom Davenport at Babson predicted back in twenty seventeen — that the average managing director's pay at Goldman would probably get even bigger. The spread inside the building starts to look like the spread outside it.

Fewer rungs on the ladder. Taller rungs.

And this is where it reaches you. When a bond trade costs a bank almost nothing to execute, the spread it charges gets thinner. Thinner spreads flow into the funds holding your pension, your index tracker. You never see it. You just, very slightly, keep more.

So the robots' winnings land in my retirement account.

A sliver of them. And the bill lands on a career that used to start with a twenty-two-year-old answering a phone.

That's a trade nobody voted on.

Nobody ever does. There was no night the machines took the bond desk. They've been taking it one quiet ticket at a time... for twenty years.