AI in Investment Banking: The Systems Five Big Banks Built for Themselves

Since 2023, JPMorgan, Goldman Sachs, Morgan Stanley and Citi have each put an AI assistant of their own making in front of large parts of their workforce, and Bank of America has extended an assistant platform that predates them all. That is what AI in investment banking mostly means: a chat window the bank controls, with outside AI models behind it and the bank deciding what data they can see.

For a junior banker, the change is the same at all five. The first version of the research note, the peer list or the pitch page now often starts as machine output, and what the analyst adds is the request that produced it and the check that clears it.

What AI do the big banks use?

Each bank has one system that most of its people meet first. Here they are side by side, in each bank's own words and dates.

How is a junior banker's job changing?

The systems differ, but what they ask of the people at the bottom of the organization does not. Each of the five banks has described the same shift in its own words: the machine produces the first version, and the person's work moves to specifying it and signing it off.

Read together, the five lines describe one move: from producing the work to directing and vetting it.

What a pitch request looks like now

Put the tools the banks have described side by side, and a first-year's pitch request runs roughly like this:

1. The research pull. A research assistant, such as the one Morgan Stanley built for its bankers and traders, answers from the firm's own reports and links each one it drew on.
2. The peer list. A reasoning model proposes comparable companies, fetches their filings and ranks them, in about ten minutes by Goldman's account.
3. The pages. Bots draft the company overview and the working group list, and generators lay out pitch pages inside the bank's PowerPoint ribbon, as at Bank of America.
4. The check. The analyst opens every linked report, tests the peers against the brief, checks each figure and its date, and decides what the client actually needs to see.

That fourth step is the one a candidate can rehearse before any offer, on tools anyone can open.

Why don't banks just give staff ChatGPT?

The banks that have explained the choice wanted the chat window without giving up control of three things: the data, the record of what was asked, and the model itself.

A public chatbot, as the banks saw it in 2023

• Answers drawn from the open internet
• Prompts and pasted data leave the bank
• No compliance record of who asked what
• A single vendor model, nothing else

A bank's own AI

• Answers drawn from research and data the bank controls
• Client data never used to train an outside model
• Prompts logged and screened for compliance
• A model the bank picks, and can replace

JPMorgan built its own so that its data would never train an outside company's model. Morgan Stanley's partner agreed to keep none of the data the firm sends it, and Goldman runs controls on what staff put into their prompts.

The fourth line is where the banks part company. Agreeing that the bank chooses the model still leaves two questions: how many labs to choose from, and whether the plumbing came before the models or after them.

Three routes to a bank's own AI

Three answers to those two questions cover all five banks.

A front door onto several labs

JPMorgan, Goldman Sachs and Citi each put a tool carrying the bank's own name in front of more than one outside lab's models. JPMorgan designed LLM Suite from the start so it could swap providers as they improve. Goldman runs automated tests that weigh accuracy against cost to decide which model to favor for each kind of question.

Citi built on Google's Vertex AI platform under an October 2024 agreement, and David Griffiths, its chief technology officer, has said he wanted options across several AI providers. The appeal of the route is that a better model can go in underneath while staff keep the tool they already know.

One lab, a tool for each job

Morgan Stanley went deep with a single partner instead: OpenAI, which the firm's co-presidents called its exclusive partner in a March 2024 memo. It built a separately named tool for each kind of work, from advisors' questions to legacy code. The one Anthropic model on record there is the Claude Mythos Preview, for cyber defense; Ted Pick, the chief executive, told analysts in April 2026 that the firm was permissioned on it.

A platform built before the chatbots

Bank of America started earlier, and from the customer's side. It launched Erica in its mobile app in 2018 on a platform designed to change models over time, and has put later assistants for staff and clients on the same base, an approach it calls "build once, reuse." When large language models arrived, it rebuilt that base for them in 2025 rather than starting over.

What the route means for a new joiner

The route decides what a new joiner has to learn. At a front-door bank, one general tool covers most tasks. At Morgan Stanley, the skill is knowing which tool belongs to which desk. At Bank of America, new tools arrive team by team on a shared base, so what a junior gets depends on the business they join.

Which AI tools are built for investment bankers?

The firmwide assistants are built for every kind of employee. A smaller set of tools is aimed at the deal team itself:

• Morgan Stanley: AskResearchGPT. Since 2024 it has served investment banking, sales and trading, and research staff, and a patented workflow turns its answer into an email draft in one click.
• Goldman Sachs: Banker Copilot. Built for some of the firm's investment bankers, it helps them compile research through a conversational interface.
• Bank of America: a GenAI Suite for bankers. At its November 2025 investor day the bank showed a proprietary chatbot connected to its research and market commentary, working beside Microsoft Copilot and next to the overview bots and page generators from the pitch example above.
• Citi: the leading models, plus specialist pilots. Vis Raghavan, its head of banking, said in May 2026 that bankers work with "Citi AI and the leading LLMs" and pilots of specialist banking apps.
• JPMorgan: the firmwide platform itself. Its bankers work in the same assistant as the rest of the firm, which has added PowerPoint generation and Excel analysis modules on top of its base.

What else are the banks rebuilding with AI?

The desk tools are one track. The other is quieter: whole processes redesigned around AI, and the lists the banks give lean toward operations and control work rather than the deal team.

• Goldman Sachs launched One Goldman Sachs 3.0, an operating model it describes as propelled by AI, starting with six workstreams that include client onboarding and KYC, lending and regulatory reporting.
• Citi started with just over 50 of its largest and most complex processes, ranging from KYC to loan underwriting, according to its January 2026 earnings call.
• Bank of America's chief technology and information officer, Hari Gopalkrishnan, said in November 2025 that the return comes from rebuilding whole client journeys, each spanning dozens of processes and thousands of employees, rather than from saving minutes on single tasks.
• Morgan Stanley said on its January 2026 call that a documentation check once run by two teams checking each other now pairs one human team with one AI team.

What comes after the chat window?

Rebuilding processes at that scale needs software that carries a task through several steps on its own, and all five banks have named that kind of agent as the next step:

• JPMorgan had begun deploying agentic AI for complex, multistep tasks by September 2025, according to an internal road map CNBC reported.
• Goldman Sachs made Devin, Cognition's autonomous coder, broadly available to its engineers by February 2026, when Anthropic engineers had spent six months at the bank building Claude-based agents for trade accounting and client onboarding.
• Citi introduced Arc, its platform for building agents, in April 2026, starting with its own developers on well-defined use cases.
• Morgan Stanley's Pick described "super agents" for the wealth business in April 2026, and a client agent inside the equities electronic trading platform.
• Bank of America's Moynihan said in July 2026 that its tools ran from productivity aids to more advanced agentic workflows.

For a junior, agents stretch the same shift further. More of a task runs without a prompt at each step, and the person's part is setting the goal and answering for the result.

What does bank AI mean for engineers?

Each of the five has been concrete about what AI means for its developers:

• Goldman Sachs gave its 12,000 developers AI coding assistants and in July 2025 put the result at 20% more productive on average.
• Bank of America's coding assistants serve 18,000 developers, with a 20% productivity lift in the parts of the development cycle it has focused on, by its November 2025 account.
• Citi's AI-driven code reviews passed one million in 2025, and Jane Fraser, its chief executive, credits them with about 100,000 hours of developer capacity a week.
• Morgan Stanley built DevGen.AI to turn legacy code into plain-English specifications; in its first five months it worked through nine million lines and saved an estimated 280,000 hours.
• JPMorgan's engineering postings for its platform team name Azure OpenAI and AWS Bedrock, agent-to-agent communication, the Model Context Protocol and agentic orchestrators.

Two of those banks point at the same pair of standards. Alongside JPMorgan's postings, Goldman's chief information officer named MCP and A2A in a July 2025 essay as protocols emerging to standardize how an agent communicates with people, data sources and other agents. For an engineer heading to a bank AI team, that overlap is a useful map of what to learn first.

How does each bank use AI?

The table at the top is the short version. Each bank's own guide goes further into what it built, the numbers behind it and what it asks of a new joiner.

JPMorgan

The guide to JPMorgan's LLM Suite follows the bank from its decision to block ChatGPT to a platform in front of nearly the whole firm, and sets out every user figure JPMorgan has given with what each one measures. It separates LLM Suite from the narrower tools candidates tend to confuse with it, such as Connect Coach and IndexGPT, and covers what the bank's executives have said about AI and jobs.

Goldman Sachs

The guide to Goldman's GS AI Assistant explains why the bank chose a menu of models over a single lab, how the assistant went from a pilot to the whole firm in about five months, and what the tools around it do, from Banker Copilot to Devin. It is also where Marco Argenti's case for describing, delegating and supervising is set out in full.

Morgan Stanley

The guide to Morgan Stanley's AI takes the tool family one job at a time, from the advisors' Assistant and the Debrief note-taker to AskResearchGPT, with a seat-by-seat table of what each takes off a junior's desk. It also explains evals, the routine of grading a model against expert answers before any use case ships, which is a habit an outsider can learn without access to any tool.

Citi and Bank of America

The guide to Citi and Bank of America sets the two banks' opposite routes side by side: Citi's family of tools on outside labs' models and what it plans next, and Bank of America's reuse of the platform behind Erica. Its seat-by-seat table covers what each bank's tools do across investment banking, markets, payments, client service and wealth.

Will bank AI shrink the analyst class?

Of everything above, what candidates ask about first is the number of analyst seats, and the banks' statements on the record describe trims in places rather than a vanished class. What each bank has said about its starting classes, what the published hiring numbers show and where the much bigger figures in circulation come from are worked through in the analysis of whether AI will replace investment banking analysts.

How should a candidate talk about bank AI?

No one outside a bank can open these tools, so an interview can test understanding but not experience. Three habits cover all five banks:

1. Match the answer to the bank. Know which route the bank you are meeting took and one reason it gave; an answer built for JPMorgan's front door does not fit Morgan Stanley's single partner.
2. Quote a figure with its date and what it counts. Access, daily use, prompts and active users are different numbers, and naming the one you mean shows you read the bank's own words.
3. Bring one example of the check. Run a real task through a public chatbot such as ChatGPT, check every figure against the source, and keep a note of one mistake you caught.

The interview readiness tool turns those habits into practice answers, and the guides cover the software and the roles beyond the banks.

AI in investment banking, in short

AI in investment banking is not one product. It is five banks' own systems, built three ways: a front door onto several labs' models at JPMorgan, Goldman Sachs and Citi, a single partner with a tool for each job at Morgan Stanley, and a platform that predates them all at Bank of America.

What those systems ask of a first-year is the same everywhere. The machine's draft arrives fast; the request that shaped it and the check that clears it belong to the analyst, and both can be practiced long before an offer.