The honest state
Finance was always going to be early on LLMs because the work is text-heavy, document-heavy, and the marginal labor cost of a senior analyst is hundreds of dollars per hour. As of 2026, every G-SIB has either built an internal LLM platform or signed a major frontier-model contract. The deployed surface is wider than the press tells.
Who is shipping (named)
Bloomberg LP published BloombergGPT (50B-parameter finance-domain LM, March 2023) but quietly pivoted to fine-tuned commercial models for production document tasks. JPMorgan filed a USPTO trademark for "IndexGPT" in May 2023 — a planned investment-advisory LLM — and runs "LLM Suite," a firm-wide ChatGPT-style internal tool deployed to 200,000+ employees by mid-2024 (per public CIO statements at the Q2 2024 earnings call). Goldman Sachs deploys an internal coding assistant across the developer base. Morgan Stanley partnered with OpenAI in 2023 for a wealth-advisor knowledge agent that searches the firm's 100,000+ research documents, replacing a manual archive workflow. Mastercard Decision Intelligence Pro (October 2023) uses a custom transformer model trained on transaction-graph data for real-time fraud scoring, claimed to roughly double detection rates per Mastercard's public release.
Asset managers: BlackRock Aladdin integrated GPT-4 in 2023 for portfolio-management workflow; BNY Mellon has Eliza, an internal AI assistant. Citadel + Two Sigma + Jane Street are visibly hiring LLM researchers, less visibly deploying. Quant trading remains dominated by classical models — LLMs aren't a fit for microsecond-latency execution.
The five real use-cases that work
(1) Document extraction — pulling structured data from 10-K filings, prospectuses, loan documents, derivatives contracts. Saves thousands of analyst-hours per year per firm. (2) Research summarization + RAG over internal archive — Morgan Stanley's wealth-advisor agent is the canonical example. (3) Coding copilots for the dev team — every major bank deploys this. (4) KYC / AML adverse-media screening — LLM-powered name screening against news and watchlist databases (Quantexa, ComplyAdvantage, Featurespace are the named vendors). (5) Customer service — chat agents for retail banking, deployed by every neobank and most majors. Stripe's GPT-4 customer-support deployment was the early public example.
Receipts
- Wu, Irsoy, Lu et al. BloombergGPT: A Large Language Model for Finance. arXiv:2303.17564, March 2023.
- JPMorgan Chase USPTO trademark application 97817313 (May 2023) — "IndexGPT."
- JPMorgan Q2 2024 earnings call — Jamie Dimon and Lori Beer (CIO) on LLM Suite deployment to 200,000 employees.
- Morgan Stanley + OpenAI partnership announcement, March 2023; ongoing wealth-management AI deployment.
- Mastercard, Mastercard Launches Decision Intelligence Pro, October 2023 — GenAI-powered fraud detection.
- SEC FinTech Forum + SEC AI Risk Disclosure Rule (December 2023) — disclosure obligations for AI in investment advisory.
- FINRA Regulatory Notice 24-09 (June 2024) — guidance on use of generative AI by member firms.
What the sector still cannot do
Reliable autonomous trading from LLM signals — published academic results show LLMs can extract sentiment from news, but no public benchmark demonstrates risk-adjusted returns above classical baselines. Direct customer financial advice without human review — SEC + FINRA expect a registered human in the loop. Loan underwriting without explainability — Equal Credit Opportunity Act (ECOA) requires adverse-action explanations; pure black-box LLM denial is illegal in the US.
Regulatory + compliance reality
SEC AI Risk Disclosure Rule, FINRA Notice 24-09, OCC Bulletin 2023-15 (third-party risk including AI), CFPB statements on adverse-action notices. EU's AI Act classifies credit scoring as "high-risk" — meaning conformity assessment, documentation, human oversight, post-market monitoring. The compliance overhead is real and is the main reason finance is methodical rather than fast.