Will AI Replace Investment Banking Analysts: What Banks Have Said

Will AI replace investment banking analysts? Not on the record. Every bank that discloses a hiring number describes a trimmed starting class, not a vanished one, even as executives describe AI doing an analyst's classic first-year work.

In January 2025, Goldman Sachs CEO David Solomon told a room at Cisco's AI summit that artificial intelligence could complete 95% of an IPO prospectus in a few minutes, a filing that used to take a six-person team two weeks. Eight months later, JPMorgan's chief analytics officer stood in front of a CNBC camera and had the bank's own AI tool build a client-ready deck in about 30 seconds.

Then look at what the same banks did with their next hiring class. Goldman still hired thousands of entry-level employees and interns in 2026, and Bank of America held both its classes flat. That gap, between what executives say the tools can do and what banks actually did about hiring, is the real story.

AI is taking over the analyst's first-year workload, not the analyst, and the bar for the seats that remain is something a candidate can prepare for months before an interview.

Is AI already replacing junior bankers?

JPMorgan's own careers page describes a full-time investment banking analyst as someone who will "write reports, build updated financial models, and support multi-billion dollar transactions." The page does not mention AI once. What the banks' own executives, and the vendors selling them software, say is a lot more specific.

JPMorgan's own Asia Pacific head of investment banking, Paul Uren, put it plainly: "We're finding that AI streamlines the preparation of content and materials, as well as helping bankers engage with more clients more efficiently," adding that the bank is "in the early phase adopting AI tools throughout our investment banking business globally."

OpenAI's VP of Product, Nick Turley, described the same shift on CNBC's live demonstration of its banking product: "We're effectively teaching ChatGPT to research like an analyst and back up its conclusions like an analyst as well."

He compared it to an older shift in the same seat: "In the same way that Microsoft Excel transformed the industry and allowed them to produce better analysis faster, you will see technology like this do the same."

Anthropic pitches Claude for Financial Services for due diligence, market research, competitive benchmarking and "generating institutional-quality investment memos and pitch decks," aimed at the exact deliverables a first-year analyst spends late nights on.

Notice what is missing from that table: judgment about whether a deal makes sense, a client conversation, a decision about what to recommend. That gap between what the tools do and what a banker still has to judge is where the next section picks up.

Checking the machine: a worked example

Take a typical first-year task: pulling comparable-company multiples for a mid-cap payments business ahead of a pitch. A tool built for exactly this, in the style of what OpenAI and Anthropic now sell to banks, can assemble a peer set, calculate EV/EBITDA and EV/revenue multiples, and drop a formatted table into the deck within minutes.

The output looks finished. The analyst's job starts there, not before it.

Three things get checked, every time:

• The peer set. Did the model include a company from an adjacent sub-sector, a payments-adjacent hardware vendor sitting in a software comp set, because the two share a data-provider tag?
• The pricing date. A multiple moves with the exact day a market price was pulled, and a tool working off a stale price feed will reuse a print from three weeks earlier.
• The share count. A convertible note or an unexercised option pool changes the fully diluted share count a multiple should be built on, a gap a model trained to sound confident will not flag on its own.

That is a research task, a modeling task and a judgment call, run three times over on one exhibit.

What the banks are actually saying about hiring

That checking pattern is what training programs now describe. What individual banks say about the hiring itself is more scattered, and it is worth reading bank by bank before trusting any summary of it, including this one.

Jamie Dimon

JPMorgan is the bank whose executives have said the most about AI and jobs, across several years. Its 2025 shareholder letter states plainly: "AI will definitely eliminate some jobs, while it enhances others," alongside a promise that "our firm will have definitive plans on how we can support and redeploy our affected workforce." The same letter adds a line rarely repeated alongside the eliminate-jobs one: "There is a huge workforce shortage for many well-paying white- and blue-collar jobs."

Dimon has added to that picture on several occasions since:

• July 2026, on JPMorgan's Q2 earnings call: the bank had counted "almost 1,000 use cases today," with "the really important ones" numbering about 50, spanning "risk, fraud, marketing, hedging, prospecting, note-taking, idea generation, document reading. And it's kind of just starting."
• Same call, on AI efficiency: "We have had discrete areas where we did reduce jobs by 30% or 40%. And most of those people were offered jobs elsewhere." He named no department.
• May 2026, on Bloomberg Television in Shanghai: "I think we will be hiring more AI people and fewer bankers in certain categories, and it will make them more productive." The qualifier, "in certain categories," is the part that keeps getting dropped when the line travels.

Reuters reported that Dimon also pointed to the bank's roughly 10% annual staff turnover as the mechanism, letting JPMorgan manage the shift through retraining and attrition rather than mass layoffs.

David Solomon

Ask Goldman's CEO the same question over ten months and the answer sharpens each time, which is itself informative:

• October 2025, to Axios: "You're going to see changes in the way analysts, associates and investment bankers work," but on cutting headcount outright, "I don't think it works that way."
• January 2026, on Goldman's Exchanges podcast: "I'm not in the job apocalypse camp," adding, "if we get this right, I don't think it significantly lowers the number of people we have."
• June 2026, on Bloomberg's Odd Lots: "You're going to see nuanced changes that probably to some degree reduce the number of people that we start with over the next few years," he said, "but probably not what you and I would call dramatically."
• August 2026, to Goldman's incoming interns: told them to challenge the model rather than trust it, because judging "whether something makes sense" is still a human job.

That last line has company inside Goldman. Jacqueline Arthur, the bank's head of Human Capital Management, put it this way: "what's much harder to instill are qualities like judgment, adaptability, and critical thinking."

Chris Churchman, a Goldman partner who co-chairs the firm's Global Banking & Markets AI working group, sounded a similar note on the same Exchanges podcast, warning of a "huge danger" that bankers "outsource our reasoning to these models" and lose the tacit knowledge that "was never written down" in the first place.

None of that caution shows up in Goldman's own headcount memo from October 2025, which promised to "constrain headcount growth" with a "limited reduction in roles" aimed at client onboarding, lending and regulatory reporting, not investment banking. Total headcount stood at 48,300 at the end of September 2025, about 1,800 more than at year-end 2024. President John Waldron has described the firm as "a human assembly line" where "digital agents will be our robots," with headcount staying "roughly stable."

Morgan Stanley

Morgan Stanley has said far less about analyst hiring specifically, even where it has said a great deal about AI generally. CEO Ted Pick told his second-quarter 2026 call that "the accelerating adoption of artificial intelligence" was one of two defining themes of the year, with no remark on headcount attached.

As a named design partner on OpenAI's banking product, the firm said, through a spokesperson: "We're working alongside OpenAI to bring that intelligence into how we research companies, develop analysis, and prepare advice." It has said nothing connecting AI to analyst-class size.

Citi

Citi is doing two things that are easy to conflate. It made AI prompt training mandatory for 175,000 of its roughly 229,000 employees, with 6.5 million prompts submitted by staff in the same period. Separately, a restructuring plan unveiled in early 2024 is cutting about 20,000 roles overall. CEO Jane Fraser's January 2026 memo to staff, written as that restructuring continued, was blunt about the standard behind it: "We are not graded on effort. We are judged on our results."

Bank of America

Bank of America is the other bank willing to put a number on its entry-level class. CEO Brian Moynihan told CBS's Face the Nation that the bank gets 200,000 applications and hires 2,000 people, and that when he asks the new graduates if they are scared of AI, "they say they are. And I understand that." His advice to them was simply to "harness it." Its most recent cycle held both its intern and full-time classes at 2,000 apiece.

Evercore and Lazard

The advisory boutiques echo the same split between enthusiasm and reassurance. Evercore, also an OpenAI design partner, said it has "an opportunity to apply frontier intelligence" to client questions.

Lazard's CEO, Peter Orszag, was asked directly at a Semafor event whether AI doing the spreadsheet and PowerPoint work would eventually shrink M&A fees. His answer: "I hope not." He described the bet instead as getting junior bankers "to master those skills and move up the value chain more into direct interaction with clients."

Line up every bank's public position and the split is consistent: retraining and productivity talk, and either silence or reassurance on headcount.

Every bank that publishes a number at all describes a trim measured in single digits, not the two-thirds figure that keeps circulating, which the next section traces to its source.

How many banking jobs will AI actually cut?

Two-thirds of junior banking jobs are about to disappear, or so the number that follows this topic everywhere claims. It is not one number, and none of its versions is a disclosed cut that any bank has confirmed.

The New York Times reported in April 2024, citing anonymous sources, that incoming analyst classes at Goldman Sachs, Morgan Stanley and other banks "could" be cut by as much as two-thirds. Both named banks denied it on the record within days: a JPMorgan spokesman said there were "no changes to our incoming analyst classes" planned, and a Goldman spokesperson said the firm had "no current plans to alter our incoming analyst classes."

More than a year later, in September 2025, CNBC reported that "one proposal being discussed at a major investment bank" would change the ratio of junior bankers to senior managers from 6-to-1 to 4-to-1, with half of the remaining juniors moved to lower-cost cities such as Bengaluru or Buenos Aires. CNBC never named the bank.

A follow-up piece did its own arithmetic on that ratio change and printed a headline naming JPMorgan and a two-thirds cut, an attribution the original reporting never made. Here is that arithmetic, worked through:

Current ratio, 6 to 1

60

Proposed ratio, 4 to 1

40

Onshore, after half moves to lower-cost cities

20

Illustrative. It shows how a ratio change CNBC reported as under discussion at an unnamed bank in September 2025 produces a two-thirds figure once offshoring is layered on top. No bank has confirmed adopting either step.

That arithmetic is internally consistent and entirely hypothetical: a proposal at an unnamed bank, run through a blogger's own numbers, wearing a named bank's logo by the time it reached recruiting forums.

The same round number gets confused with at least three figures that measure something else entirely. A McKinsey partner said in June 2026 that banks were cutting junior classes "as much as two-thirds," naming no bank and no method.

Accenture's research, reproduced in a Citi report, actually measures the share of US banking work time spent on tasks its model rates as highly automatable: 54%, not a jobs count and with no time horizon attached. For capital markets, the segment nearest investment banking, the same chart puts the figure at 40%, lower than the banking-wide number that gets quoted instead.

Goldman Sachs's own 2023 research found roughly two-thirds of US occupations exposed to AI to some degree, a measure of exposure that Goldman itself distinguishes from elimination. A separate estimate, Bloomberg Intelligence's "as many as 200,000" bank job cuts over three to five years, comes from a survey of 93 bank technology chiefs about their entire workforce, weighted toward back and middle office roles, not analysts.

Strip away the estimates and the surveys, and the hard numbers a bank has actually put its own name to are far smaller and far duller than two-thirds. Goldman Sachs took in about 2,600 summer interns in 2025 and about 2,500 in 2026, a drop of roughly 4%. Its July entry-level hiring runs at a similar scale, around 2,500 people, down from more than 3,000 during the pandemic-era hiring surge of 2021.

Applications, meanwhile, kept climbing: over 360,000 in 2025, up from 315,126 the year before, against an acceptance rate under 1%, itself down from roughly 5% a decade earlier.

2,500

Goldman interns, 2026

down from about 2,600 in 2025

2,500

Goldman entry-level hires, 2026

down from over 3,000 in 2021

2,000 and 2,000

BofA interns and full-time hires, 2026

most recent cycle

None of it carries an AI label. The pattern across every bank willing to publish a figure is the same: single-digit movement, or none.

What happens to young workers, measured

If the disclosed class-size numbers are this small, where does the real pressure on young hires show up? A team at Stanford's Digital Economy Lab set out to answer exactly that question using US payroll data, and their answer is a mechanism rather than a body count.

Employment among 22-to-25-year-olds in the most AI-exposed occupations sat about 19% below where it would be had it tracked employment among similarly aged workers in less-exposed occupations, a gap that had been 15% a year earlier. The researchers found "the adjustment appears to operate primarily through reduced hiring of young workers rather than increased separations."

Goldman Sachs's own economics team reached a related conclusion in a cross-country study covering more than 800 occupations: a 10% increase in an occupation's AI exposure was associated with only a 0.1 percentage point drag on annual headcount growth across France, Canada and the US, but "junior workers may face stronger headwinds to hiring due to AI adoption" than senior ones.

Neither study names investment banking specifically. Both describe the same mechanism a candidate actually needs to plan around: the squeeze shows up as a narrower door each year, not as existing analysts being shown out of it.

What automation has already changed inside a bank

Two specific stories show what that narrower door looks like in practice: one still unfolding, one already finished.

JPMorgan's LLM Suite, in brief

The tool behind Derek Waldron's deck-building demonstration has a name of its own: LLM Suite, JPMorgan's in-house wrapper that routes across several outside model providers, built so the bank controls how its own people use them. Its rollout, the models behind it, and what Dimon has said about it are covered in full on the dedicated page on JPMorgan's LLM Suite.

What matters here is simpler. JPMorgan's own leadership is treating AI-drafted work as a present fact of the job, not a future scenario, and building the infrastructure to make that the default rather than the exception.

What Goldman's cash desk already went through

This is not the first time a Goldman Sachs trading floor has had this conversation. In 2000, the bank's New York cash equities desk employed 600 traders. By 2017, according to remarks its then deputy CFO Marty Chavez gave at a Harvard academic symposium, only two human equity traders remained, supported by roughly 200 computer engineers.

Chavez told the same audience that Goldman had found "four traders can be replaced by one computer engineer" in currency trading, and that the bank had "already mapped 146 distinct steps taken in any initial public offering of stock," many of them, in his words, "begging to be automated."

In 2017, automating an IPO was a research idea a bank executive floated at an academic conference. In 2025, that same bank's CEO said AI could already do most of one. In 2026, Goldman still hired thousands of people into the entry-level job that, on the loudest version of the story, was supposed to have disappeared by now. Automation reshaped what a Goldman trading floor does for a living without eliminating the floor.

What skills do analysts need now?

None of this calls for waiting to see what a bank decides. Training programs already show what the bar looks like, and the preparation it calls for is not exotic.

UBS's Graduate Talent Program builds AI literacy directly into its structure: an "AI Fluency Pathway" built around "real-world use cases, responsible application and sound judgment." Citi's approach is broader: its mandatory prompt-training program, set for 175,000 employees, teaches what the head of learning, Peter Fox, called "the possibilities of great prompting versus basic prompting to generate impactful" results, with experienced staff finishing the course in under 10 minutes and beginners in about 30.

Training providers that work directly with new-hire programs at major banks describe a curriculum built around one repeated pattern: build the skill manually, apply AI to the identical problem, then compare the two outputs and discuss where the model added value, where it introduced risk, and where a person's judgment was the thing that actually mattered.

Wall Street Prep and Financial Edge, which say they train analysts at nine of the top ten global investment banks, describe that same curriculum covering governance alongside mechanics: hallucination detection, restrictions on material nonpublic information, and knowing when AI is categorically off-limits.

In practice it comes down to five things:

1. Learn to build one comparable-company table and one simple LBO model by hand before you rely on a tool to do either. You cannot audit an output you could not have produced yourself.
2. Practice describing a specific, concrete time an AI tool gave you a wrong or misleading answer, and how you caught it. Vague enthusiasm for "using AI" answers a question nobody asked.
3. Be ready to explain, specifically, how you used an AI tool on a real task and how you checked its output. Treat this as preparation, not optional extra credit.
4. Read what a bank's own leadership has actually said publicly about the job. It is a steadier signal than a headline number with no name attached to it.
5. Build one small project that demonstrates checking a model's work, not just prompting one. A short writeup of a model's mistake and how you found it says more than a certificate.

A structured walkthrough of exactly these questions lives on the interview readiness tool, and the wider library of guides covers what individual banks have built for themselves in house.

The bottom line

So, will AI replace investment banking analysts? Not on the evidence banks themselves have put on the record. What it replaces is the version of the job that used to take a first-year six hours of pulling comps by hand or building a deck overnight before a client meeting. What survives, and what banks are now visibly hiring and training for, is the version of the job that catches the machine's mistake before a client ever sees it.

That is a harder bar than showing up able to run a model. It rewards exactly the kind of candidate who can say, specifically, what a first-pass AI output got wrong and why, which is a skill built long before the first interview, not discovered inside one.