Goldman Sachs' Secure AI Platform: Behind-Firewall LLM Deployment & Impact

Goldman Sachs has successfully deployed an internal artificial intelligence platform, the GS AI Platform, marking a significant step in the financial giant’s embrace of advanced technology. The initiative, driven by a desire to boost productivity across the firm, prioritizes stringent security, compliance, and governance controls, setting a precedent for other highly regulated industries.

At its core, the GS AI Platform operates entirely behind Goldman Sachs’ robust firewall, hosting a suite of large language models including OpenAI’s GPT-4, Google’s Gemini, Meta’s Llama, and Anthropic’s Claude, alongside proprietary internal models. Employees access these capabilities through a simple chat interface, akin to public-facing AI tools, allowing them to initiate new conversations and leverage the latest models. The platform’s flexible architecture supports multi-model orchestration, intelligently routing tasks—such as code requests or document summaries—to the most suitable model without requiring users to retrain. This approach ensures high-quality results across diverse use cases and allows for seamless model swapping.

A cornerstone of the platform’s design is its rigorous security and compliance framework. All AI interactions pass through a secure compliance gateway that applies prompt filtering, data anonymization, and policy checks. This prevents sensitive information from being transmitted to the models and ensures outputs adhere to firm and regulatory rules. Data in transit to any model APIs is encrypted, and sensitive prompts or responses are masked within the system. The platform maintains a comprehensive audit trail of all AI interactions, providing compliance teams with detailed records of information exchanged, models used, and user identities. Access to specific models and databases is meticulously controlled based on employee role, department, and use case, while token-level filtering analyzes every prompt to strip or replace sensitive data like client names or account numbers before processing, preventing data leaks and blocking disallowed content.

One of the earliest and most impactful applications of the platform has been in assisting software developers. Goldman Sachs has integrated AI coding assistants directly into popular Integrated Development Environments (IDEs) like VS Code and JetBrains. These AI tools provide real-time code suggestions, completions, explanations, and can even propose bug fixes, translate code between languages, or generate boilerplate code and test cases. To ensure safety, all AI-generated code is sandboxed and subjected to the firm’s standard code review processes and automated testing pipelines before deployment, ensuring human oversight and quality control. The firm also offers both Microsoft’s and Google’s code models internally, providing redundancy and allowing for performance comparisons.

Beyond off-the-shelf solutions, Goldman Sachs has extensively customized and fine-tuned models for internal use cases. A crucial aspect involves feeding the AI models with Goldman’s vast repository of proprietary data, including financial texts, code repositories, and research archives. This grounds the AI’s knowledge in the firm’s specific context. Open-source and internal models are trained on this data, ensuring outputs align with internal standards, abbreviations, and historical context. The system also leverages Retrieval-Augmented Generation (RAG), enabling the AI to pull relevant internal documents in real-time to answer queries with precise, source-backed information. Furthermore, specialized variants like a “Banker Copilot” or “Research Assistant” are tuned for department-specific needs, and a multi-size model strategy reserves larger, more complex models for truly challenging problems while using smaller, faster models for simpler tasks.

The organizational impact of the GS AI Platform has been substantial. Developers report over 20% faster coding cycles and a 15% reduction in post-release bugs. The time required for tasks such as IPO document drafting has been dramatically reduced from weeks to minutes, with AI handling approximately 95% of the work. Document translation and regulatory comparisons, once hours-long endeavors, now take mere seconds. Since its widespread release to over 46,500 employees in June 2025, the platform has achieved over 50% adoption, with a target of 100% usage by 2026. This success is attributed to robust change management, including “AI champions” in each business unit, training workshops, and clear messaging from executive leadership—including CEO David Solomon and CIO Marco Argenti—that AI augments rather than replaces jobs. New hires also leverage AI as a tutor, accelerating their onboarding into complex codebases and internal processes.

Looking ahead, Goldman Sachs is piloting Devin, an autonomous AI software engineer developed by Cognition. Unlike existing AI assistants that require step-by-step instructions, Devin can take a high-level goal, devise a plan, write code, test it, and present a solution for review. This pilot focuses on automating less appealing tasks like updating legacy code or migrating systems, aiming to clear backlogs and accelerate delivery by potentially tripling or quadrupling output compared to current AI tools. The trial is also a critical test of Devin’s ability to operate within Goldman’s stringent compliance framework, with the potential for future integration into the GS AI Platform, allowing employees to delegate complex tasks for autonomous completion.

Goldman Sachs’ AI strategy offers a compelling case study for Chief Information Officers in regulated industries. It demonstrates that with a thoughtful architecture and robust controls, even sensitive sectors like finance can effectively harness generative AI to automate routine work, unearth insights, and enhance decision-making without compromising security or compliance. The firm’s “behind-the-firewall” approach has enabled its entire workforce to access advanced AI models, fostering a collaborative mindset where AI is seen as a powerful partner, poised to redefine work processes at the bank.