Goldman limits Claude Code and Copilot to engineers: how to do it

Goldman Sachs has taken Anthropic's Claude Code and Microsoft's GitHub coding tools away from its bankers and traders and left them with engineers, Bloomberg reported on Monday, November 2, citing internal guidance issued in October. The Decoder and PYMNTS, reporting on the Bloomberg story, say bankers and traders spent months using the tools to watch trading flows and to produce content. Techstrong says designated software engineers and the quantitative "strats" who sit between trading and development keep access. Goldman has 12,000 engineers.

Spokesperson Nick Carcaterra told Bloomberg the firm offers some AI tools to staff so it can find where they work best, and that after a trial period it weighs factors including productivity and risk management before deciding who keeps them. In July, chief information officer Marco Argenti had praised "vibe coding" by staff outside engineering, who built prototypes and handed them to developers to finish. According to Techstrong, Goldman still runs its own in-house assistant built on models from OpenAI, Alphabet and Anthropic. Goldman is also the lead underwriter for Anthropic's planned share sale.

Why it matters to IT buyers

Techstrong puts the risk plainly: a coding agent runs with the permissions of the person using it. On a trading desk that raises questions about material non-public information, information barriers and who reviews code that no engineer wrote.

Goldman's change is narrower than JPMorgan's, where chief executive Jamie Dimon said in October that outside AI agents will not be allowed inside the bank, according to PYMNTS. Goldman keeps the tools, but only for people who write software.

Limiting access by role depends on how each vendor sells seats.

GitHub Copilot: license by team where you can

GitHub's documentation gives an enterprise two ways to hand out Copilot. The first is to assign licenses directly to users or enterprise teams. GitHub says this route is currently available only for Copilot Business. The second is to switch Copilot on for whole organizations and let each organization's owners hand out licenses to their own members.

For a Goldman-style split on Copilot Business, the first route is the one to use. Copilot Enterprise licenses can only be handed out through organizations, so on that plan the split has to follow organization lines. Under Billing and licensing, open Licensing, click Manage next to Copilot, choose the Enterprise Teams tab and assign licenses to a team such as "engineering." GitHub says access then follows team membership: people gain or lose Copilot as they join or leave the team. With Enterprise Managed Users, that team can be synced to a group in your identity provider, so the HR system that knows who is an engineer also decides who has Copilot. Before you start, set the "Policies for enterprise-assigned users" policy, which GitHub lists as a prerequisite.

Then close the side door. Under "Organization access" you can turn Copilot off for organizations, which stops organization owners from assigning licenses on their own. Skip this, and the owner of any organization where Copilot is enabled can still hand licenses to its members. GitHub's page does not say whether turning it off removes licenses already given out that way, so check the seat list afterward.

Licensing decides who has Copilot at all. Policies, on the enterprise "AI controls" tab, decide what licensed users can do. GitHub lists policies for Copilot CLI and for model access, and a policy that can block Copilot in the cloud for "all users in your enterprise's repositories," licensed or not. Two details matter. Users licensed directly by the enterprise are not covered by the "Let organizations decide" option; the "Policies for enterprise-assigned users" setting sets their defaults, so set the enterprise-level policies deliberately. Where a user gets Copilot from several organizations in one enterprise and they set a policy differently, GitHub says the least restrictive one usually applies. And GitHub says not every policy applies on every surface, such as IDEs, the website and the CLI, so test each one where people actually work.

Claude Code: it depends on your seat type

Anthropic handles access at the Claude account level. Its admin guide says SSO, SCIM provisioning and seat assignment all live there, and that a separate layer, managed settings, controls what Claude Code may do once someone has it.

Whether you can withhold Claude Code from some staff depends on which plan you bought. Anthropic's help center says every Team seat includes Claude Code. New and self-serve Enterprise plans also include it in the single Enterprise seat. Older Enterprise plans split seats: Chat-only seats leave it out, while Chat + Claude Code and Premium seats include it. On those plans the help center says an owner buys or reassigns Claude Code seats under Organization settings, then Organization. Anthropic's Enterprise setup tutorial describes its tiers as Standard, without Claude Code, and Premium, with it, and says seat assignments are limited to primary owners and owners; its Settings, Members, "View tier breakdown" screen shows how many seats of each tier are free. SCIM provisioning hands out seats by tier availability. Anthropic's admin guide says a developer whose seat lacks Claude Code gets an error saying they have not been added to the organization.

On a Team plan, or a newer single-seat Enterprise plan, the help center article does not describe a per-user switch for Claude Code. Ask your Anthropic account team what control applies before you assume a seat change will do it. Until you have that answer, the managed settings below are what you can enforce.

What managed settings can enforce

Managed settings override a developer's own configuration, with a few security-related exceptions Anthropic lists. They can come from the Claude admin console (Team or Enterprise plans), from MDM through a macOS plist or the Windows HKLM registry key, or from a managed-settings.json file. Anthropic warns that the Windows per-user registry key under HKCU can be written without admin rights, so do not rely on it for enforcement.

The controls that matter for a role split:

• forceLoginMethod and forceLoginOrgUUID pin Claude Code to a sign-in method and to your company's organization. Claude Code checks the organization for claude.ai sign-ins in the terminal, the VS Code extension and the Agent SDK, which blocks personal claude.ai accounts there, though not Claude Console logins.
• permissions.deny blocks specific tools and commands. Developers can add to a managed deny list but cannot remove entries from it. allowManagedPermissionRulesOnly makes the managed rules the only ones that count, and permissions.disableBypassPermissionsMode turns off the flag that skips permission prompts.
• sandbox.enabled with sandbox.network.allowedDomains limits which hosts the agent can reach. Anthropic points out that denying the WebFetch tool alone does nothing if Bash is allowed, because curl and wget still work. The sandbox closes that gap.
• allowedMcpServers, deniedMcpServers and allowManagedMcpServersOnly control which outside data sources Claude Code can connect to.
• availableModels with enforceAvailableModels restricts models.

For visibility, Claude Code exports sessions, tool use and tokens over OpenTelemetry on every provider. Team and Enterprise plans have spend limits in organization settings, and Anthropic's self-hosted Claude apps gateway keeps a per-request audit log tied to the user's identity-provider account. To confirm a policy has landed, run /status and look for "Enterprise managed settings" on the Setting sources line.

What IT teams should do

1. Keep the list of who counts as an engineer in your identity provider, not in each tool.
2. In GitHub, license Copilot to an enterprise team synced to that group, then turn off organization-level assignment.
3. Set the enterprise-level Copilot policies and the "Policies for enterprise-assigned users" default, since organization settings do not govern users licensed by the enterprise.
4. In Claude, find out which seat type you hold. On older Enterprise plans, move non-engineers to Chat-only seats. On Team or single-seat Enterprise plans, ask Anthropic how to withhold Claude Code from a group.
5. Push managed settings through the admin console or MDM, with forceLoginOrgUUID, a sandbox allowlist and an MCP allowlist, and check /status on a test machine.
6. Tell the people losing access what replaces it, whether an internal assistant or a request to engineering. Techstrong reports that some Goldman staff now have work held up, with nothing in-house yet to replace the tools.

Related: the backdoored Claude Code extension in Cursor and Copilot Business metering.

Sources: The Decoder, Techstrong.ai, PYMNTS, GitHub Docs: granting Copilot access, GitHub Docs: Copilot policies, Claude Code Docs: admin setup, Claude Help Center: Claude Code on Team and Enterprise, Claude: enabling Claude Code for an Enterprise team.