Goldman Sachs went from building custom AI tools to buying agentic platforms in 18 months.
The Summary
- Goldman's CIO Marco Argenti discusses the bank's shift in AI strategy, particularly around adopting agentic platforms like Claude Code
- The conversation reveals how quickly enterprise AI deployment changed between late 2024 and early 2026
- Goldman moved from building internal tools to integrating third-party agent platforms
The Signal
Eighteen months ago, Goldman Sachs was in build mode. Custom internal AI tools. Proprietary infrastructure. The classic enterprise playbook: if it matters, we make it ourselves. Now they are buying off the shelf. That is the story.
Marco Argenti's return to Odd Lots marks a clear inflection point in how serious enterprises think about AI agents. When one of the world's most sophisticated financial institutions stops building and starts buying, that tells you the platforms won. Claude Code and similar agentic systems have crossed the threshold from "interesting experiment" to "production-grade infrastructure."
This matters because Goldman is not some mid-market company trying to move fast and break things. They have capital, talent, and a paranoid approach to technology risk. If they trust third-party agent platforms to run inside their systems, those platforms have matured faster than most people tracking this space realized. The build versus buy calculus shifted dramatically, and it shifted in the direction of integrated agent platforms.
The timing is revealing too. Late 2024 was still the era of companies announcing AI initiatives and showing off custom chatbots. Early 2026 is the era of companies quietly deploying agents that actually do work. The gap between those two moments was filled with a lot of corporate learning about what AI can actually handle at scale, and apparently, the answer was: more than expected, but only if you use the right tools.
The Implication
Watch who else stops talking about their internal AI labs and starts talking about their agent platform partnerships. That is the leading indicator for where production AI is actually working. If you are building enterprise tools, the message is clear: the window for custom solutions just narrowed. The platforms are winning on speed and capability.
Source: Bloomberg Tech