Goldman Sachs Bans Bankers From AI Coding Tools, Restricts Access to Engineers
Goldman Sachs just revoked AI coding access for bankers and traders, proving that controlling who uses AI is as critical as the AI itself.
Business · Source: PYMNTS
What happened
Goldman Sachs abruptly revoked access to advanced AI coding tools for its investment bankers and traders. Under revised internal guidance issued in October, tools like Anthropic's Claude Code and Microsoft's GitHub are now restricted exclusively to designated software engineers and quantitative strats. This removes autonomous coding capabilities from the front office entirely.
The restriction marks a massive policy reversal. Just months ago in July, Goldman Chief Information Officer Marco Argenti publicly praised grassroots vibe coding. Non-technical employees were independently assembling functional software solutions. Bankers relied on these tools to model market positions, track client trading flows, and draft content before passing their creations to professional developers for refinement.
The sudden clampdown sparked intense frustration among traders who lost their new automation tools. Goldman spokesperson Nick Carcaterra stated the shift followed a period of experimentation to assess productivity and risk management. The bank decided to put the tools only in the hands of those best suited to leverage them. Ironically, this operational clampdown happens while Goldman serves as the lead underwriter for Anthropic's planned initial public offering.
Key facts
- 12,000 — Number of professional developers at Goldman Sachs who retain access to AI tools
- October — Month the revised internal guidance restricted AI coding access
Why it matters
Wall Street is hitting the governance wall with generative AI. Coding agents operate under the permissions of their users. When bankers write and execute unvetted machine-generated code on sensitive trading desks, it creates massive security and regulatory nightmares. Safeguarding material non-public information becomes impossible if banks cannot audit what the autonomous software is doing. Maintaining strict data barriers is a legal requirement, not an option.
This sets a harsh precedent for enterprise AI adoption. Companies are realizing that giving everyone access to frontier models is a massive liability. Similar caution is sweeping the financial sector. JPMorgan Chase CEO Jamie Dimon recently warned of rogue agents, banning outside AI agents from the bank. Bank of America and Citigroup also placed strict limits on third-party agents accessing consumer deposits. The future of enterprise AI is not free-for-all access, but strict role-based deployment.
For builders
Build strict governance and audit trails
Enterprises will not buy AI agents that execute code without absolute oversight. Founders who build compliance, data barriers, and audit logs into their AI tools will win lucrative enterprise contracts. Banks pay for risk mitigation first and productivity second.
Target specific technical roles
General-purpose AI access is shrinking in highly regulated industries. Build specialized tools for designated engineers and quants rather than trying to sell a one-size-fits-all solution to the entire workforce. The buyers are IT leaders, not end users.
Create internal workflow bridges
Non-technical staff still want AI automation but lack the permissions to run it. There is a massive opportunity to build safe, sandboxed platforms. These platforms can let business users request AI automations that engineers can easily review and approve.
My take
I have always said AI power is not the problem, but who holds the keys is. Goldman Sachs just proved that handing autonomous coding agents to bankers is a compliance disaster waiting to happen. If you build enterprise AI, stop selling magic to the masses and start selling absolute control to the IT department.
Original reporting: PYMNTS. This is my rewrite and opinion.