Goldman CIO Marco Argenti on the Warp-Speed Improvements in AI

What It Covers

Goldman Sachs CIO Marco Argenti returns to Odd Lots to detail how the firm deploys AI across 47,000 employees in 2026, covering token cost management, agentic developer tools, legacy software disruption, regulatory frameworks, and how AI is reshaping the skills required from every worker at the firm.

Key Questions Answered

• •AI ROI Measurement: Track project delivery timelines rather than headcount reduction to measure AI value. Goldman measures success by projects finishing ahead of schedule, treating surplus developer capacity as optionality — either absorbing more backlog or reducing costs — rather than immediately cutting staff to capture savings.
• •Token Budget Centralization: Build a model gateway that routes queries to the Pareto-optimal combination of quality and cost before worrying about individual token anxiety. Goldman's GSAI platform intelligently routes simple queries to cheap local models and complex ones to frontier models, preventing unmonitored API sprawl and unexpected CFO-level billing shocks.
• •Buy vs. Build Equation Shift: The cost calculus for software procurement has fundamentally changed. Simple internal applications that previously required months and millions to build can now be prototyped in hours. Goldman has already terminated third-party software contracts where internal builds became viable, with the pendulum swinging back toward build for smaller-scope tools.
• •Agentic Worker Skills Framework: AI turns every employee into a manager requiring three core competencies: explain tasks clearly, delegate work across specialized agents in parallel, and supervise outputs critically. Goldman actively recruits and trains for these skills — ideation, delegation, and quality judgment — rather than traditional individual-contributor technical execution.
• •Data Quality as AI Differentiator: Data curation determines AI output quality more than model selection. Goldman wired hundreds of data sources into its GSAI assistant and built Legend AI, a lakehouse tool that converts data to MCP server connections in two to three clicks, enabling complex multi-dimensional client portfolio queries that previously took days or weeks.