Goldman Sachs took AI coding agents from its traders. Copy the policy, not the panic

In July, Goldman Sachs' chief information officer was praising traders and bankers who built their own tools with AI. By October, most of them had lost access to the AI coding agents that made it possible. Internal guidance now restricts Anthropic's Claude Code and Microsoft's GitHub to the bank's engineers, Bloomberg reported, as summarised by State of AI.

It is easy to read this as a story about a big bank being cautious. It is more useful as a template for any company deciding who should use which AI tools.

The short version

• Goldman did not ban AI. It separated chat assistants, open to everyone, from agents that write and run code, now limited to engineers.
• The risk is not the model. It is unreviewed code running with an employee's full access to files and client data.
• Write your AI policy by role and by what the tool can touch, before your own staff decide it for you.

What Goldman actually changed

According to the report, traders had been using the tools to monitor client trading flows, model market moves and positions, and produce written and graphic material. Strats, who work between trading, research and engineering, kept access. Goldman's in-house AI assistant, which uses models from several companies, is still available across the firm.

So the line is not "AI yes or no". The bank's roughly 12,000 developers, out of about 47,000 staff, can use agents that write and execute code. Everyone else keeps an assistant that answers and drafts, but doesn't run programs on their machine.

A spokesperson said the bank makes AI tools available to staff "to determine the best use cases". In other words: try widely, then decide.

Why a coding agent is different from a chatbot

A coding agent reads files, writes programs and runs them, usually with the same access as the person using it. When an engineer writes code that way, it normally goes through review and testing before it touches anything important.

When a trader builds a script to watch client flows, it can end up in daily use without any of those checks. Nobody may fully understand how it handles unusual cases. On a trading desk, that can mean losses. In your company, it might mean a spreadsheet macro that quietly emails the wrong client list, or an automation that reads every HR folder because its creator could.

The cost question nobody budgets

Agents are billed by usage, and the report notes that several companies that handed them to everyone saw spending climb. Amazon shut down an internal leaderboard after employees ran agents on pointless tasks to rank higher.

Here is an illustrative budget. A 40-person company gives an AI agent subscription to 30 people at around $29 a month each, the figure the report cites for Accenture's per-employee allowance. That is $870 a month, or more than $10,000 a year, before anyone has measured what it brought. Giving it to the 8 people with clear, repeated use cases costs $232 a month and is much easier to review.

A three-line policy you can write this week

• Everyone: an approved chat assistant for writing, summarising and research, with a clear rule on which client data may be pasted into it.
• Selected roles: agents that automate tasks, for people with a defined use case and someone who reviews what they build.
• Engineers only: agents that write and run code on company systems, inside the normal review and testing process.

Then decide what each tool is allowed to read. An agent can only leak or break what it has access to, so the safest AI policy starts with tidy permissions on your documents. That is where Holol-GED helps: secure, searchable document management, so contracts, HR files and client records sit behind clear access rules instead of a shared drive anyone's tools can read.

Try widely, then draw lines

Goldman let thousands of people experiment, learned where the tools paid off, and then drew lines by role. That is a sensible order for a 30-person company too. Just don't skip the second step.

Sources

• State of AI, citing Bloomberg: Goldman Sachs Is Pulling Claude Code From Its Bankers and Traders (Hacker News discussion)