The Signal #130 — Dakota’s read on the AI news that actually matters to people running a business.

A big bank gave AI coding tools to everyone. Then it took them back. The story behind that reversal is more useful to operators than the headline makes it sound.

What happened

According to reporting from Bloomberg, covered by State of AI, Goldman Sachs updated internal guidance in October that restricts Anthropic’s Claude Code and Microsoft’s GitHub Copilot to the bank’s engineering staff. Traders and bankers, who had been using those tools to monitor client trading flows, model market positions, and produce written and graphic materials, no longer have access.

The carve-out is specific. A group called Strats, who sit across trading, quantitative research, and software engineering, still have access. Everyone else in those business functions does not.

This follows a pattern. Goldman had already removed Claude from its internal platform for bankers in Hong Kong back in April. That restriction was based on contract terms and location. The October restriction is based on role.

The timing has a certain irony to it. As recently as July, Goldman’s chief information officer Marco Argenti was publicly praising the arrangement where traders and bankers built first-version tools with AI and handed them to engineers to be improved. He called it a natural fit, letting the people closest to a problem build something without waiting for engineering bandwidth. Goldman employs roughly 12,000 engineers out of about 47,000 total staff. The October guidance pulls the coding agents back to that group.

Goldman’s in-house AI assistant, which draws on models from multiple companies, is still available firm-wide. The restriction applies specifically to third-party agents that can write and run code on their own.

Why it matters

The article makes the core risk plain: Claude Code reads files, writes programs, and runs them, typically with the same level of access as the person using it. Code that an engineer writes with it usually goes through review, testing, and deployment controls before it touches anything sensitive. A script a trader builds to watch client flows can end up in daily use without passing through any of those checkpoints.

That gap is the problem. And it is not a Goldman-specific problem.

Consider a real estate brokerage where agents start using an AI coding tool to pull and slice MLS data. Or a healthcare administration office where billing staff use one to automate claim tracking. Or an e-commerce team where merchandisers build scripts to watch competitor pricing. In each case, the people closest to the data are building tools that touch sensitive systems, and no one with engineering oversight has looked at what those tools actually do.

Goldman also noted a cost dimension. Agent usage is billed by the token (a token being a small chunk of text the AI reads or writes, and what most AI services use to calculate your bill). Several companies that opened agent access broadly watched spending climb fast. Amazon shut down an internal leaderboard after employees ran agents on meaningless tasks just to rank higher on it. Accenture, according to the article, now funds 29 dollars a month of Claude usage per employee as a cap. Uncapped agent access across a large staff is a budget line that can move quickly.

What most people get wrong

Most operators read a story like this and assume it is a regulated-industry problem. Banks have compliance rules, information barriers between deal teams and trading desks, audit requirements. So the lesson must not apply to them.

That framing misses the operational part of the story, which has nothing to do with financial regulation.

The risk Goldman identified is simpler than compliance. It is that people without engineering training were running software that writes and executes code on machines that also hold client information. The article notes that last month a Meta engineer said he now ships code he cannot fully account for, and writing software is his job. A trader using an agent to model positions is less likely to understand how that code handles edge cases. On a trading desk, an edge case can mean losses. In an agency, it might mean a client’s data going somewhere it should not. In a SaaS company, it might mean a script quietly breaking a downstream process nobody monitors.

The question is not whether your industry has regulators watching. The question is whether anyone in your organization understands what the AI-generated code is actually doing before it runs on something that matters.

Goldman’s answer, for now, is to put that responsibility back with the people trained to carry it.

The lesson

Broad access to capable tools is not the same as a plan. Goldman gave traders and bankers AI coding tools, saw real productivity from it, then pulled access back after a risk review. That is not a failure. That is what governing a new capability actually looks like in practice.

For operators, the useful question is not whether to restrict AI agents. It is whether you have thought through who is running them, what systems those agents can touch, and whether anyone qualified has seen what they produce before it becomes a daily process.

Access policies are not a one-time decision. Goldman’s moved twice in the same year on this. Yours probably will too.

If you are working through where to draw those lines in your own operation, the resources at xovionlabs.com are a reasonable place to start thinking it through.