AI in investment banking: what Goldman, Morgan Stanley, JPMorgan and the rest actually deploy.

What does AI in investment banking actually mean?

On the record of the investment banks profiled here it means five things: research assistants and agents that draft reports and answer questions over a research library (Morgan Stanley, RBC, TD Securities, UBS, Deutsche Bank), banker-materials tools that build pitchbooks and client decks from prior work (Deutsche Bank's Farsight, RBC's AidenBanker), markets-operations agents for email triage, document processing, reconciliation and surveillance (Barclays, Citi, Morgan Stanley, Deutsche Bank), pricing and quoting models for FX, bonds and derivatives (ING, BBVA, Santander, Société Générale, Barclays) and coding assistants and agents for the engineering floor (Goldman, Citi, JPMorgan, Bank of America, Morgan Stanley). 22 of the 140 banks on this site record one of those, 21 in production.

Strip the vendor language away and the investment-banking record on this site describes five kinds of system. The first is the research assistant: a retrieval layer over the bank's own research library and market data, with a language model drafting or answering on top. Morgan Stanley's AskResearchGPT (October 2024, 70,000-plus reports), TD Securities' OpenAI assistant with text-to-SQL for sales and trading (about 1,000 users in October 2025, planned for all of global markets, no fine-tuning needed), RBC's AidenResearch and QuickTakes (reports drafted up to 95% faster), Deutsche Bank's dbLumina on Vertex AI with inline citations and audit logging, and UBS's Global Research team, which won Celent's 2026 Model Sell Side award for generative-AI automation, are all this shape. The second is the banker-materials tool: Deutsche Bank's Farsight learns each team's prior work, standards and formats and keeps the data inside the bank; RBC's AidenBanker handles pitchbook assembly and meeting preparation.

The third is markets operations, which is where the agentic systems live: Barclays classifies, enriches and routes about 120,000 Global Markets emails a day with Claude; Citi has processed 4,400-plus markets documents with AI; Morgan Stanley's FIXR runs three cooperating agents across finance, risk, operations and trade capture; Société Générale's My Cases reads post-trade client emails and handles margin calls twice as fast; Deutsche Bank's voice surveillance transcribes at over 90% accuracy with a tenfold wider lexicon at 65% lower cost. The fourth is pricing and quoting, the oldest machine-learning use on the desk: ING's reinforcement-learning FX pricer, BBVA's algorithmic bond, swap and FX quoting, Santander's Automated Pricer, Société Générale's Trading Assistant for requests for quotation (designed in 2017), Barclays' BARXBot. The fifth is engineering: Goldman's Devin fleet, Citi's 40,000 developers on Devin, JPMorgan's mainframe refactoring, Bank of America's 19,000 developers, Morgan Stanley's DevGen.AI. Advisory judgement, valuation and negotiation appear in none of the 35 records.

Which banks deploy AI in investment banking, and for what?

32 investment-banking, markets or research records are in production across 21 banks, led by Goldman Sachs, Morgan Stanley, JPMorgan Chase, Citi, Bank of America, RBC, Deutsche Bank, Barclays, UBS, BNP Paribas, Société Générale, BBVA, Santander and ING. The table lists every record with the bank's own name for the tool, its status, the date of the evidence and the source.

The table lists every bank-profile record that describes an investment-banking, markets or research use of AI, strongest status first and largest bank first within a status; retail broking, wealth, custody and asset-management records are left out and appear on the retail and corporate pages. “In production” means the bank or a named vendor says it is live; the date is the earliest sourced event on the bank's timeline for the record. The US ranking by whole disclosed record is on the AI disclosure index.

Reading across the rows: the universal banks with a large markets business (Goldman, Morgan Stanley, JPMorgan, Citi, Bank of America, Deutsche Bank, Barclays, BNP Paribas, UBS, RBC) disclose the firm-wide platform first and the desk tools second; the European banks with strong markets franchises but smaller US footprints (Société Générale, BBVA, Santander, ING) disclose the pricing and operations models in most detail, often with before-and-after numbers; and the research floor is where generative AI reached production earliest and most visibly, with 15 records on research alone.

What do Goldman, Morgan Stanley, JPMorgan, Citi and Bank of America actually run?

Goldman Sachs: GS AI Assistant for all 46,000 employees, coding agents beside 12,000 developers, Marquee AI in internal pilot, a Firmwide AI Risk Committee. Morgan Stanley: AskResearchGPT over 70,000-plus reports, FIXR agents saving about 1,500 controller hours a week, DevGen.AI with about 16 million lines modernised, gen-AI access for about 98% of 80,000 staff. JPMorgan Chase: LLM Suite with 150,000 weekly users and 175-plus use cases in production in the commercial and investment bank. Citi: markets document processing freeing 1,700-plus hours a month, Devin for 40,000 developers, the Arc agent platform. Bank of America: the TED issuance score in CashPro, coding assistance for about 19,000 developers, the Orchestra model-routing layer.

The five largest US investment banks by the breadth of what they disclose, with the bank's own name for each system, the job it does, its status and the figure the bank gives. Every row is on the bank's profile with its source; the profiles also hold the numbers that cut the other way, such as JPMorgan's CFO putting annual AI spend closer to $700 million and calling the figure “a little stale”.

Two differences between the five are worth knowing. Goldman and Citi have put autonomous coding agents on production work (Devin at both); JPMorgan, Morgan Stanley and Bank of America describe assistants and in-house platforms rather than agent fleets, although JPMorgan has agents in production for software testing, private-banking research screening and payments and plans long-running agents for 2026. And on model choice, Goldman and Bank of America both run a routing layer that lets the firm pick among GPT, Gemini, Claude and open-weight models per task, while Morgan Stanley's named tools are built on OpenAI models; AI agents in banking follows the agent side across all 140 banks.

What results do investment banks report from AI?

RBC's research coverage per analyst rose from 14.5 names in 2024 to over 17.5 in October 2026; Morgan Stanley answers client research inquiries in a tenth of the time and saves about 1,500 controller hours a week with FIXR; Citi's markets document processing created 1,700-plus hours a month of oversight capacity; Barclays routes about 120,000 Global Markets emails a day with Claude; Bank of America says a banker can cover 50 clients instead of 15; Goldman's CIO cites up to 20% developer efficiency; MUFG staff report a 20 to 30% workload reduction on selected research tasks; Goldman's Q2 2026 revenue of $20.3 billion, up 39%, came on what it called an AI-driven markets boom.

The banks' own figures as reported, with date and source. They measure different things (coverage per analyst, hours saved, documents processed, users, revenue) and should not be averaged; the strongest are the ones with a stated before-state and denominator (RBC's 14.5 to over 17.5 names per analyst, Morgan Stanley's tenth of the time), the weakest are productivity percentages from a pilot. Goldman's revenue figure is in the table because the bank attributed the quarter to an AI-driven markets boom, not because AI tools produced it.

Will AI take over investment banking?

The banks say no and act like partly. No bank on this site discloses an AI system that originates, values or negotiates a deal; what they disclose is output per person rising (RBC's coverage per analyst, Bank of America's 50 clients per banker, Morgan Stanley's 1,500 controller hours a week) and group headcount falling by attrition (Wells Fargo down 8% year on year, Citi at 219,000, BNY down 7%, Standard Chartered's more than 7,000 roles affected, HSBC's reported 20,000). JPMorgan's CEO says the bank has displaced people with AI and redeploys them; Wells Fargo's CEO says automation will mean tens of thousands of fewer positions; Goldman's research puts US displacement at over 9% of workers over ten years. The honest answer is that junior analyst work is being automated and the number of analysts a bank needs per dollar of revenue is falling; the banker is not being replaced.

The question gets 100 searches a month and the honest material to answer it with is thin, so this section uses only what banks have said in filings, earnings calls and interviews on the record. Three kinds of number exist. Output per person: RBC's analysts cover more names, Bank of America's bankers cover more clients, Morgan Stanley's controllers do less reconciliation by hand. Group headcount: Wells Fargo is down 8% year on year after 24 consecutive quarterly declines, Citi is at 219,000 after 5,000 fewer in a quarter, BNY is down 7%, Standard Chartered has more than 7,000 roles affected by Reuters' count, HSBC around 20,000 by Bloomberg's; none of these is an investment-banking figure and most are attributed to restructuring and attrition rather than to AI. Forecasts: Goldman's own research raised its estimate of US workers displaced by AI over ten years to over 9%, about 15 million, across all industries.

What the chief executives say, in their words. The pattern is consistent: Dimon says people have been displaced and redeployed; Scharf says automation means tens of thousands of fewer positions, through attrition where possible; Fraser says it will take some jobs away; Elhedery says generative AI will destroy certain jobs and create new ones; Bank of America's markets co-president says the biggest risk to implementation is people fearing they will be replaced. Against that, Morgan Stanley promoted 184 new managing directors in its 2026 class and RBC is growing its Vancouver hub to 1,000 jobs by 2029. The “take over” framing misses what the records show: the analyst's first two years of work (model updates, comps, first drafts, meeting prep) is the part being automated, and the number of analysts per dollar of revenue is the number to watch.

What do the SEC, FINRA, FCA and the other market regulators say?

That the existing rules apply and that they are watching. The SEC withdrew its predictive-data-analytics conflicts proposal in 2025, names AI in its FY2026 exam priorities and has fined advisers for AI washing; FINRA's Notice 24-09 and 2026 oversight report map supervision, communications and third-party rules onto generative AI; the CFTC's Staff Advisory 24-17 does the same for derivatives; the FCA's 2024 AI Update rests on the Consumer Duty, the senior managers regime and PRA SS1/23; ESMA's 2024 statement applies MiFID II; IOSCO's 2026 supervisory toolkit sets shared expectations for AI in capital markets; the Fed's SR 26-2 keeps machine-learning pricing, hedging and surveillance models inside model risk management.

No securities regulator has written an AI rule for investment banks. What exists is a set of statements that the existing rules apply (best interest and fiduciary duties, supervision, communications with the public, recordkeeping, market abuse, the Consumer Duty, MiFID II), enforcement against firms that overstate their AI, examination priorities that name AI, and the model-risk and third-party guidance that reaches any machine-learning pricing, hedging or surveillance model. The documents below are the ones an investment bank's AI programme is most often examined against; the tracker holds 238 documents from 41 authorities, and the full markets set is on the trading and markets page.

Three things to take from the table. First, the SEC's withdrawal of the predictive-data-analytics proposal in June 2025 removed the one rule that would have been specific to AI in broker-dealers and advisers; what replaced it is enforcement, with the Delphia and Global Predictions settlements as the template, and the Investor Advisory Committee's December 2025 recommendation on AI disclosure. Second, FINRA's Notice 24-09 is the document a US sales-and-trading compliance team actually works from, because it maps generative AI onto the supervision and communications rules that already govern the desk. Third, the UK and EU route is prudential and conduct rather than AI-specific: PRA SS1/23 for the models, the Consumer Duty and MiFID II for the client, and the Bank of England's Financial Policy Committee for the system-wide view of AI in trading. The model-risk side is on model risk management; the governance side on AI governance in banking.

Which vendors and models do investment banks use?

Frontier models behind in-house platforms: OpenAI (Morgan Stanley, TD Securities), Anthropic's Claude (Barclays, Goldman's platform), Google (Deutsche Bank, Citi) and Cognition's Devin (Goldman, Citi). 13 published vendors in this site's directory match that stack; the ones most often named by the banks here are Microsoft Copilot (16 banks), ChatGPT Enterprise (8 banks), Claude (8 banks), Gemini Enterprise (6 banks).

Investment banks buy models, not applications. The platforms are in-house (GS AI Platform, LLM Suite, Stylus, Aiden, Farsight, dbLumina, FIXR, DevGen.AI) and the models underneath are rented: OpenAI at Morgan Stanley and TD Securities, Anthropic at Barclays and inside Goldman's platform, Google at Deutsche Bank and Citi, Cognition's Devin at Goldman and Citi. The table lists the published vendors in this site's AI tools directory that match that stack, with the number of banks in the US set that name them on the record. A bank naming a vendor is a disclosure, not a market share.

The exceptions prove the rule: BNP Paribas keeps its corporate-and-investment-bank assistant, LLM@CIB, available to more than 65,000 employees and is integrating Gemini models; MUFG staff built more than 1,800 custom GPTs in four months after OpenAI-led workshops; UBS uses OpenAI and Synthesia models to turn analyst notes into avatar videos, with more than 35 of about 720 analysts opted in and output capped at about 1,000 videos a year by studio capacity. For what the same banks run in treasury, KYC and fraud, follow the sibling pages.