Bloomberg/Getty Images
- Marco Argenti, chief information officer, said engineers need to transfer "tribal knowledge" to AI.
- Now that all developers are using AI, Argenti said the focus is on making tools specific to Goldman.
- AI is also transforming mentorship structures across lines of business.
Marco Argenti is facing a new problem: how to communicate the subtleties of Goldman Sachs' engineering culture to an AI bot.
Argenti, the firm's chief information officer, said that since the firm's more than 12,000 developers are all using AI, the focus is on making their AI tools Goldman-specific. All of the bank's developers have access to updated agentic technology, including Claude and Devin, Cognition's AI coding assistant. They now must figure out how to "mentor" the AI, much as they would a new employee who doesn't yet know Goldman's standards. An AI agent that's familiar with Goldman's environment — its data standards or security protocols, for example — can produce better work, faster for engineers to review.
"The transfer of the institutional knowledge into the AI is the biggest question," Argenti said. "How does an experienced GS AI look versus a naive AI?"
Transferring "tribal knowledge"
Goldman spent around $6 billion on AI last year and is, like its peers, under mounting pressure to show that its investments are paying off. Jamie Dimon, the CEO of JPMorgan, said spending on AI is a prerequisite to staying competitive at this point.
Developers are often the most sophisticated AI users at banks, and Argenti said that they now need to teach AI agents the "tricks and tribal knowledge" that the technology doesn't intuitively grasp. Goldman's "engineering tenets," as described in a blog post on the firm's website, include "innovate incrementally" and "look around corners," subjective directives that AI tools can't grasp as easily as rote instructions.
The firm has drafted "skills" — reusable bundles of instructions for performing specific tasks — to capture developers' technical knowledge about Goldman, like the design principles, data models, and environments used. One skill, for example, explains how to migrate information to the cloud; the "cloud fast track" skill teaches AI what good cloud migration looks like, specifically at Goldman. Thanks to the skill, AI tools are more helpful to those tasked with cloud migration, much like an experienced employee would likely be more helpful than a new hire.
"The unwritten rules are the ones that are actually harder to capture. That's why we try to systematize them by doing these evals," Argenti said.
As part of the effort, Argenti said, Goldman has studied its internal processes by conducting interviews and analyzing outputs, among other measures. The bank regularly updates its skills as its internal processes evolve, creating a "constant loop of improvement," in which the agents' code becomes both better and more specific to Goldman.
Mentorship is changing
At this point in the bank's AI rollout, engineers aren't spending as much time coding but instead making sure their agents are doing a good job according to firm standards. It's a profound change, Argenti said, and one that will likely become relevant to other lines of business.
AI is also changing mentorship among humans at the bank, Argenti said. Whereas junior employees usually turn to managers for advice, they're now mentoring firm veterans on AI use, both in structured and unstructured ways. Other firms are leaning into similar peer-to-peer programs — at Citi, thousands of employees have volunteered to serve as "AI accelerators" for their colleagues.