Judging Judgment: How Junior Bankers Learn, and Get Evaluated, When AI Drafts the Work

On October 8, Kevin Sneader, who runs Goldman Sachs’ Asia-Pacific business and sits on its Management Committee, told the Milken Institute Asia Summit in Singapore that the bank’s new joiners now manage AI agents from the day they start. According to Bloomberg’s report, as picked up by TechRadar and Notebookcheck, he said young bankers have to make the most of the “virtual army” they now have.

Six weeks earlier, a different Goldman partner raised the other side of the same coin. Chris Churchman, who leads Marquee, Goldman’s digital platform for institutional clients, said on the firm’s Exchanges podcast that “there’s a huge danger here that in the era of AI, we outsource our reasoning to these models.” His goal, he said, is that a new joiner ten years from now would be known for the quality of their reasoning.

Put those two comments side by side and you have the question every analyst class now faces. For the last two years the debate has been about headcount: will banks need as many juniors? That question is not settled. But for the analysts and associates who do get a seat, a more practical question has arrived. If the first draft of the model, the comps and the deck comes from a machine, how do you learn the job, and how will your bank know you have learned it?

We have written before about how AI tools are pushing review work up to associates and VPs. This piece is about the other end of that chain: the junior banker who is now expected to direct the tools, check them and be accountable for what they produce.

What Is Actually Changing in the Junior Seat

The analyst job has always been an apprenticeship disguised as a production role. You built the model at 1 a.m. because someone had to, and in the process you learned how a business makes money, how a capital structure fits together and which numbers a senior banker will ask about first. The grind was the training.

AI tools change who does the first pass. Banks have been building toward this for a while. Citi said in April that 180,000 of its employees have access to its AI tools and more than 80% use them regularly, and it introduced Arc, a platform for building and running AI agents across the firm, starting with defined use cases built by its developers. Citi described the banker’s role as moving “from coordinator to architect and advisor.” Goldman’s own research, cited in our Rogo piece, suggested AI could automate roughly a quarter of current banking work hours, with modeling and deck formatting among the most exposed tasks.

What does not change is accountability. When a number in a board book is wrong, nobody blames the software. The analyst who sent it, the associate who reviewed it and the VP who signed off still own it. Citi said every one of its agents will be “monitored, auditable and governed.” In practice, the people doing the monitoring at the deal level will be juniors.

So the junior role is not disappearing so much as changing shape. Less time producing. More time directing, checking and explaining. That sounds like a promotion. It can be, but only if the analyst actually understands what the model did.

The Apprenticeship Problem

Here is the tension Churchman put his finger on. He described how junior traders at his former firm learned partly by handling routine client pricing requests under senior supervision. That work can now be automated. “We can absolutely automate that,” he said, “but then do we get the senior traders that fully understand?”

The same question applies in banking. An analyst who has built fifty LBO models by hand knows, almost by instinct, when a leverage number looks wrong. An analyst who has only ever reviewed AI-built models may not, unless someone has deliberately taught them what to look for. Judgment comes from repetition, from mistakes caught early and from sitting close enough to senior bankers to hear how they think about a deal.

Churchman said humans should stay responsible for high-stakes decisions and that the goal is to “empower human reasoning, not delegate it.” He also acknowledged, as reported, that Goldman has not yet fully worked out what that model should look like. That candor matters. If one of the most technology-forward banks on the Street is still figuring this out, most groups are too. Juniors should not assume someone has designed their training around the new tools. In many groups, nobody has.

From a recruiter’s seat, this is where we expect to see the widest gap open up over the next few years. Two analysts with the same title, the same bank and the same deal list could come out of their first two years with very different skills, depending on whether anyone made them understand the work the tools were doing.

How Banks Will Judge Juniors Now

If production speed matters less, what will staffers, VPs and review committees look at instead? Based on what we hear from hiring managers, a few things could carry more weight.

• Catching errors before anyone else does. The analyst who finds the bad formula, the stale share count or the comp that does not belong will stand out quickly. With AI output, the error rate per page may fall, but errors can be harder to spot because the work looks finished.
• Explaining the numbers. Can you walk a VP through why the model shows what it shows, without opening the file? Can you say which three assumptions drive the answer? That is the clearest sign you understand the work rather than just delivered it.
• Knowing when not to trust the tool. Seniors will notice who pushes back when output looks plausible but wrong, and who forwards it as is.
• Client-ready judgment earlier. If juniors are freed from some formatting work, banks will expect them to think about what the client actually needs from a page, not just whether the page is clean.

None of this is new. Good analysts have always been judged this way. What changes is that these skills become the main way to stand out, rather than one way among several. “Fast and tireless” was a real advantage when the work was manual. It is a smaller one now.

What Changes in Recruiting and Interviews

Banks are starting to say this out loud. UBS will require applicants for its 2027 graduate and intern intake in Global Banking and Markets to show they can use AI to improve outcomes and efficiency, according to a September report by the Financial Times, as summarized by The Next Web. Interviews will include questions about how candidates actually use AI tools. Successful hires then join an internal training program UBS calls the “AI Fluency Pathway.” UBS told the FT that AI skills complement, rather than replace, academic and social abilities.

We would expect more banks to move in this direction, in their own way. For candidates, that could mean:

• Technical interviews that test reasoning, not recall. If anyone can generate a DCF in seconds, the useful question becomes “this output says X; what is wrong with it?” or “which assumption would you challenge first?” Candidates who can only recite the steps could struggle.
• Case and modeling tests that look more like review exercises. Being handed a flawed model and asked to find the problems is a natural way to test the skill banks now need most.
• Specific AI questions. Be ready to describe a real task you used an AI tool for, what you gave it, what came back and how you checked it. Vague enthusiasm will not help. A concrete, honest example will.

For lateral and buy-side recruiting, the same logic applies one level up. Private equity firms have always tested whether a banker really understands the deals on their resume. As more of the build is automated, we expect those conversations to probe harder: why the structure was chosen, what the downside case looked like, what you would have done differently. Bankers who were close to the decision-making will have better answers than those who were close to the software.

What It Looks Like by Level and Platform

Analysts. The opportunity is earlier exposure to the substance of a deal. The risk is skipping the reps that build instinct. If your group lets the tools do the first pass, rebuild key pieces yourself at least once so you know how they work.

Associates. You are now the main quality check on both the analyst and the machine. That is more review work, as we noted in our Rogo piece, and it is also where you can show you are ready for VP. Associates who teach analysts how to check AI output, rather than just fixing it themselves, will be noticed.

VPs and directors. Your role as a trainer matters more, not less. If juniors learn judgment by working close to senior bankers, the VPs who make time for that will build stronger teams, and those teams will be the ones juniors want to join.

By platform. Bulge brackets have the largest technology budgets and the most formal rollouts, which means more tools, but also larger classes where individual coaching can get thin. Elite boutiques tend to run leaner deal teams with juniors close to senior partners, which could make them strong training grounds if the tools are used well. Middle-market firms vary widely. Some have adopted third-party tools quickly; others are still mostly manual. When you compare offers, it is fair to ask how a group uses AI on live deals and how it trains juniors to check the output.

What It Means for Your Career

The headcount question will play out over years, and we will keep tracking it. The judgment question is here now. A few practical takeaways:

• Treat AI output as a draft from a talented but unreliable analyst. Check it the way a good associate checks your work: line by line on the numbers that matter.
• Keep doing some work by hand. Not to prove a point, but because building a model from scratch is still how you learn what a model should look like.
• Stay close to senior bankers. Join the calls. Read the markups. Ask why a page changed. That is where judgment is passed on, and it is the part of the apprenticeship no tool replaces.
• Be ready to talk about it. In reviews, in lateral interviews and in buy-side processes, expect questions about how you used AI and how you made sure the work was right. Have a concrete example.
• Choose seats with this in mind. When weighing an offer or a move, the group’s approach to training could matter as much as its deal flow.

Sneader is right that new bankers now start with tools earlier generations did not have. Churchman is right that the tools could weaken the reasoning they are meant to support. Both can be true. The juniors who do best will be the ones who use the tools to do more, and still know the work well enough to tell when the tools are wrong.

Related reading

• Rogo Was Supposed to Shrink the Junior Class. Instead, It’s Creating More Work for Associates and VPs
• Citi Moves to a Two-Year Analyst Program: What It May Really Mean for Junior Bankers
• AI in Investment Banking: The 2025 Revolution of Junior Work

Sources

• Bloomberg: Young Bankers to Manage AI Agents From Start, Goldman Exec Says (Oct. 8, 2026; headline)
• TechRadar: Top Goldman Sachs executive says new workers will be managing ‘virtual army’ of AI (Oct. 8, 2026)
• Notebookcheck: Goldman Sachs expects new employees to manage AI agents from day one (Oct. 8, 2026)
• Moneywise: Goldman Sachs partner warns of AI cognitive atrophy (Aug. 27, 2026; citing CNBC)
• HCAMag: Goldman Sachs partner unveils ‘real danger’ in working with AI (Aug. 25, 2026)
• Citi: Introducing AI Agents, the next phase of Citi’s AI journey (Apr. 30, 2026)
• The Next Web: UBS now wants AI skills from its junior bankers, FT reports (Sept. 7, 2026)
• Prospect Rock Partners: Rogo article (Goldman research on share of banking hours exposed to AI)