Goldman’s AI clampdown is really a production problem

Wall Street has spent a week treating Goldman Sachs ’ decision to pull AI coding tools from its traders as a story about cost or security or a frontier-model backlash. James Harrison , chief executive and co-founder of Genesis Global, a software provider to the financial services industry, told FinAi News that the analysis misses what actually happened: The tools worked fine. They just couldn’t ship. Goldman confined Claude Code and GitHub Copilot to its roughly 12,000 engineers last week, citing productivity and risk. Meanwhile, the traders who had used them to monitor flows, model positions and generate content lost access. [caption id="attachment_57375117" align="alignright" width="397"] (Courtesy/Bloomberg)[/caption] To Harrison, the restriction is not a one-off crackdown, but rather a symptom of a problem every bank is about to hit — the gap between what AI can build on a desk and what a bank can safely run in production. “Business demand for technology has always outrun engineering capacity, and AI has blown that gap wide open,” Harrison said. Traders can whip up a workflow or data tool in an afternoon, and it works — for them. But functioning on a desk and running in a bank are very different things, he explained. "Security, resilience, audit, integration and compliance sign-off don’t come free with a clever prompt," he said. Goldman Sachs did not respond to a FinAi News request for comments. The problem Restricting use of AI, Harrison argues, is where both of Goldman’s concerns originate. Many trader-built apps at Goldman end up on IT’s desk to be rebuilt, which is the productivity problem, while the distance between a working prototype and a production-grade system is the risk. He points to a mismatch in the technology itself. "General purpose AI is ‘probabilistic,’ and it's brilliant at well-understood, well-documented problems," he said. Capital markets software is deterministic: The trade either books correctly or it doesn’t. "Asking a probabilistic tool to produce a deterministic outcome on its own is where things come unstuck," he said. " That's not a knock on the tools. It's a question of using them for the right job." Taking the tools away doesn’t make the demand disappear, Harrison said. “It just sends the business back to the engineering queue,” he said. AI agents have given almost everyone that ability a build tools, Andrew Bud , CEO of digital identity company iProov , told FinAi News . The tech is racing ahead of risk mitigation. "Larger banks can deploy new security architectures that manage and mitigate the risks created by deployment of agents, but that is limited," Bud said. "It's slower than the rate at which people are deploying agents." Many FIs will look to slow AI usage until proper risk modules are ready to tackle the influx of tools being submitted, he said. Banks don’t need more AI code Harrison is equally skeptical of equating AI progress with raw code output. “Nobody ever bought a car for the number of bolts in it,” he said. The constraint has never been a shortage of ideas — tokenization, 24/7 trading, fixed income electronification, compliance, risk — only a shortage of places near the front of the engineering queue. And the cost of more code outlasts the prototype, Harrison explained. “Code is technical debt." Every line must be tested, supported and maintained for years. If AI writes 10 times more code, banks inherit 10 times more to look after. The goal, he said, is the opposite: more problems solved with less new code, built on components that are already proven. Cost also plays a role in AI restriction, Harrison said. Paying by the token for a frontier model to reinvent the same wheel on every desk “isn’t a strategy, it’s a habit,” he said. If most of what a desk needs already exists as production-grade components, banks pay AI only for what’s unique. The ROI debate misses the same point, in his view. “A prototype that never goes live costs money and earns nothing,” he said. Register here for early-bird pricing for the FinAi Banking Summit, taking place March 1-2 in Denver. View the full event agenda here .

Goldman’s AI clampdown is really a production problem

Trader-built apps work on the desk but can’t run in a bank — the hard part is shipping

ING said to be seller of $500 million data center loans

Related Posts

ING said to be seller of $500 million data center loans

Claude now plugs into your bank account — through Plaid

AnswerRocket replaces controller with AI agent, saving $100K annually

EMERGING FINTECH DIRECTORY

FinAi Podcast

SPONSORED

The Next Generation of Cyberthreats: Powered by AI

Build an Antifragile Strategy to Outperform the Market

How AI and Product Experts Turn Fuzzy Requirements Into Focused Dev-ready Roadmaps

Connect

Welcome Back!

Retrieve your password

Continue Reading with FinAi News Premium - Less than $2/Day

Goldman’s AI clampdown is really a production problem

Trader-built apps work on the desk but can’t run in a bank — the hard part is shipping

ING said to be seller of $500 million data center loans

Related Posts

FinAi Podcast

Connect

Welcome Back!

Retrieve your password

Unlock This Article

Continue Reading with FinAi News Premium - Less than $2/Day