Goldman Sachs Restricts Claude Code to Engineers Only, a Bellwether for AI Deployment in Banking

According to Bloomberg, Goldman Sachs restricted access to Claude Code and GitHub to engineers only last October. On the surface this looks like a minor internal IT policy tweak, but it reflects the core tension regulated industries face when deploying generative AI tools: the tug-of-war between the lure of efficiency and the need for risk control.

What Happened: A Low-Key Permission Rollback

Goldman Sachs' access restrictions on Claude Code and GitHub took effect last October and were first reported by Bloomberg. The report did not detail the exact scope of the restrictions, but it confirmed two key facts: the restricted tools include Anthropic's coding assistant Claude Code and the GitHub platform; the restricted group is non-engineer employees.

This means Goldman's approach is to let engineers—who have code review skills and technical judgment—use AI coding tools, while non-technical staff are temporarily excluded. This is not an outright ban but a selective approach to permission management.

Claude Code is an agentic coding tool from Anthropic, distinct from traditional code-completion plugins. It runs in the terminal and can autonomously read the context of an entire code repository, plan multi-step operations, directly modify multiple files, and execute shell commands. Once authorized, it can even complete the full development task from locating a problem to committing code. This "agentic" capability means it can reach a far broader range of internal assets than a single-file editor, including private repositories, build scripts, environment variables, and internal API credentials. For financial institutions, the more capable the tool, the greater the need to narrow permissions.

Why Banks Are Especially Cautious About AI Coding Tools

The Risk Surface Beyond Code

The capability boundary of tools like Claude Code has long exceeded "completing a line of code." It can read repository context, execute commands, modify multiple files, and in agent mode even autonomously complete development tasks. For banks, this means AI tools may come into contact with sensitive assets such as core banking systems, customer data processing logic, and internal API keys.

When non-engineer employees use such tools, they lack the judgment to assess the risks of code changes and struggle to identify security vulnerabilities that AI-generated code may introduce. Goldman's restriction essentially draws a line on the question of "who can safely use this tool."

Hard Constraints of Regulatory Compliance

The financial industry is bound by multiple regulatory frameworks with explicit requirements for data retention, model explainability, and third-party tool auditing. When deploying generative AI tools in banking, compliance teams need to answer questions such as: Will the tool send code or data externally? Can logs be fully retained? Is model behavior reproducible?

By narrowing usage to engineers, Goldman can build audit experience on a relatively controllable sample before deciding whether to expand. This "narrow first, broad later" path is not uncommon in regulated industries.

Data residency refers to the requirement that data be stored physically or logically within a specific jurisdiction. Financial regulators typically require that customer data and critical system logs not be transferred across borders at will, while inference for most cloud-based AI tools happens on vendor servers, potentially involving cross-border data flows. As a result, when evaluating generative AI tools, banks focus on whether regional deployment, local inference, or private cloud options are available. Insufficient data residency capability often directly causes a tool to fail compliance review—another threshold beyond permission granularity.

Implications for Regulated Teams

Goldman's approach offers a reference template. For teams in similarly heavily regulated environments, the following dimensions are worth including in their own AI tool launch checklist:

Role-based tiered authorization. Define AI tool usage levels by employees' technical capabilities and job responsibilities rather than opening or banning access across the board. Engineers, data analysts, and business staff can be matched to different tool capabilities and data access scopes.

Define data boundaries upfront. Before launch, specify which repositories and data directories AI tools are not allowed to access, and enforce this through technical means rather than policy alone.

Full audit log coverage. Ensure all AI tool operations are traceable, including session records, code changes, and command execution history, to meet compliance review needs.

Gradual expansion mechanism. Set clear evaluation checkpoints and, after accumulating sufficient usage data and risk cases among engineers, assess whether to open access to other roles.

The Signal Matters More Than the Event Itself

The news that Goldman Sachs restricted Claude Code is noteworthy not because the restriction itself is particularly strict, but because it comes from a top global investment bank. The AI deployment strategies of leading financial institutions are often referenced by peers. Goldman's decision to tighten permission boundaries as early as last October shows that the cautious attitude of regulated industries toward generative AI tools is shifting from "whether to use them" to "how to use them safely."

For AI tool vendors, this is also a reminder: in heavily regulated sectors such as finance, healthcare, and law, beyond product capabilities, permission granularity, auditability, and data residency options are becoming key factors determining whether a tool can enter enterprise procurement lists. Whether Claude Code can expand from engineers to a broader scope within Goldman Sachs will depend on whether these supporting capabilities can keep pace.

One-line Summary

Goldman Sachs restricts Claude Code to engineers only, reflecting the efficiency-risk balance of AI deployment in regulated industries.

Paragraph Summary

Last October, Goldman Sachs restricted Claude Code and GitHub access to engineers only, as first reported by Bloomberg. This low-key permission rollback reflects the core tension in regulated industries' deployment of generative AI tools: the tug-of-war between the lure of efficiency and risk control. The restriction stems from the fact that AI coding tools' capabilities now go beyond code completion and can reach sensitive assets such as core banking systems, customer data processing logic, and API keys, while non-technical staff lack the judgment to assess code risks. Hard constraints from financial regulation on data retention, model explainability, and third-party tool auditing have also prompted Goldman to adopt a "narrow first, broad later" strategy. The article distills a launch checklist across four dimensions—role-based tiered authorization, upfront data boundary definition, audit log coverage, and gradual expansion—and notes that the cautious strategies of leading investment banks are a bellwether: permission granularity and auditability are becoming key factors for AI tools entering enterprise procurement.