The restrictions follow an internal review of productivity and risk management, Goldman spokesperson Nick Carcaterra told Bloomberg. They mark a retreat from the “vibe coding“ approach Chief Information Officer Marco Argenti promoted in July, when he described employees building prototypes with AI that the bank’s roughly 12,000 engineers would then refine.
That model promised to close the gap between spotting a business problem and shipping a fix. The ban frustrated some traders who had used the tools to watch client flows, model positions and produce written and graphic material, a person familiar with the matter told Bloomberg.
The difficulty is that financial software cannot be judged only by whether it runs. An application can produce plausible numbers while applying the wrong pricing assumption. A coding agent with access to internal systems can also pull confidential data, link restricted databases or push changes outside the bank’s approval process.
MUFG, BNY and Deutsche Bank Keep the Tools but Add Controls
Japan’s Mitsubishi UFJ Financial Group (MUFG) built an in-house alternative to Claude Code after finding the tool could work around its information-access limits and compliance safeguards, WatersTechnology reported in July. Stu Taylor, MUFG’s head of global markets AI outside Japan, told the publication the bank keeps a human in the loop on every AI tool, and autonomous coding agents make that harder to enforce.
BNY, unlike Goldman, lets employees outside engineering build on its Eliza platform, inside an environment designed around the bank’s data, legal and compliance requirements. BNY lists more than 400 AI-enabled solutions in production, and artificial intelligence now writes roughly 40% of its code, PYMNTS reported in April.
Deutsche Bank, like Goldman, keeps the tools with professional developers. More than 15,000 of them use GitHub Copilot and Google Gemini Code Assist, saving an estimated 1.5 to 2.5 hours a week each, the bank said. Developers stay responsible for the code they commit and must complete training on safe AI use first.
No CFO Surveyed Would Give AI Agents Full System Access
The caution matches what finance leaders tell PYMNTS Intelligence. Among executives at firms already using or piloting AI agents, 43% would give agents moderate system access with human approval, and none would grant full access, according to the PYMNTS Intelligence report “AI Agents in the Enterprise,” a July 2025 survey of 60 CFOs at U.S. companies with at least $1 billion in revenue.
Spending is still rising. Among financial services and insurance firms, 85% plan to increase AI budgets over the next 12 months, and 30% name data quality as their biggest barrier, according to the PYMNTS Intelligence report “Financial Services Pulls Ahead in the Enterprise AI Race,” a March survey of 60 senior technology executives at $1 billion-plus enterprises.
The split matters commercially. AI coding vendors are selling the ability to build applications from plain-language instructions, which would expand their market beyond engineers. Banks may instead concentrate spending on specialized users or proprietary platforms that give them control over data and execution.
Goldman has not disclosed a security incident or quantified the productivity findings behind its decision. The move does not show that employee-built applications failed. It does challenge the assumption that making software easier to create makes it easier for a bank to deploy.
The dividing line on Wall Street is not whether AI should write code. It is who can authorize what that code does, which systems it can reach and who is accountable when it goes wrong.
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.
