Half of the 46,000 employees at Goldman Sachs now have access to artificial intelligence. By the end of this year, chief information officer Marco Argenti expects even more will be able to tap into AI in hopes of boosting their productivity—yet still not everyone at the firm.

“We have the entire organization that needs to somehow re-tune and re-tool itself for AI,” says Argenti. “But, I think we’ve been very, very, very intentional with regards to driving people change management.”

The measured approach at Goldman Sachs, ranked #35 on the Fortune 500, reflects Argenti’s view that AI technology is rapidly evolving and still comes with a lot of uncertainty. Goldman’s AI steering group and risk and control teams work to determine which of the dozens of AI proposals should be tackled and how that can be done responsibly.

One example is the experimentation Goldman is doing with agentic AI, which still isn’t fully deployed. AI agents are meant to work autonomously or with little human oversight, performing multi-step reasoning or task completion. These agents could, theoretically, perform compliance checks or help process customer transactions. But the AI agents also need specific training and can hallucinate, resulting in errors in the results they produce. Goldman says it is still assessing what additional controls it needs to effectively and safely use agentic AI.

Because Goldman operates in a highly regulated sector, the industry historically prefers to build technology in-house, giving these institutions more control to protect sensitive customer financial data. That’s changed with the rise of cloud, and more recently generative AI, and a vast majority of financial institutions have deployed at least one generative AI product, frequently partnering with external vendors.

Roughly one out of every four Goldman Sachs employees is an engineer, and this group was the first Argenti targeted when deploying generative AI tools. Argenti gave those workers access to AI coding assistant tools, including GitHub Copilot and Gemini Code Assist. Goldman has conducted competitions inspired by reality TV entrepreneurial competition show Shark Tank so that developers could share their most creative uses of AI.

Argenti measures the return on investment from these copilot tools in a few ways, including frequency of use and the acceptance rate of code generated by GitHub and similar tools. 

Broader use of generative AI within the company came with the launch of GS AI Assistant, which rolled out last year and has expanded to 10,000 employees including bankers, traders, and asset managers. This tool, which Goldman anticipates will be available to nearly all employees by the end of 2025, can summarize documents, draft emails, analyze data, and create personalized content. 

GS AI Assistant was built to be multi-modal, utilizing large language models from Gemini, OpenAI and Llama, with Argenti exploring adding LLMs from other AI hyperscalers. Argenti says he doesn’t want to rely on just one vendor and is giving the firm the flexibility to use a model that may be better for coding, while a rival offering is stronger at reasoning. Goldman also factors in how easy the LLMs are to modify and how expensive they are to run.

“All of those considerations got us to the point where we want to continue to plug-and-play with those models,” says Argenti.

For workers not in engineering, Goldman is tracking usage rates and sends out surveys to get feedback to make continued improvements to GS AI Assistant. The company has sought to promote champions from asset and wealth management, private banking, and trading—not the technologists—to get buy-in. “People might be afraid or skeptical when you drive technology first,” says Argenti.

Argenti joined the firm as a partner and co-CIO in 2019 and fully took on the role in 2022 after his co-CIO, George Lee, became co-head of the geopolitical and technology insights arm Goldman Sachs Global Institute. Prior to joining Goldman, Argenti was vice president of technology at Amazon Web Services for six years and also held leadership roles at telecommunications company Nokia.

A lot of his earlier work at Goldman focused on enabling the firm’s employees to work from home as a result of the global pandemic. But he also wanted to shift the culture of engineering to be less like how a bank thinks about technology, which tended to favor creating bespoke solutions for each separate division, and more like a tech giant that creates one tool to be shared across the firm. 

This newer way of thinking is reflected in the open-source data management and governance platform Legend, which publicly launched in 2020. Goldman built Legend internally over 10 years prior to the public launch, giving both technologists and non-technical users the ability to develop data-centered applications and derive insights from that data.

Data is a key component of Argenti’s AI strategy, which he calls a three-leg stool that should also represent the AI technology itself and the people who use it. Good quality data is needed for the right output from LLMs, but changing people’s behavior is equally important. 

“It’s about amplifying capabilities and in the hands of the best people, I think you’re going to get the best results,” says Argenti.

John Kell

Send thoughts or suggestions to CIO Intelligence here.

This story was originally featured on Fortune.com

Source link