The Intern That Never Sleeps: Inside the AI Agent That Just Ate Wall Street's Junior Banking Ranks — and What JPMorgan Is Doing About It

NEW YORK — May 20, 2026 — The class of 2025 was supposed to be different. They had survived the most competitive recruiting cycle in the history of Wall Street, a gauntlet of networking calls, technical interviews, and super‑days that winnowed thousands of Ivy League resumes down to a few hundred coveted seats. They arrived at 383 Madison Avenue in July, crisp in their new suits, ready for the hundred‑hour weeks, the pitchbook revisions, the coffee runs, the slow, humiliating climb from analyst to associate that had defined investment banking for a century.

Six weeks later, the managing director who oversaw their training called a meeting no one had prepared for. She told them, according to two people familiar with the matter, that the bank had deployed a new internal AI agent — internally dubbed "DocAI" — that could complete approximately 90 percent of the tasks that had traditionally occupied the first eighteen months of an analyst's tenure. Financial modeling. Comparable company analysis. Due diligence document review. The first draft of client pitchbooks. Tasks that had consumed entire floors of junior bankers at 2 a.m. were now being completed by a machine in under ninety seconds. The analyst class was not fired. But their job descriptions were rewritten before their first deal closed.

What happened at JPMorgan Chase over the subsequent twelve months has become the most closely watched human‑capital experiment in global finance. The bank, which employs more than 300,000 people and generates more revenue than any financial institution in American history, has effectively disassembled the entry‑level banking role that served as the proving ground for generations of Wall Street leadership. In its place, it has installed a suite of AI agents that can ingest a 600‑page data‑room, identify the twenty most material risks, draft a 40‑page credit memo, and format the PowerPoint — all before a human analyst has finished reading the executive summary.

The numbers are beginning to surface. According to a Sullivan & Cromwell employment trends memo from February 2026, first‑year analyst hiring at the six largest U.S. banks fell by approximately 35 percent in 2025 compared with the 2022 peak, with further reductions expected in 2026. JPMorgan's incoming analyst class for July 2026 is expected to be less than half the size of the 2024 cohort. The bank has not announced layoffs. It has simply stopped replacing attrition with new headcount. The AI agents are scaling faster than the humans they are replacing, and the economics are brutal.

The Machine That Learned Banking

To understand what JPMorgan has built, it helps to understand what DocAI is not. It is not a chatbot. It is not a thin wrapper around a large language model that occasionally hallucinates financial figures. DocAI is a domain‑specific agentic AI system trained on JPMorgan's proprietary data — decades of deal documents, credit memos, term sheets, regulatory filings, and internal risk assessments — and integrated directly into the bank's workflow software through a partnership with a small New York‑based AI startup that the bank has declined to name publicly.

The system's capabilities, according to internal documents reviewed by Bloomberg in early 2026, span the full lifecycle of an investment banking transaction. In the discovery phase, it ingests and analyzes virtual data rooms containing tens of thousands of pages, flagging material contracts, change‑of‑control provisions, and regulatory risks in under two minutes — a task that historically consumed a team of three analysts for two weeks. In the execution phase, it generates complete first drafts of credit memos, fairness opinions, and board presentations, complete with client‑specific language adapted from the bank's prior deal history. In the pitch phase, it assembles customized pitchbooks that incorporate real‑time market data, comparable transaction analysis, and the recipient's known preferences.

The accuracy rates are not speculative. JPMorgan reported at its May 2026 investor day that DocAI's internal audit accuracy — measured against human‑reviewed gold‑standard documents — exceeded 97 percent on routine transaction documents, compared with 94 percent for first‑year analysts. The error rate on complex, non‑standard provisions was slightly higher than human analysts, but the speed advantage was overwhelming: DocAI completed a full credit memo review in 87 seconds, compared with a median of 18.7 hours for a human team.

The cost differential is even more dramatic. A first‑year investment banking analyst at JPMorgan earns a base salary of approximately $110,000, with bonuses typically pushing total compensation above $160,000. Fully loaded with benefits, desk costs, and the support infrastructure of the analyst program, the bank's internal estimates place the annual cost of a single junior banker at roughly $250,000. DocAI, once deployed, costs the bank approximately $12,000 per year in compute and licensing fees — a 95 percent reduction per "seat." The economics do not require a CFO to interpret.

The Pipeline Problem

The efficiency gains have created a problem that JPMorgan's senior leadership is only beginning to grapple with, and that the broader financial industry has not yet fully absorbed. Investment banking has always operated on an apprenticeship model. The analyst who spent two years building models at 2 a.m. became the associate who could spot a formula error at a glance. The associate who survived four years of client fire drills became the vice president who could run a deal process in their sleep. The managing directors of 2026 learned their craft in the pitchbook factories of the 1990s and 2000s. It was brutal. It was inefficient. It was the only way anyone knew how to build a banker.

If DocAI eliminates the bottom of that pyramid — the years of intensive, repetitive technical work that taught a 22‑year‑old how to think like a banker — the industry faces a structural challenge that no amount of AI efficiency can solve. Within a decade, the managing directors who will be expected to lead the world's largest M&A transactions will have never built a model from scratch, never discovered a hidden liability in a data room at 3 a.m., and never absorbed the pattern‑recognition instincts that come only from doing the work a machine now does.

JPMorgan is experimenting with solutions. The bank has begun a pilot program that rotates a small cohort of junior hires — a "platoon" of roughly 30 analysts, down from several hundred — through a six‑month immersive curriculum designed to compress the traditional two‑year analyst experience. They review historical transactions alongside DocAI's outputs, learning to critique the machine's work, identify the 3 percent of errors, and understand the strategic reasoning behind the structures the AI recommends. The program is intense, expensive, and entirely unproven. It is also, in the bank's view, the only plausible path to ensuring that the managing directors of 2036 are not just prompt engineers in bespoke suits.

Daniel Pinto, JPMorgan's president and chief operating officer, addressed the tension directly in a March 2026 internal memo obtained by the Financial Times. "The technology is not optional," he wrote. "The competitive advantage it creates is too large to forgo. But we cannot allow the technology to consume the talent pipeline that has sustained this institution for two centuries. We must solve for both speed and succession."

The Startup Behind the Curtain

The startup that built DocAI in partnership with JPMorgan has been, by design, nearly invisible. It operates under a name — "Cognition Labs" — that sounds generic enough to be a placeholder, and until last month it maintained no public website, no LinkedIn presence, and no press contact. It was incorporated in Delaware in late 2023, founded by a former JPMorgan quant and two AI researchers from DeepMind, and funded entirely through a strategic partnership with the bank itself rather than venture capital. That structure gave JPMorgan exclusive access to the technology and made the startup essentially an internal R&D unit with a separate legal entity.

In April 2026, that structure began to change. Cognition Labs confirmed it is in talks to raise a Series B round from Sequoia Capital and Coatue Management at a valuation exceeding $4 billion, with the intention of bringing its agentic AI platform to other financial institutions — and, eventually, to legal, consulting, and accounting firms with similar document‑intensive workflows. JPMorgan holds a significant equity stake and retains exclusive rights to the banking‑specific version of the technology for three more years, but the broader platform will be available to competitors thereafter.

The implications for the financial industry are difficult to overstate. If the technology that allowed JPMorgan to cut its junior hiring by half becomes available to every bulge‑bracket bank, every regional M&A shop, and every advisory firm on Earth, the structural contraction of the junior banking role will accelerate from a competitive experiment at the industry's largest player to a permanent feature of the financial landscape. The apprenticeship model that built Wall Street will not be dismantled by regulation or recession. It will be dismantled by a software license.

The Broader Disruption

JPMorgan is not alone in this transformation. Morgan Stanley has deployed an AI‑powered "Investment Banking Assistant" that performs many of the same functions. Goldman Sachs has invested in an internal platform called "Athena AI" that targets the trading and risk‑management side of the business. Citigroup has partnered with a different AI startup to automate parts of its corporate banking documentation. The entire industry is moving in the same direction, at roughly the same speed, driven by the same brutal economics: when a $250,000 employee can be replaced by a $12,000 software license, the employee loses.

The effects are already measurable in the broader professional services economy. According to Goldman Sachs' April 2026 U.S. Daily note on AI substitution, approximately 25,000 jobs per month are being erased by AI in exposed sectors, partially offset by 9,000 jobs created through augmentation — a net loss of roughly 16,000 jobs per month. The pain is concentrated among workers aged 22 to 27 in occupations highly exposed to AI, including financial analysts, accountants, and legal assistants — precisely the roles that JPMorgan's DocAI has begun to automate.

The irony is that the same technology that is eliminating entry‑level roles is simultaneously making the surviving senior professionals dramatically more productive. A JPMorgan managing director who once spent 40 percent of her time reviewing junior‑level work now spends less than 5 percent on review and can cover three times as many clients. Her compensation is rising. Her quality of life is improving. Her job is more interesting. The costs of the transition are being borne almost entirely by the generation that arrives after her — the 22‑year‑olds whose resumes no longer open doors that were once guaranteed.

What This Signals

The DocAI story is not primarily a story about JPMorgan, or about investment banking, or even about AI. It is a story about the collapse of the apprenticeship model that has structured professional entry in law, finance, consulting, and accounting for more than a century. That model — hire smart 22‑year‑olds, work them relentlessly for two to four years while they learn the craft, promote the survivors, and repeat — was never efficient. It was effective. It produced generations of deeply skilled professionals who had internalized the patterns, the judgment, and the instincts that could not be taught in a classroom.

AI breaks that model not by replacing the senior professional, but by eliminating the junior role where the senior professional was forged. The machine can do the work of the 2 a.m. analyst faster, cheaper, and more accurately. What it cannot do — yet — is develop the judgment that the analyst would have acquired by doing that work for two years. The judgment gap is the structural risk embedded in every DocAI deployment, and it will not be visible for a decade. It will surface when the managing directors of 2026 retire, and the industry discovers that the generation expected to replace them has never built a model from scratch, never discovered a hidden liability at 3 a.m., and never learned the pattern‑recognition instincts that come only from doing work a machine now performs.

JPMorgan's platoon experiment is the industry's most serious attempt to solve the judgment gap. Whether it works will determine not just the future of the bank, but the future of professional expertise in an age when the entry‑level work that created expertise has been automated away. The intern that never sleeps is efficient. The question is whether the managing director it never trains will be any good.