Faison

Goldman Sachs just gave the AI industry a reality check. After months of allowing bankers and traders to experiment with advanced AI coding tools, including Anthropic’s Claude Code and Microsoft’s GitHub tools, Goldman is tightening access. The new October guidance largely restricts these tools to engineers. Quantitative specialists, known as strats, are among those who retain access. Why does this matter? Traders had been using AI coding agents to monitor client trading flows, model market movements, analyze positions, and generate written and visual materials. In other words, people who don’t traditionally write software were starting to build their own tools. Some weren’t happy when access was restricted. Goldman’s explanation was straightforward. The firm had completed an experimentation period and was evaluating productivity alongside risk management. It now wants these tools in the hands of people best equipped to use them over the long term. But here’s where the story gets interesting. Goldman isn’t abandoning AI. Its internal GS AI Assistant is reportedly being used weekly by approximately 40,000 of its 46,200 employees. The firm has also reported productivity improvements of at least 20% among its roughly 12,000 engineers using AI coding tools. CIO Marco Argenti had even praised nontechnical employees experimenting with so-called vibe coding earlier this year. Now the firm is pulling back on access to some of those same tools. There is a bigger lesson here for the AI industry. Getting employees to experiment with AI is relatively easy. Turning experimentation into reliable production software is a different problem. Banks have proprietary data, regulatory obligations, compliance requirements, cybersecurity risks, and accountability standards. A coding agent that works brilliantly in a demonstration still has to survive those realities. And there’s another question. If traders and bankers can use AI to build tools that previously required engineers, where does that leave the traditional technology department? Goldman appears to be drawing a line between experimentation and production. That may be a sensible decision. But it also exposes the gap between what AI vendors promise and what large institutions are actually comfortable deploying. The AI revolution isn’t being stopped. It’s being given a risk committee, an approval process, and restricted access. Welcome to Wall Street.

Software is dead It's now the era of personal software and everyone just builds what they need whenever they need it I think