Writing

Goldman Sachs just took Claude Code away from about 35,000 of its own people.

By ·

In July their CIO was still bragging about the tools traders built with it.

So it looks like a bank giving up on AI. I think it's the smartest AI call a bank has made this year.

The traders used it to track client flows and to model their positions. That's useful. But a coding agent runs what it writes with the same access as you, so a script a trader made can end up running every day on a machine full of client data. And nobody ever looked at it.

The agents they kept are the ones for client onboarding. Those are run by their own engineers (Anthropic even had people at the bank for six months to build them).

Building with AI is quick now, but getting it live at a bank or an insurer isn't. You have compliance, you have all the bureaucracy, and that's where the time goes.

At my previous employer the first version of our AI coverage check took about 3 months. Then making it run in real time was another 3 months on the roadmap, and most of that was about rules and data.

So if you want to give your ops team agents, sort this out first. Who owns what the agent builds? What can it touch? And who checks it before it hits a customer's file? (No, "the person who built it" doesn't count.)

I think most companies that gave everyone an agent this year will end up doing the same as Goldman.