Code and Content Gen AI is among the most adopted and highest RoI AI use cases among enterprises
Everyone’s probably already heard that Goldman Sachs built an internal AI platform called GS AI platform but here’s how they did it.
TLDR
- Built behind the Firewall - GS’ AI platform hosts GPT - 4, Gemini, Llama, Claude, and internal models all within their networkRailguards all along - Encryption, prompt filtering, role-based access, audit logs, human-in-the-loop approachProductivity gains across GS - >50% adoption among 46k employees and a productivity increase of 20% among coders, 15% reduction in post-release bugsBacked by execs - CEO David Solomon and CIO Marco Argenti (hired from Amazon) are gunning for 100% adoption among employees by 2026
Goldman Sachs wanted to allow their employees to converse with large language models to boost productivity across the firm with emphasis on security, compliance and governance controls.
In this article we’ll go through the platform’s architecture, security measures, developer integrations, model customization, organizational impact and next steps
Architecture: Secure Multi-Model AI Behind the Firewall
A GS employee uses the GS AI interface through a chat interface much like how we use ChatGPT where they can start new conversations.
“a very simple interface that allows you to have access to the latest and greatest models” - Marco Argenti, CIO, GS
Technical stack and orchestration: GS AI Platform supports local or secure API deployments of models like OpenAI’s GPT variants, Google’s Gemini, Meta’s LLaMA, and Anthropic’s Claude. Its flexible architecture can add new models and route tasks to the best fit code requests to coding models, document summaries to language/finance-tuned models ensuring high-quality results across use cases. This method of multi-model orchestration means that GS can swap out models without retraining the users.
Use of proprietary data: All queries are routed through an internal gateway that adds proprietary data and context before reaching the model. Using retrieval-augmented generation (RAG) and fine-tuning, responses are generated primarily from GS’ own up-to-date, domain-specific knowledge. Initially trained on Goldman data within models from OpenAI, Meta, Google, and others, the system will increasingly integrate more internal context as additional firm data is indexed.
Security and Compliance
All AI interactions pass through a secure compliance gateway that applies prompt filtering, data anonymization and policy checks so that no sensitive information is sent to the models and outputs comply with firm and regulatory rules. Encryption is used for data in transit into any model APIs, and sensitive prompts or responses are masked within the system.
Compliance and audit trails: The platform maintains an audit trail of all AI interactions allowing compliance teams to check the information given to or generated by AI, which model was used and who was the person running the interaction.
Access control: AI limits access to certain models and databases based on employee role, department and use-case. For instance a research analyst can get access to financial data while a developer might get access only to codebases.
Token-level filtering: Every prompt is analysed to strip or replace sensitive data (e.g., client names, account numbers) before sending them to external models. Combined with real-time compliance scanning of both inputs and outputs, this prevents leaks, blocks disallowed content.
AI in the SDLC
One of the earliest and most impactful uses of Goldman’s AI platform is to assist software developers and engineers in coding tasks. Goldman deployed AI coding assistants within VS Code and JetBrains IDEs so developers can get code suggestions, completions, and explanations right as they write code.
The AI Developer Copilot is capable of tasks like: explaining existing code, suggesting bug fixes or improvements, translating code between programming languages, and even generating boilerplate code or test cases on the fly.
To integrate this safely, Goldman sandboxed the AI’s coding suggestions and instituted additional checks. All code generated by the AI goes through the normal code review process and automated testing pipelines before being merged or deployed, ensuring that any mistakes are caught by human developers or QA tools.
GS offers both Microsoft’s and Google’s code models internally, so they could compare their performance and ensure redundancy (if one model had an outage or limitation, another could be used).
Model Customization and Domain Specific Tuning
Goldman Sachs didn’t simply take off-the-shelf AI models - they customized and fine-tuned models for internal use cases to maximize performance and safety. One key aspect of this is feeding Goldman’s extensive internal data (financial texts, code repositories, research archives, etc.) into the models, so that the AI’s knowledge is grounded in Goldman’s context.
Fine-tuning: Open-source and internal models are trained on Goldman’s proprietary codebases, research, and market data, making outputs align with internal standards, abbreviations, and historical context.
RAG: The AI can pull relevant internal documents in real time via platforms like Legend to answer queries with precise, source-backed information.
Role-based behaviour: Access and model capabilities are segmented by user clearance. Specialised variants (e.g., Banker Copilot, Research Assistant) are tuned for department-specific needs.
Multi-size model strategy: Smaller models that could handle less complex tasks quickly, allows them to reserve the giant models for truly hard problems.
Organizational Impact and Cultural Change
- Developer productivity: 20%+ faster coding cycles; tasks that took 5 days now done in 4, with fewer bugs.Dramatic time savings: IPO document drafting cut from weeks to minutes (AI does 95% of work); document translation & regulatory comparisons reduced from hours to seconds.Error reduction: AI catches anomalies in reports, code, and financial models, reducing manual mistakes with a 15% reduction in post release bugs Widespread adoption: Opened to 46,500+ employees in June 2025; >50% adoption today with a goal of 100% usage by 2026Change management success: AI “champions” in each business unit, training workshops, and strong messaging that AI augments rather than replaces jobs.Faster onboarding: New hires use AI as a tutor, speeding up learning on codebases, models, and internal processes.
“Leveraging AI solutions to scale and transform our engineering capabilities as well as to simplify and modernize our technology stack” - David Solomon, CEO, Goldman Sachs
The Next Phase: Devin
Goldman Sachs is piloting Devin, an AI software engineer built by Cognition, as part of its move into autonomous AI tools. Unlike an AI Assistant, which waits for you to tell it what to do step by step, Devin can take a goal, figure out the steps, write the code, test it, and hand it back for review.
Right now, the pilot is aimed at the kind of work developers don’t love - updating old code, migrating systems, cleaning up legacy frameworks, and cranking out boilerplate. The idea is to clear backlogs and speed up delivery. Developers still stay in the loop, assigning Devin tasks and checking its work before anything goes live.
Goldman’s CIO, Marco Argenti, thinks this could mean 3-4x faster output compared to today’s AI tools. If it works, the bank could roll out hundreds of these agents and use them for other areas like operations, research, or finance.
The trial is also a test of whether this kind of AI can work inside Goldman’s tight compliance rules. If Devin proves itself, it could be plugged into the GS AI Platform so employees could ask the AI to just get things done, not just assist. That could change how a lot of work gets done at the bank.
Yes, We Can
Goldman Sachs’ AI strategy shows how a large, regulated enterprise can embrace transformative technology without compromising security or compliance. The firm’s behind-the-firewall approach allows the entire workforce to access advanced AI models. Early results are impressive with productivity lifts on the order of 20% in key functions. Equally important is the change in mindset - Goldman’s workforce is increasingly treating AI as a collaborator, and the firm is training its people to leverage and supervise AI effectively. Executive leadership is fully aligned with these changes, clearly articulating that AI is central to Goldman’s strategy for innovation, efficiency, and competitiveness in the coming years.
GS AI platform offers a case study for CIOs in regulated industries. It demonstrates that with the right architecture and controls, even sensitive sectors like finance can harness generative AI to automate grunt work, surface insights, and enhance decision-making
