Skip to content
SpendByTeam

Journal Governance

Goldman Sachs restricts Claude Code: how companies control AI coding tools

Goldman Sachs limited Claude Code to engineers. What was reported, why companies restrict AI coding tools, and the controls a team can put in place instead of a ban.

6 min read

In early November 2026, Bloomberg reported that Goldman Sachs had limited Claude Code and GitHub’s AI coding tools to its engineers. Most traders and bankers lost access. The story spread quickly because it looked like a big bank stepping back from AI coding.

It’s more useful to read it as a governance decision. Any company rolling out Claude Code ends up answering the same questions Goldman did: who gets it, what it may touch, and what it costs. This post covers what was reported and the controls you can use before you reach for a ban.

What was reported

Based on coverage of the Bloomberg report (Yahoo Finance, PYMNTS, Quartz):

  • Internal guidance updated in October 2026 limited advanced AI coding tools, including Claude Code and GitHub’s tools, to engineering staff — reportedly about 12,000 engineers.
  • Before that, employees across the business had been able to use them for several months.
  • One report says “strats” — hybrid trading, quant and engineering roles — kept access.
  • The bank reportedly cited a review of productivity and risk. A spokesperson described it as a routine step rather than a reversal. The review’s findings weren’t published.
  • Goldman’s in-house AI tools reportedly stayed broadly available.

Most coverage traces back to a single Bloomberg article, so treat the details above as reported, not confirmed.

Why companies restrict AI coding tools

When a company narrows access, the reasons usually fall into three groups:

  1. Risk. A coding agent can read files, run commands and call networks on the user’s machine. Outside engineering, people are less likely to understand what they’re approving.
  2. Value. Seats and usage cost money. If non-engineers use a coding agent for spreadsheets and emails, the company is paying coding-agent prices for work a chat assistant could do.
  3. Visibility. Often the real trigger is that nobody can answer “who is using it, on what, and what does it cost?” When you can’t measure it, restricting it is the easy option.

The first is a permissions problem. The second and third are measurement problems. Each has a fix that’s lighter than taking the tool away.

Controls that work before a ban

1. Central settings for permissions

Claude Code supports managed settings: a settings file deployed by IT that users can’t override. It can deny specific tools and commands, block bypass-permission mode and set which actions always need approval. That addresses most of the risk argument without removing access.

The “don’t ask again” option is worth a policy of its own. When a developer approves a command permanently, it runs without prompting from then on. Decide which commands are fine to auto-approve, put them in managed settings, and deny the rest centrally instead of leaving it to each person’s judgment.

2. Access by role, not all-or-nothing

Goldman’s reported split — engineers yes, bankers no, strats yes — is a role-based rollout. You can do the same on a smaller scale: give Claude Code to people who write and review code, and a chat assistant to everyone else.

3. Spend visibility by team and project

This is the control most companies skip, and the one that makes the others easier to defend. Claude Code reports cost and token usage per session through OpenTelemetry. With a repository attribute attached, you can answer:

  • Which teams and projects use it most?
  • Which sessions are outliers?
  • Is a seat worth it for this person?

In our own usage of Claude Code, the most expensive 10% of sessions made up 59% of the cost, and the cheapest 75% made up 15% (see the full breakdown). Without per-session data, those outliers look like “AI is expensive” — and that’s how blanket restrictions get approved.

4. Budgets and alerts instead of shut-offs

A budget per team or project, with an alert at 80% and 100%, turns surprise into a conversation. The team lead sees the number before finance does.

A simple rollout policy

If you’re writing your own policy, this covers most of it:

Area Rule
Who Engineers and code reviewers. Others by request.
Permissions Managed settings: deny network and destructive commands by default, block bypass mode.
Data Metrics only. No prompts or code leave developer machines for monitoring.
Cost Spend tracked per team and project. Monthly budget with alerts.
Review Quarterly: usage by team, top sessions, seat value.

Where SpendByTeam fits

SpendByTeam handles the visibility part: it reads Claude Code’s cost metrics — never code or prompts — and shows spend by repository, project, team and client. Spend by team and project is free for up to 3 developers; budget alerts are on the Team plan. It doesn’t replace managed settings for permissions — use both.