Goldman Sachs AI Assistant: Why the Bank Built Its Own and What It Changes

GS AI Assistant is Goldman Sachs's in-house generative AI tool: a chat window, built on the bank's own AI platform, that lets employees put questions to models from OpenAI, Google and Anthropic without client or firm data leaving Goldman's systems. It reached about 10,000 people by January 2025 and every employee in June 2025.

For a junior banker, it changes where the work starts. The first draft of a summary, a list of comparable companies or a translated research note arrives in minutes, and the job becomes describing the task precisely, then checking what comes back before anyone relies on it.

June 2025

Firmwide launch

after a pilot of about 10,000 staff

1 million+

Prompts in July 2025

a record month

GPT, Gemini, Claude

Model families on offer

staff can choose

12,000

Developers with AI coding tools

by July 2025

What is GS AI Assistant?

Goldman's own definition is exact. Marco Argenti, the chief information officer behind it, calls GS AI Assistant "an internal natural language conversational application that enables end users to access firm-approved Large Language Models (LLMs) in a safe and secure way." In plainer terms, it looks like ChatGPT, but Goldman chooses the models behind it, wraps them in its own controls, and lets their answers draw on Goldman's data.

Two layers are worth keeping apart. The GS AI Platform, which debuted in 2024, is the plumbing: guardrails, connections to the firm's data, and access to outside models. GS AI Assistant is the application most employees see on top of it, and several narrower Goldman tools run on the same platform.

When Goldman opened it to everyone in June 2025, the bank's next earnings call described it as "the first generative AI-powered tool to reach the scale allowing for safe, secure, and responsible access to firm-approved external large language models."

What does GS AI Assistant change for a junior banker?

Take a classic analyst request: find the comparable companies for a client, pull their filings, and rank them by exposure to a particular risk. Argenti used almost exactly that example to explain how the assistant handles a hard question. A reasoning model breaks the request down, looks for plausible comparables, fetches the filings and ranks them, and, in his words, "the user doesn't mind waiting 10 minutes for an answer."

Ten minutes for a first pass is the whole change in one line. The work does not disappear. It moves to the two ends of the task: the brief that goes in, and the judgment applied to what comes out.

Before the assistant

• Build the peer list from a blank spreadsheet
• Write the client email from scratch
• Read the whole document to find the three facts that matter

With GS AI Assistant

• Brief a reasoning model, then audit its ranked list
• Edit a first draft the tool produced
• Check the summary against the source before it travels

Describe, delegate, supervise

Argenti has spelled out what this asks of people at the bottom of the organization. In a July 2025 essay, he wrote that agentic AI will require "even the most junior employees and individual contributors to master three foundational management skills: describing a task clearly, delegating it effectively to an AI agent, and supervising the results."

Two months later he ranked them. Supervision is "the most critical," he said, because handing work to an AI "without understanding what the A.I. produces, and not being able to critique and correct, is a recipe for failure."

That is a manager's job description handed to a first-year analyst, and Argenti draws a hopeful conclusion from it: the shift makes early-career workers "more critical than ever," because they have grown up alongside the technology. His analogy is an airline that fills flights with junior pilots because the autopilot can land the plane. That is safe only if the juniors are properly equipped to supervise.

Checked like anyone else's work

Goldman's bar for trusting the output is not perfection. Argenti's test is whether an AI is "at least as precise as a human" doing the same job, and he wants AI put through the same scrutiny and controls as people: code reviews for code, human reviewers for documents. For a junior, the checking step is not a courtesy. It is the control.

Why did Goldman build its own AI assistant?

Goldman built rather than bought for three reasons: to keep the data inside, to stay free to choose the model, and to hold the output to the firm's own standard.

Keep the data inside

Staff using the assistant talk to large language models "firewalled within Goldman," as Fortune described the launch, so sensitive information does not escape to an outside provider. Argenti set three conditions before opening it to the whole firm:

• Data he trusted. "There's no quality AI without good quality data," he said.
• Use cases that clearly added value, tested by pilot users at every level, from engineers and analysts to partners.
• Safety, meaning "protecting all information, reducing hallucinations, and protecting from potential new vectors of attack."

Stay model independent

The platform's founding tenets, as Argenti set them out in September 2025, put "Be model independent" second, right after safe and compliant access.

Staff can pick the model suited to the task in front of them, and Goldman can fine-tune models on its own data in a controlled way instead of building one of its own. The commercial logic is simple: one model may be better at coding while a rival reasons better, and they differ in cost and in how easily they can be modified.

Teach it the Goldman way

The third reason is cultural. "For the AI to have a very specific identity that reflects the tenets, the values, the knowledge and the way of thinking of the firm is extremely important," Argenti said in January 2025. The aim is an assistant that picks up the habits of an experienced Goldman employee, such as checking a number against more than one data source.

What does GS AI Assistant do day to day?

With the controls in place, the everyday uses are mostly unglamorous. The launch memo put the core in a single line: "summarizing complex documents and drafting initial content to performing data analysis." Around that core, the uses described since fall into a handful of types.

Business Insider spoke to seven Goldman employees, from analyst to partner, who described the assistant as part of daily life at the firm. One engineering associate said he used it every day "for getting a head start on traditionally time-consuming tasks," and that it was "saving me hours every week."

It is not one identical tool for everyone. At the firmwide launch, Fox Business reported that developers, investment bankers, research analysts and asset and wealth management staff each had a version tailored to their work. And to win over non-technical staff, Goldman recruited champions from asset and wealth management, private banking and trading rather than from its technology teams.

Banker Copilot, Devin and Goldman's other AI tools

Those tailored versions sit inside a wider family of tools, most of them built on the same platform.

Banker Copilot

Banker Copilot is the tool closest to an investment banking analyst's desk. It gives eligible bankers data on clients and deals that only authorized users can see, analyzes corporate filings "for hard-to-find cues that bankers would want to know about," and drafts long documents. Through 2025 it was a pilot: a group numbering in the dozens had it in May, and in September Argenti said it was "not scaled" and still being refined before any wider rollout.

From coding assistants to Devin

Engineers were first in line, and roughly one in four Goldman employees is one. They were the first to get AI coding assistants, including GitHub Copilot and Gemini Code Assist, and Goldman measures the return by how often the tools are used and how much of the suggested code engineers accept. In July 2025, Argenti put the result at developers being "on average, 20% more productive."

That same month, Goldman became the first major bank to use Devin, Cognition's autonomous software engineer, according to Cognition. Argenti called it "our new employee who's going to start doing stuff on the behalf of our developers." Its work is the drudgery engineers would rather not do:

• migrating a code base from Java 8 to Java 21
• finding every dependency a new framework touches, and updating it
• moving code from one programming language to another

Each change arrives as a merge request that a human reviews, then passes the same CI/CD controls as any developer's code before it reaches production. By February 2026, Devin was broadly available to Goldman's engineers.

Agents for accounting and onboarding

The next agents work outside engineering. In February 2026, Argenti said Anthropic engineers had spent six months embedded at Goldman, building Claude-based agents for two jobs: accounting for trades and transactions, and vetting and onboarding clients. He called them "a digital co-worker for many of the professions within the firm that are scaled, are complex and very process intensive," and said Goldman was still "in the early stages."

The firm was "surprised," he said, at how capable Claude was at tasks besides coding, in accounting and compliance work that means parsing large amounts of data and documents while applying rules and judgment. Investment banking pitchbooks are among the tasks he named as possible next targets.

Which AI models does GS AI Assistant use?

Goldman described the models behind the assistant four times in 2025.

Staff can choose a model themselves, and Goldman runs automated tests that weigh accuracy against cost, then favors one model over another for a given kind of user or question. That is why a research request like the comparable-companies example above goes to a slower, more expensive reasoning model, while a quick question, such as what an acronym means, goes to a cheaper one, and a presentation goes to a model that can generate charts and images.

In April 2026, Reuters and the Financial Times reported that Goldman had withdrawn Claude from its bankers in Hong Kong after a strict reading of its contract with Anthropic, while Gemini and ChatGPT stayed on the platform. That is the design working: one model left the menu for one office, and the assistant carried on with the others.

How many Goldman employees use GS AI Assistant?

That menu reached the whole firm quickly. Goldman's dated figures measure different things: who could open the tool, and how much people used it.

Access went from about 10,000 people to the whole firm in about five months.

January 2025, pilot

about 10,000

June 2025, firmwide launch

about 46,000

Both bars count access. The June bar is the size of Goldman's workforce at the time, about 46,000, all of whom got the assistant at the firmwide launch.

July 2025 was a record month for usage, with more than a million prompts, and the number of questions per employee was still climbing. The growth came from use as well as adoption: people who already had the tool were using it more.

What Goldman's executives have said about AI and jobs

Those usage numbers raise the obvious question of what happens to the people doing the prompting. Goldman's leaders have framed the tools the same way throughout: as added capacity, with the firm's advantage still in its people.

David Solomon, Goldman's chief executive, drew the most useful distinction on the bank's January 2026 earnings call. One kind of change, he said, is handing "very smart, very productive people" the models and tools and letting them work out day by day how to be more productive. That is GS AI Assistant's job.

The other kind is redesigning how the firm operates. That is One Goldman Sachs 3.0, the operating model "propelled by AI" set out in Goldman's 2025 annual letter, starting with six workstreams, among them client onboarding, lending and regulatory reporting.

What it means for headcount and analyst classes is on the page asking whether AI will replace investment banking analysts, including Solomon's view that Goldman will probably start fewer people over the next few years, "but probably not what you and I would call dramatically."

What should a candidate know about GS AI Assistant?

Nobody outside Goldman can open GS AI Assistant, so what counts is describing it accurately and showing the habits it rewards: a precise brief going in, and a skeptical check coming out.

What to practice

1. Run a real finance task, such as summarizing an earnings call transcript, through ChatGPT or Gemini, then check every figure against the source and note how long the checking took compared with the drafting.
2. Keep one written example of an AI mistake you caught, whether a stale price, a wrong share count or an invented source, and how you caught it.
3. Practice writing briefs: turn a vague request into a prompt that states the output format, the sources to use and the checks to run.
4. Read what Goldman's executives have said on the record, so any answer about AI at Goldman rests on their words rather than a headline.

Solomon's advice to Goldman's 2026 interns makes the same point in one line: "you have to challenge what the models give you." The interview readiness tool walks through preparing those answers, and the wider guides cover what other banks built, including JPMorgan's LLM Suite.

For engineers moving into finance

Engineers get a sharper version of the same brief. The GS AI Platform lets teams build applications on top of it with guardrails, data access and compliance already in place, so developers do not start from scratch with every application.

The coding agents change the job from writing every line to directing the work. Argenti expects engineers "to have the ability to really describe problems in a coherent way and turn it into prompts … and then be able to supervise the work of those agents." Staffing a project, as he put it, now means deciding "I'm going to need X amount of humans, X amount of virtuals."

His July 2025 essay adds a product mindset: "To be a great engineer now means also being a great product manager: understanding the customer, defining the roadmap, identifying risks and biases, and designing compensating controls."

The bottom line

GS AI Assistant is Goldman's way of giving every employee frontier AI models without giving up control of the data, the choice of model or the standard the output must meet. Built on a platform the bank owns, it offers staff a choice of outside models, with Goldman's own tests deciding which to favor for each kind of question, and it has been in every employee's hands since June 2025.

For a junior banker, that means the first draft is no longer the job. Describing the task well, and catching what the model got wrong before anyone relies on it, is the job, and it is a skill a candidate can start building long before a first day at Goldman.