How Goldman Sachs and JPMorgan Rolled Out AI to Their Bankers

Forty thousand people. Out of about forty-six thousand two hundred. That's how much of Goldman Sachs now opens the firm's own AI assistant every single week... and asks it millions of questions a month.

Millions? That's not a pilot. That's a habit.

It's plumbing now. Which leaves the question everyone in finance is asking very quietly. If the machine drafts the deck, builds the model, reads the data room... what exactly is the twenty-three-year-old analyst for?

Hold that, because I have opinions. First, what does the thing actually DO?

Goldman announced the firmwide launch in an internal memo in mid twenty twenty-five. Their chief information officer, Marco Argenti, called it an important moment... the first generative AI tool at the firm to reach that scale. They'd spent about a year quietly piloting it with ten thousand people first.

A year of secret testing before the press release. That is very Goldman.

Argenti's pitch was reducing repetitive work. Drafting reports. Analyzing large datasets. A force multiplier, is the phrase he used. And the whole thing sits inside Goldman's own infrastructure, so client and trading data never touches an outside model.

Honestly that matters more than the chat window. Confidentiality IS the product in banking.

Now cross the street. JPMorgan put out something called LLM Suite in the summer of twenty twenty-four. Eight months later, two hundred thousand employees were using it.

Two hundred thousand. In eight months. Nobody changes a work habit that fast. How?

They made it opt-in. Their chief analytics officer, Derek Waldron, said it created healthy competition... driving viral adoption. People wanted it because the person next to them had it. They use it for client-ready presentations, earnings transcripts, comparing financial documents.

Okay, and the payoff? Every bank claims a payoff.

The documented number is small, and I love it for being small. Three to six hours saved per person per week. Waldron's also said on average people get an hour or two back.

So the great AI revolution is... a Friday afternoon.

Times two hundred thousand people, that's a lot of Fridays. And the ambition is louder than the number. Waldron talks about building the world's first fully AI-connected enterprise.

Nothing ominous in that sentence. None at all.

And it's everywhere. Citi's chief executive Jane Fraser said in mid-October that nearly a hundred and eighty thousand employees across eighty-three countries have access to the bank's own AI tools, and that automated code reviews alone save a hundred thousand developer hours a week.

Developer hours, though. That's engineers. Is anybody saying it about BANKERS?

And that's the honest gap. The hard disclosed numbers are adoption and hours. The stuff about analyst classes shrinking? Much softer. Treat it that way.

So no bank has stood up and said, we're hiring half as many juniors.

No. What IS on the record is structural. In October twenty twenty-five Goldman announced a program called OneGS three point oh. Redesigning onboarding, know-your-customer checks, reconciliations, vendor management... around teams of people, software and AI agents.

People, software and agents. They're writing the machine into the org chart.

And there are now more than a hundred specialized versions of that assistant, each layered on a corner of the bank's own data. That's the real shift. Not one chatbot... a hundred little ones.

Here's my pushback, though. A comps analysis, a discounted cash flow, a pitch book... yes, repetitive. But that's also how you LEARN. If the model does the reps, where does the thirty-five-year-old banker come from?

That's the tension inside these firms. Except the apprenticeship was never really about the spreadsheet. It was about sitting in the room while someone senior tells a chief executive a number he does not want to hear.

And you can't hand that to a model. Because the model can't be blamed.

That's the whole thing. A bank gets paid for judgment, for relationships, for confidentiality... and for being the name on the fairness opinion when it all goes wrong. An agent can draft the page. It cannot carry the liability.

Unlike trading desks, where the machines genuinely did take the chairs.

Completely different. Trading automated because the task was measurable and instant. Advisory is slow and relational, and deal volume comes from a chief executive deciding to do something brave on a Tuesday. No model generates that.

So from out here, what do I actually believe?

Believe the adoption numbers. Banks disclose them, regulators watch the model risk and the client data. Be skeptical of any headcount claim that isn't in a filing. One is measured. The other is a vibe with a press release attached.

A vibe with a press release. That's most of finance news.

And here's why it reaches your desk, whatever your desk is. The same split is coming. The drafting goes. The deciding, the vouching, the standing behind it... stays. For now.

Until someone builds a machine that can take the blame.

Nobody has. And that, more than any model, is what the analyst is still for.