Goldman Sachs CIO says AI's next job at the bank is making money

At Goldman Sachs, a workday once began when the first email arrived. Today it might begin with a query to the bank's in-house assistant, or with an agent, according to chief information officer Marco Argenti, who said AI touches pretty much everyone at the firm ^[5]. He joined Goldman in 2019 from Amazon Web Services, where he was vice president of technology ^[4].

Goldman counted usage first. It measured AI through proxies such as how often developers committed code, and nobody could trace those to dollars, Argenti said ^[9]. What the tools actually changed shows up in delivery time. A three-month project finished in two months takes a third less time ^[17]. Over a year of work, that is 1.5 projects delivered in the time one used to take ^[18].

Argenti spoke with Bloomberg's Tom Mackenzie at Wave by Vento in Turin on Friday and framed the third phase as a change of goal ^[2]. "How do you make money because of AI, not only how do you save money because of AI?" he said ^[3]. He also said the growth wave is "emerging right now," and described it as AI helping the company grow as well as become more efficient ^[8]. The money evidence he gave is a budgeting result: projects that fell below the line in zero-based budgeting now fund themselves ^[10]. His account does not include a revenue figure tied to AI.

Developers are the group the model is built for first, and they are about a quarter of Goldman's staff ^[19]. Argenti said their job now is to spell out what a good result looks like and to run resources and priorities, much as an entrepreneur would ^[13]. On whether that means a smaller workforce, he said the option could be open to the bank. But any engineering backlog contains far more work than each planning cycle pays for, and as long as there is appetite to grow, there will be plenty of work first, he said ^[11].

Goldman assumes its models will make errors, as people do, because a statistical machine will not give the same result every time ^[14]. Argenti compared the setup to a kindergarten, where you remove the sharp edges instead of handing each child a safety policy ^[14]. The bank also reads a model's chain of thought and uses other AI models to challenge its work ^[16].

Goldman's waves turn into a test with one pass condition each. For developers, it is retiring activity measures like commit counts and tracking delivery time on funded work ^[9]. In the process wave, the test is whether a workflow step can be deleted outright, the question Argenti said to ask of every step; speeding up the same step does not pass ^[7]. The growth wave passes when a project that missed the budget line gets funded from its own results ^[10]. By Argenti's account, Goldman is entering the third while still rethinking its processes in the second ^[2][7].