TL;DR: Anthropic released ten pre-built AI agents on May 5 that automate the work junior investment bankers do at 3am. Pitchbooks, credit memos, KYC screening, month-end close. The same week, the head of Claude Code told CNN that “software engineering will go away by the end of the year.” Goldman, Citi, JPMorgan, Visa, and AIG are already using the agents. Anthropic’s share of US enterprise AI spending climbed to 40% while OpenAI’s collapsed to 27%. Best for: understanding why banks aren’t replacing analysts yet but are quietly slowing recruiting. Not ideal for: anyone hoping the AI jobs story is overblown.
A week ago, the head of Claude Code told CNN that “software engineering will go away by the end of the year.”
Two days later, Anthropic shipped ten AI agents that automate the work of junior investment bankers.
Nobody is connecting those two events. They should be.
What Actually Launched
On May 5, Anthropic walked onto a stage in New York with Jamie Dimon and announced ten pre-built AI agents aimed at the most labor-intensive work inside a bank. Pitchbooks. Credit memos. KYC screening. Earnings reviews. Month-end close. Statement audits. The job that armies of 24-year-olds in suits do at 2am for $130K base plus bonus.
The agents ship as plugins for Claude Cowork and Claude Code, or as fully managed services running on Anthropic’s own infrastructure. They run on Claude Opus 4.7, which sits at the top of Vals AI’s Finance Agent benchmark at 64.37%. Hand one of them a target list and it spits back a comps model in Excel, a pitchbook drafted in PowerPoint, and a cover note in Outlook. The whole thing carries context across the four apps so the analyst (the one who still exists) doesn’t have to re-explain anything as the work moves between them.
Anthropic also rolled out full Microsoft 365 integration the same day. Excel, PowerPoint, and Word add-ins are live. Outlook is in beta. Moody’s plugged its entire ratings platform in as a native app. Dun & Bradstreet, Verisk, Fiscal AI, Third Bridge, Guidepoint, and IBISWorld all came along for the ride.
The data partners already in the stack include FactSet, S&P Capital IQ, MSCI, PitchBook, Morningstar, LSEG, and Daloopa. Anthropic isn’t selling a chatbot to banks. It’s becoming the operating layer Wall Street runs on top of.
Who’s Already Using It
Goldman Sachs. Citi. Visa. AIG. JPMorgan (Dimon was literally on stage with Amodei per Fortune’s coverage). Anthropic disclosed that financial institutions now make up roughly 40% of its top 50 customers. Anthropic’s share of US enterprise AI spending climbed from somewhere south of OpenAI’s to 40% by early 2026, while OpenAI’s collapsed from 50% to 27% over the same window.
This isn’t a pilot. This is deployment.
The Quiet Part
Boris Cherny (Anthropic’s head of Claude Code) told CNN in March that the term “software engineer” would be obsolete by the end of 2026. He said “builder” was the more fitting title going forward.
Everyone treated that as a hot take. It wasn’t. It was a roadmap.
Because if you’re the company that just shipped agents that build pitchbooks, run earnings analysis, and reconcile general ledgers (the literal output of an analyst’s first three years on the desk), and you also believe software engineering as a job is on the way out, what you’re saying is: every workflow that exists today in front of a screen is in our sights. Finance happens to be where the unit economics work first because the salaries are highest and the work is the most templated.
Anthropic’s chief economist Peter McCrory put it on the record at the same event. AI is now being used for at least a quarter of tasks in roughly half of all US jobs. A year ago that number was a third. He projected AI could add 1.8 percentage points per year to US labor productivity over the next decade, which sounds nice until you remember productivity gains don’t have to mean more jobs. They can also mean fewer.
When a Wall Street Journal report was thrown at the panel suggesting AI could trigger 20% to 30% unemployment, Dimon and Amodei both said the same thing: we don’t know.
That’s a hedge, not reassurance.
What Banks Are Actually Buying
The pitch to a Goldman or a JPMorgan is clean. You pay junior analysts six figures to spend their first two years building pitchbooks at 3am. Anthropic’s Pitch agent does it in minutes for the price of an API call. The math is so obvious it doesn’t need a slide.
The marketing language Anthropic uses is “alongside human analysts.” The agents are positioned as augmentation. Analysts review, iterate, and approve before anything goes to a client. There’s a “staircase of autonomy,” in Chief Commercial Officer Paul Smith’s words, that runs from basic research assistance all the way up to fully autonomous. Anthropic is selling banks the entire staircase, not the bottom step.
The “staircase” is the giveaway. It’s a roadmap for how this rolls out. Year one, the agents help analysts do more in less time. Then the next year analysts get reviewing work instead of building it. And finally year three, you don’t need as many analysts.
Banks haven’t said this part out loud yet. They don’t have to. Hiring just slows. Recruiting freezes. The 2027 analyst class is smaller than the 2026 one. Nobody gets fired. Nobody writes a press release. The job category just quietly shrinks.
This is the same playbook Meta ran with engineering. Same playbook Oracle ran with everyone. AI doesn’t have to replace you to take your job. It just has to make the case for not hiring your replacement.
The Part Nobody’s Asking
Here’s the question that didn’t get asked at the New York event: if Claude Opus 4.7 scores 64.37% on Vals AI’s Finance Agent benchmark, what happens when it scores 90%?
At 64% you still need analysts to review the output. At 90% you don’t. And the gap between those two numbers historically gets closed in about 18 months in this industry.
The agents are also explicitly designed to run autonomously as Claude Managed Agents on Anthropic’s platform. Long-running sessions. Per-tool permissions. Managed credential vaults. Full audit logs. Anthropic built the production infrastructure for autonomous deployment before the autonomous version was even good enough to use. They’re building the road for a car that doesn’t quite exist yet.
It will.
Why Finance First
Investment banking has the highest salaries, the most templated workflows, the most structured data, and the most explicit approval chains. Every junior banker’s job is to take inputs (a target list, a 10-K, a comps universe) and produce outputs (a model, a deck, a memo) under tight rules and tight deadlines. That’s a perfect agent task. Compare that to a therapist or a kindergarten teacher and the difference becomes obvious.
Banks are also the customers most willing to pay. A bank that saves $50M in junior salaries by adopting Claude doesn’t care if the contract is $5M a year. The ROI is so absurd it doesn’t even need a finance team to sign off (which is funny in its own way).
Finance is the beachhead. Legal is the next likely target. Anything where the work is structured, the salaries are high, and the output can be templated.
What This Means
A trillion-dollar industry just got a new vendor that quietly redrew its labor curve. The story isn’t ten new AI agents shipping. The story is that the company most plausibly building the next operating layer of work just told you, in plain English, what the next two years look like.
And the people whose jobs are about to change first are the ones least likely to read the press release.
