What proprietary AI has Goldman Sachs built in-house?
Goldman Sachs has built its own AI platform and assistants but not its own foundation model. CIO Marco Argenti said in October 2023 that the firm had no plans to build a large language model from scratch and would instead fine-tune existing models and use retrieval-augmented generation on its own data. What it built is the layer around those models: the GS AI Platform, a centralized internal platform with technical, legal and compliance checks, content filtering, logging of every interaction and a single chat interface the firm developed, which in June 2024 fronted Microsoft-hosted GPT-3.5 and GPT-4, Google Gemini and Meta Llama. On it runs the GS AI Assistant ('we built and we're deploying at scale our GS AI assistant', Argenti said in March 2025), which reached about 10,000 staff in January 2025 and all 46,000 employees by July 2025, with more than a million prompts a month. Argenti keeps the models 'plug-and-play' so the firm can pick a vendor by task. Other in-house systems: Marquee AI, which attributes each statement to a line in a source (internal only as of August 2026), the AlphaAI investing platform (July 2026), and Legend, a data platform built internally over 10 years before it was open-sourced in 2020. Agents are bought or co-built: Cognition's Devin from July 2025 and Claude agents built with embedded Anthropic engineers (February 2026). The work sits under Daniel Marcu, Global Head of AI Engineering and Science, with named heads of GS AI platform engineering and of AI Research.[1][2][3][4][5][6][7][8][9][10][11][12][13]
Data as of Oct 22, 2026 · Goldman profile updated Oct 13, 2026 · 27 sources cited on this page · 140 banks profiled on this site · Markdown
Which AI platforms and assistants has Goldman Sachs built, and what do they run on?
'Built on' names the models behind each system as the firm or the press reported them. 'Since' is the earliest public date.
| Platform or model | What it does | Built on | Since | Status |
|---|---|---|---|---|
| GS AI Platform ('GS AI') | Centralized internal gateway: technical, legal and compliance checks; a front-end server that filters inappropriate content; every interaction logged; one chat interface the firm developed; designed with guardrails 'to experiment safely at higher speed' | Third-party models: Microsoft-hosted GPT-3.5 and GPT-4, Google Gemini, Meta Llama (June 2024) | In place by October 2023 | In production[1][3][2] |
| GS AI Assistant | 'Your desktop for AI': summarizes documents, drafts emails, analyzes data, translates code; meant to give 'very GS-specific answers' and, over three to five years, reason more like a Goldman employee | OpenAI, Gemini and Llama models fed with Goldman data (January 2025), Anthropic, Mistral and Cohere under evaluation; GPT, Gemini and Claude choices by July 2025 | Rolled out in 2024 | All 46,000 employees; more than a million prompts a month (July 2025)[4][5][7][6] |
| Grounding layer | Fine-tuning plus retrieval-augmented generation on firm data; output cross-checked against sources to reduce hallucinations and prompt-injection risk | Existing third-party models, not a model trained from scratch | October 2023 | In use[1][4] |
| GitHub Copilot on GS AI | Integrated into the software development life cycle; generated code passes the same quality checks as other code; Gemini Code Assist also offered | Microsoft GitHub Copilot; Google Gemini Code Assist | June 2024 | 9,000-plus developers, about 80% of engineers, at launch[3][7] |
| Marquee AI | Breaks a question into research topics and attributes every statement to a specific line in a source, through a grounding framework the Marquee team built | Frontier models; vendor not disclosed | 2026 | Internal only (August 2026)[8] |
| AlphaAI | Asset-management investing platform built to find dispersion within sectors, drawing on public and private markets data and over 100 scaled AI use cases in portfolio companies | Not disclosed | July 2026 | Launched; chaired by Lou D'Ambrosio[9] |
| Legend | Data-modeling and governance platform; staff search firm files in natural language | Goldman's own code, built internally over 10 years; open source through FINOS | 2020 (public) | In production[7][10][14] |
| Claude agents | Trade and transaction accounting; client vetting and onboarding | Anthropic Claude, co-built with Anthropic engineers embedded for six months | February 2026 | Pilot[12] |
| Devin | Autonomous software engineer in developer teams | Cognition (bought; Goldman holds no stake) | July 2025 | Up to 20% efficiency gains reported by the CIO (June 2026)[11][15] |
Which teams build Goldman Sachs' in-house AI?
| Person or team | Role | Since |
|---|---|---|
| Marco Argenti, CIO | Owns the platform strategy; joined as partner and co-CIO in 2019 after six years as a vice president of technology at Amazon Web Services; sole CIO from 2022 | 2019[7] |
| Daniel Marcu | Global Head of AI Engineering and Science, hired from Amazon 'to help develop and refine artificial intelligence platforms and products', per the memo | January 2025 (memo); March 2025 (joined)[13][16] |
| Rahul Sharma | Head of GS AI platform engineering, reporting to Marcu | January 2025[13] |
| Bing Xiang | Head of AI Research within Engineering, reporting to Marcu | January 2025[13] |
| Archana Vemulapalli | Global Head of AI Product Management and Strategic Relations; leads engagement with frontier AI labs | April 2026[17] |
| Chris Churchman | Head of Marquee and co-chair of the Global Banking & Markets AI working group; leads Marquee AI | August 2026 (on the record)[8] |
| Engineering | Roughly one in four of 46,000 employees is an engineer; Argenti puts his engineers at 12,000 | 2025-2026[7][14] |
| AI steering group | With risk and control teams, decides which of dozens of AI proposals to build and how | March 2025[7] |
Why does Goldman Sachs build the platform but buy the models?
Argenti set the line in October 2023: 'I might be completely wrong, but I don't believe at this point... that it is necessary to start from scratch.' Instead the firm fine-tunes existing models and uses retrieval-augmented generation, because 'a lot of stuff out there is generic, but there is data that we have, that is the most important thing.' Fortune reported the same position in June 2024: integrate GPT-3.5, GPT-4, Gemini and Llama into Goldman's own platform rather than build a model.[1][3]
The platform is multi-vendor by design. Argenti told Fortune in March 2025 that he does not want to rely on one vendor, that one model may be better for coding and another for reasoning, and that the firm weighs how easy each model is to modify and how expensive it is to run: 'All of those considerations got us to the point where we want to continue to plug-and-play with those models.'[7]
Chris Churchman, who leads Marquee AI, put the build rule as 'don't bet against the model': build what a model cannot learn, such as the firm's entitlements, mandates and the way its data sets connect, because reasoning scaffolds and retrieval pipelines built around older models' limits were made 'totally pointless' by reasoning models, longer context windows, function calling and MCP.[8]
For agents Goldman buys and co-develops rather than builds alone: Devin comes from Cognition, in which it holds no stake, and the accounting and onboarding agents were built on Claude with Anthropic engineers working inside the bank.[11][12]
What has Goldman Sachs not disclosed about its proprietary AI?
No model trained by Goldman from scratch is on the record, and the firm's stated position since 2023 is that it does not need one. It has not said which fine-tuned models run in production or what share of the GS AI Assistant's million-plus monthly prompts goes to each vendor.[1][6]
Headcount for GS AI platform engineering and AI Research is not published; the January 2025 memo names their heads only. No AI platform budget is disclosed. Goldman holds AI patents granted in 2026 on embedding-model fine-tuning and embedding-free retrieval-augmented generation, but has not said which in-house system uses them.[13][18][19]
What does the Goldman profile record on this topic?
The rows below are the entries of the Goldman Sachs AI strategy profile that bear on this page, as of Oct 13, 2026; the profile holds the full record of 37 dated moves, 13 use cases and 62 sources.
| Use case | What the record shows | Status | Source |
|---|---|---|---|
| GS AI Assistant | Firm-wide conversational assistant on the GS AI Platform. | In production | [6][20] |
| Devin and coding agents | Autonomous software engineer piloted in developer teams. | Pilot | [20] |
| Legend natural-language research | Staff search the bank's files in natural language without knowing where a document sits. | In production | [14] |
| Coding agents on production work | Devin is broadly available to engineers and hundreds of agents work beside 12,000 developers; Argenti cites up to 20% efficiency gains and says developers now set specs and review output. | In production | [12][21][15] |
| Prompt controls on the GS AI Platform | Controls check what information staff enter in prompts and flag it; the platform lets users choose among current GPT, Gemini and Claude models. | In production | [6] |
| Marquee AI for institutional clients | Natural-language assistant on the Marquee platform that breaks a question into research topics and attributes statements to source lines; available internally only as of August 2026. | Pilot | [8] |
| AlphaAI for investing | Asset-management platform for finding dispersion within sectors, drawing on public and private markets data. | Rolling out | [9] |
| Frontier-lab and AI-application vendor relations | A Partner-level head of AI product management leads strategic engagement with frontier labs and AI application providers, alongside embedded Anthropic engineers and the Devin deployment. | In production | [17][12] |
| Date | Move | Detail | Source |
|---|---|---|---|
| Jul 30, 2026 | AlphaAI launched for AI-led investing | Lou D'Ambrosio chairs AI for Asset Management; the memo cites over 100 scaled AI use cases inside portfolio companies. | [9] |
| May 8, 2026 | Argenti measures idea-to-production speed, not AI usage | The CIO says tracking individual use is the wrong metric for his 12,000 engineers; Legend lets staff search firm files in natural language. | [14] |
| May 4, 2026 | Founding partner of an Anthropic enterprise-services firm | With Anthropic, Blackstone and Hellman & Friedman; the firm, later named Ode, embeds Anthropic engineers to bring Claude into companies' operations. | [22][23] |
| Apr 6, 2026 | Archana Vemulapalli hired to lead AI product management | Partner and Global Head of AI Product Management and Strategic Relations, from AMD; she also leads engagement with frontier AI labs and application providers. | [17] |
| Mar 20, 2026 | Solomon's letter flags AI-exposed private credit | 'Exposure to software companies that may be adversely affected by AI' named as a reminder that the credit cycle has not been repealed. | [24] |
| Feb 6, 2026 | Anthropic engineers embedded to build agents | Six months of co-development in trade accounting and client onboarding; Goldman confirms the report. | [12][25] |
| Jul 25, 2025 | GS AI Assistant reaches the whole workforce | 46,000 employees; a million prompts a month; Devin agents joining developer teams. | [6] |
| Jul 11, 2025 | Devin pilot announced: 'hundreds of Devins', maybe thousands | Argenti expects Cognition's agent to raise developer productivity three to four times over earlier AI tools; Goldman is the first major bank to use it, per Cognition, and holds no stake in Cognition. | [11] |
| Jun 23, 2025 | Argenti memo: GS AI Assistant launches firmwide | About 10,000 employees already use it; it will help with summarising documents, drafting content and data analysis. | [26] |
| Mar 17, 2025 | Daniel Marcu joins as Global Head of AI Engineering and Science | A Partner in Engineering, hired from Amazon's AGI organisation, with more than 35 US patents. | [16] |
| Jan 21, 2025 | GS AI Assistant reaches about 10,000 employees | Argenti aims to cover all knowledge workers within the year; answers draw on OpenAI, Google Gemini and Meta Llama models, with Anthropic, Mistral and Cohere under evaluation. | [27] |
| Jun 27, 2024 | GS AI Platform: one entry point, several model vendors | Argenti says the first generative-AI tool, for code generation, reaches thousands of developers by month end; the platform uses Microsoft-hosted GPT-3.5 and GPT-4, Google Gemini and Meta Llama, and the bank had banned in-house ChatGPT. | [2] |
Does Goldman Sachs build its own AI models?
Not from scratch. CIO Marco Argenti said in October 2023 that he did not believe it was necessary to start from scratch; the firm fine-tunes existing models and uses retrieval-augmented generation on its own data. What Goldman builds is the GS AI Platform, the GS AI Assistant, Marquee AI, AlphaAI and the Legend data platform, all running on or around third-party models from OpenAI, Google, Meta and Anthropic.
What is the GS AI Assistant built on?
It runs on the firm's GS AI Platform. In January 2025 it drew on OpenAI, Google Gemini and Meta Llama models fed with Goldman data, with Anthropic, Mistral and Cohere under evaluation; by July 2025 staff could choose among current GPT, Gemini and Claude models. Argenti says the firm wants to keep plugging models in and out by task and cost.
Who builds AI at Goldman Sachs?
CIO Marco Argenti oversees it. Daniel Marcu, hired from Amazon, is Global Head of AI Engineering and Science; Rahul Sharma heads GS AI platform engineering and Bing Xiang heads AI Research within Engineering, both reporting to Marcu. Archana Vemulapalli runs AI product management and frontier-lab relations, and Chris Churchman leads Marquee AI.
What is Marquee AI?
Marquee AI is a natural-language assistant on Marquee, Goldman's digital platform for institutional and corporate clients. It breaks a question into research topics and attributes every statement to a specific line in a source, using a grounding framework the team built. As of August 2026 it was available internally only.
How many Goldman Sachs employees use its in-house AI?
The GS AI Assistant reached about 10,000 employees in January 2025 and all 46,000 by July 2025, handling more than a million prompts a month. GitHub Copilot on the GS AI platform launched to more than 9,000 developers, about 80% of engineers, in June 2024.
- Goldman Sachs CIO is anxious to see results from GenAI, but moving carefully | The AI Beat
- Goldman Sachs Deploys Its First Generative AI Tool Across the Firm
- Goldman Sachs CIO says AI tools may offer superhuman productivity, but 'I don't know if they can be superhumanly smart'
- AI Exchanges: CIO Marco Argenti on the future of AI in the workplace (transcript)
- Goldman Sachs wants to create AI that thinks and acts like a seasoned banker
- Goldman Sachs staff now write a million gen AI prompts a month
- Why Goldman Sachs' CIO is taking a measured approach to rolling out AI across the business
- Building AI Systems for Capital Markets (Goldman Sachs Exchanges)
- Goldman Sachs launches AI investing platform – report
- Goldman Sachs data governance code goes live on Finos
- Goldman Sachs is piloting its first autonomous coder in major AI milestone for Wall Street
- Goldman Sachs taps Anthropic's Claude to automate accounting, compliance roles
- Goldman Sachs hires Amazon exec in senior AI engineering role
- Goldman Sachs’ tech boss says tracking individual AI usage isn’t useful
- Marco Argenti gives Goldman Sachs coders an AI helper (Most Innovative People in Finance, #3)
- Daniel Marcu Joins Goldman Sachs as Partner and Global Head of AI Engineering and Science
- Archana Vemulapalli Joins Goldman Sachs as Partner and Head of AI Product Management and Strategic Relations
- US12518171B2: Data generation and retraining techniques for fine-tuning of embedding models for efficient data retrieval
- US12585896B1: Embedding-free retriever-augmented generation (RAG) architectures
- Marco Argenti gives Goldman Sachs coders an AI helper
- How Goldman Sachs Is Using Agentic AI For Software Engineering At Scale
- Anthropic Partners with Blackstone, Hellman & Friedman, and Goldman Sachs to Launch Enterprise AI Services Firm
- Anthropic, Blackstone, and Hellman & Friedman Introduce Ode with Anthropic
- Financial Services Roundup: Market Talk (Solomon shareholder letter)
- Goldman Sachs teams up with Anthropic to automate banking tasks with AI agents, CNBC reports
- Goldman Sachs launches AI assistant firmwide, memo shows
- Goldman Sachs rolls out an AI assistant for its employees as artificial intelligence sweeps Wall Street
When Goldman moves on AI, you'll read it here first.
when one of the banks we cover moves on AI, it is in the next morning's brief · six sourced stories · 7 am PT · free
plus every regulation, bank and agent page update, the morning after · leave any morning