Goldman Sachs is building AI agents with Anthropic for accounting and compliance

Goldman Sachs has been working with Anthropic for the past six months to co-develop autonomous AI agents for back-office functions, the bank's chief information officer told CNBC in an exclusive report published today.^[1]

The effort targets two specific areas: accounting for trades and transactions, and client vetting and onboarding. CIO Marco Argenti said the firm is "in the early stages" of building the agents using Anthropic's Claude model, and expects to launch them "soon," though he declined to give a specific date.

Anthropic engineers have been embedded directly within Goldman's internal teams during the development process.

How it started

The project grew out of an earlier experiment. Goldman began last year by testing Devin, an autonomous AI coding assistant, which is now broadly available to the bank's engineers. That trial raised a question Argenti put to CNBC: "Claude is really good at coding. Is that because coding is kind of special, or is it about the model's ability to reason through complex problems, step-by-step, applying logic?"

The answer turned out to be the latter. Argenti said the firm was "surprised" at how capable Claude was at tasks beyond coding, particularly in areas like accounting and compliance that require parsing large volumes of data and documents while applying rules and judgment. That finding opened the door to a broader rollout.

"There are these other areas of the firm where we could expect the same level of automation and the same level of results that we're seeing on the coding side," Argenti said.

What the agents will do

Argenti described the planned agents as a "digital co-worker for many of the professions within the firm that are scaled, are complex, and very process-intensive."

Trade accounting and client onboarding are among the most documentation-heavy functions inside a global investment bank. They involve high volumes of records, cross-referencing against regulatory requirements, and frequent exception handling. These are the kinds of workflows where speed and consistency matter, but where the source material is rarely cleanly formatted or centralized.

That the agents are being built for these areas, rather than for more visible functions like research summarization or client-facing tools, is itself notable. It suggests Goldman sees Claude's reasoning capabilities as suited to rules-driven, audit-sensitive processes, not just general-purpose productivity gains.

Workforce implications

Goldman has not framed the project as a headcount reduction initiative. Argenti told CNBC it was "premature" to expect the technology to lead to job losses for the thousands of staff currently working in the affected functions.

However, broader signals from the bank point in a more measured direction. CEO David Solomon said in October that Goldman would constrain headcount growth as part of a multiyear AI reorganization, even as trading and advisory revenue continues to grow. And Argenti acknowledged that Goldman could eventually cut out some third-party providers it currently relies on as the technology matures. "It's always a trade-off," he said.

The distinction worth watching is between job displacement and job absorption. If AI agents reduce the volume of routine work without eliminating roles, the practical effect may be slower hiring rather than layoffs, at least in the near term.

What this signals for enterprise AI

Goldman is not the first bank to experiment with large language models, but the placement of this project is significant. Most enterprise AI deployments to date have operated at the edges of workflows: drafting content, summarizing research, assisting developers. Building agents for trade accounting and compliance puts AI closer to processes that carry direct financial and regulatory consequences.

That Goldman chose to embed Anthropic engineers inside its own teams, rather than deploy an off-the-shelf product, reflects a broader pattern in how regulated industries are approaching AI adoption. Banks rarely automate without deep customization, because they need controls, audit trails, and clear accountability. Co-development is the model that allows for that.

If the agents perform reliably once launched, the project could serve as a reference point for other institutions weighing whether large language models are ready to operate inside audit-intensive systems. The question facing enterprise AI has shifted from whether models can produce useful outputs to whether organizations are willing to place them inside workflows where errors have real consequences.

Goldman's answer, at least in early form, appears to be yes.