Goldman Sachs deploys its first generative AI tool across the firm

Chief Information Officer Marco Argenti said the company’s approach to generative AI involved centralizing all proprietary uses of the technology on an internal platform, and restricting them elsewhere. “It might have slowed us down initially, but then we definitely gained a lot of velocity afterwards," Argenti said.

“Taking this centralized approach obviously has pros and cons," he added. Argenti said he had to push back against those who wanted to move faster and contend with some frustration over the bank’s decision to ban the use of OpenAI’s ChatGPT within its walls, a step taken by other firms.

CIOs across industries, facing pressure from chief executives and corporate boards to leverage generative AI, are stuck between moving fast enough to achieve competitive gains, but not so fast they leave themselves vulnerable to dangerous hiccups. A number of efforts remain in the pilot or proof of concept phase, although ideation is gradually moving to more implementation this year, said Chirag Dekate, a vice president analyst at research firm Gartner.

Goldman’s generative AI platform, known as the GS AI Platform, grew out of an existing machine-learning platform and is the single point of entry for all generative AI use at the company. Goldman’s approach also included tapping partnerships with OpenAI-backer Microsoft to use GPT-3.5 and GPT-4 models and Google for its Gemini model. The platform also uses open source models including Meta Platforms’ Llama. The ability to switch between models for different use cases is a key benefit of the approach, Argenti said.

Wall Street Journal owner News Corp has a content-licensing partnership with OpenAI.

Critically, the internal platform also allows Goldman to fine-tune the models with its own internal data in a safe way and that complies with regulations. Argenti said controls are embedded to ensure that models aren’t serving up data to employees who shouldn’t have access to it, for example.

Select Goldman employees can access the platform directly to interface with various models, and developers are using it to build custom applications on top of the models, such as a copilot assistant tool for investment bankers that searches a large body of public and proprietary documents to answer questions and extract analysis.

The other benefit of the platform is being able to build more applications faster, Argenti said. Developers don’t have to start from scratch with every application—a copilot for asset managers can lean heavily on the copilot for investment bankers and all the safety guardrails are already built-in from the beginning.

Argenti said the speed at which developers are building generative AI applications has shortened from months to weeks as a result. But at the same time, the company isn’t rushing them into production.

“AI is a new thing," he said. “Safety and responsibility is absolutely our number one priority."

The need to safeguard data, remain compliant with existing data regulations and prepare for any new AI regulations are key factors that help determine the speed with which financial services adopt AI, said Gartner’s Dekate. Financial services is one of the most regulated ecosystems, he added.

Code generation is a common starting point for many entities because it is text-centric and provides clear, tangible efficiency gains, he said.

Goldman’s generative AI coding assistant, Microsoft’s GitHub Copilot, is the firm’s most scaled-out use of AI so far. Argenti said it is giving developers about a 20% increase in efficiency. The company is also in earlier stages of rolling out an AI tool that translates documents into other languages and an application that summarizes proprietary research to provide insights to advisers.

Generative AI remains a relatively small part of the technology budget, Argenti said, with most resources dedicated to running the bank, and keeping it safe and compliant.

Argenti said he sees AI drawing more dollars in the future, but “it’s not something where all of a sudden we pivot resources from other things."