Goldman Sachs exploits Anthropic AI model to automate accounting and compliance

Goldman Sachs He’s working with AI startup Anthropic to create AI agents to automate a growing number of roles within the bank, the company’s chief technology officer told CNBC exclusively.

The bank has, for the past six months, been working with Anthrop engineers to co-develop autonomous agents in at least two specific areas: accounting for trades and transactions, and client screening and onboarding, according to Marco Argenti, Goldman’s chief information officer.

Argenti said the company is “in the early stages” of developing agents based on Anthropic’s Cloud model that will reduce the amount of time these basic functions take. He expects to launch the clients “soon,” although he declined to provide a specific date.

“Think of it as a digital co-worker for many occupations within a company that are expansive, complex and very process-intensive,” he said.

Goldman Sachs CEO David Solomon said in October that his bank was embarking on a multi-year plan to reorganize itself around generative artificial intelligence, a technology that has been making waves since the arrival of OpenAI’s ChatGPT in late 2022. Even as investment banks like Goldman see their revenues from trading and advisory activities rise, they will seek to “constrain headcount growth” amid the overhaul, Solomon said.

The news from Goldman comes as model updates from Anthropic, co-founded by a former OpenAI executive, have sparked sharp selling among software companies and their credit providers, as investors bet on who the winners and losers of AI trading will be.

Goldman last year began testing a standalone AI encryption program called Devin, which is now widely available to the bank’s engineers. But she soon found that Anthropic’s AI model could work in other parts of the bank, Argenti said.

“Claude is really good at programming,” Argenti said. “Is it because programming is kind of special, or is it more about the model’s ability to solve complex problems, applying step-by-step logic?”

Argenti said the company was “surprised” by how well Claude was able to perform tasks alongside programming, especially in areas such as accounting and compliance that combine the need to analyze large amounts of data and documents while enforcing rules and governance.

The view inside Goldman now is that “there are other areas of the company where we can expect the same level of automation and the same level of results that we see on the programming side,” he said.

The result is that by helping agents develop, clients will be onboarded faster and issues with trade settlement or other accounting issues will be resolved faster, Argenti said.

Goldman could then develop agents for tasks such as monitoring employees or preparing guidelines for investment banking, he said.

While the bank is employing thousands of people in compliance and accounting jobs where AI agents will soon be working, Argenti said it is “too early” to expect the technology to lead to job losses for those workers.

However, he said Goldman Sachs could do without the third-party providers it uses today as AI technology matures.

“It’s always a trade-off,” Argenti said. “Our philosophy now is that we are increasing capacity, which in most cases will allow us to do things faster, which translates to a better customer experience and more business.”