Tracy Alloway and Joe Weisenthal host Marco Argenti, Chief Information Officer at Goldman Sachs, for his second appearance on the podcast since August 2024. Argenti oversees AI implementation across Goldman's 47,000 employees and manages the firm's technology infrastructure.

The conversation explores how Goldman has moved beyond AI experimentation to production deployment, with concrete examples of workflow transformation and cost optimization. Topics include the GSA assistant system, agentic AI for developers, token budget management, and regulatory discussions around AI deployment in financial services.

Argenti discusses the practical realities of enterprise AI adoption, from replacing third-party software vendors to managing the 'token sticker shock' that CFOs are experiencing as AI usage scales across organizations.

From AI Experimentation to Production at Scale

Goldman has moved decisively past the experimentation phase, with 47,000 employees using the GSA system daily for complex research and analysis tasks that previously required hours or days.

The GSA system can answer multi-dimensional questions like 'How do recent geopolitical events in the Strait of Hormuz impact portfolio volatility and what rebalancing strategies are optimal?' by calling models, retrieving data, and creating execution plans.

Data quality emerged as the key differentiator between effective and ineffective AI implementations, requiring extensive work to make hundreds of data sources understandable to AI systems.

The Legend AI lakehouse tool allows users to go from query to answer in 'two or three clicks' by connecting data sources directly to the GSA assistant.

Developer Productivity Revolution Through Agentic AI

Every Goldman developer now has access to agentic AI tools including DEVIN, Claude Code, and GitHub Copilot, fundamentally changing how software development works.

Projects are consistently finishing ahead of schedule, with one cloud migration project completing two months early - 'This is how we know when things are going to work' - Argenti.

AI is transforming developers into product managers and planners rather than coders, with the most important skill becoming the ability to explain requirements clearly.

Weekend 'vibe coding' has become common, with developers creating complete applications or cloud migrations of legacy systems in hours rather than months.

The Economics of Token Consumption and Cost Management

Goldman built a centralized Model Gateway that intelligently routes requests to optimize the 'Pareto frontier of quality and cost' across different AI models.

Token costs are destined to become 'a major item of cost in any organization' comparable to human wages rather than traditional IT expenses - Argenti.

The philosophy is to eliminate 'token anxiety' for users, similar to electric car range anxiety, allowing them to focus on productivity while the platform team optimizes costs behind the scenes.

Per-unit token costs will decline but total token consumption will increase dramatically as organizations adopt agentic workflows and reasoning models.

Disrupting Legacy Software Through Internal AI Development

Goldman has already terminated contracts with third-party software providers, replacing them with AI-built internal solutions as the 'buy versus build' equation shifts dramatically.

The cost of building simple applications has dropped to weekend projects, while complex enterprise software still requires traditional development approaches.

Software disruption depends on whether underlying business processes will change - accounting and regulatory functions remain stable while developer workflows face major transformation.

Forward-deployed engineers from AI companies work directly at Goldman as 'tailors rather than fashion assistants' to accelerate implementation during rapid technological change.

Regulatory Compliance and Risk Management for AI Systems

Banks have used neural networks for decades, so AI regulation discussions focus on familiar territory of risk tiering, controls, and human supervision rather than explaining black box decisions.

Goldman employs a zero-trust model where AI cannot approve its own code - all AI-generated code goes through the same human review and CI/CD pipeline as human-written code.

Information barriers are enforced at the source level, with each AI session receiving badges that restrict data access according to the same rules governing human employees.

The centralized GSA platform took nearly two years to build specifically to handle cybersecurity, information barriers, and regulatory compliance requirements.

The Future of Human-AI Collaboration in Finance

AI is turning every employee into a manager who must explain, delegate, and supervise - the fundamental skills needed to work effectively with AI agents.

Goldman's competitive advantage lies in the 'extra ten percent' beyond what public AI can provide - proprietary data, global relationships, and complex multi-asset expertise.

The firm optimizes for 'velocity' (sustained speed over time) rather than short-term 'speed' that hits security or scalability walls.

Developer satisfaction has increased as AI eliminates repetitive tasks like library upgrades and mechanical work, allowing focus on higher-level planning and architecture.