Why Goldman Sachs Took Claude Code Away from Its Bankers and Traders
Goldman Sachs pulls Claude Code, reflecting enterprise AI governance shifting from 'can it be used' to 'who should use it.'
Goldman Sachs recently pulled the AI coding tool Claude Code from its bankers and traders, sparking heated discussion on Hacker News. The article points out that this is not a rejection of AI capabilities but a rational recalibration of enterprise-grade AI governance: agentic tools can execute commands and modify files, and in financial scenarios permission boundaries and audit requirements are extremely high, while business staff lack the ability to review AI-generated code, creating an asymmetric risk-reward profile. Three voices in the discussion collectively indicate that the dividing line for enterprise AI has shifted from 'can it be used' to 'who should use it and how to audit it.' For vendors like Anthropic, role-based permissions, operation auditing, and sandboxed execution will move from bonus features to product standards, and governance capability becomes the ticket to open the doors to highly regulated industries.
Why Goldman Sachs Took Claude Code Away from Its Bankers and Traders
Financial giant Goldman Sachs recently made a decision worth pondering: it will no longer let its bankers and traders use the AI coding tool Claude Code directly. The news sparked a nearly hundred-point discussion on Hacker News, with 37 comments reflecting a critical look at the difficulties of deploying AI in the financial industry. A top Wall Street investment bank actively taking a powerful AI tool away from frontline business staff reflects not a technological step backward, but a deep-seated challenge in enterprise-grade AI governance.
The Nature of the Pullback: Not a Ban, but a Redrawing of Boundaries
Based on the original information, Goldman Sachs' action was "pulling Claude Code from its bankers and traders." Claude Code, launched by Anthropic, is positioned as an AI coding agent capable of executing code generation, file operations, command execution, and other tasks in the terminal. Its original design scenario is closer to the daily development workflow of engineers than the office desktops of financial business staff.
When investment banks pull such tools back from the business side, it usually points to several practical constraints: permission boundaries that are hard to control, insufficient operational auditability, and business staff lacking the ability to review AI-generated code. When AI agents can directly execute commands and modify files, the financial sector's requirements for "who did what and whether it can be traced" are far stricter than in general office scenarios.
Claude Code's Product Positioning
Claude Code is an agentic coding tool launched by Anthropic in 2025. It runs in a terminal environment and can autonomously complete tasks such as reading codebases, generating and modifying files, executing shell commands, running tests, and committing Git changes. Unlike traditional code completion tools (such as early GitHub Copilot), it has the ability to "execute" rather than merely "suggest," which brings high efficiency but also introduces risks such as accidental file modifications, erroneous command execution, and leakage of sensitive data. This characteristic determines that when deploying it within enterprises, it must be accompanied by permission isolation and auditing mechanisms, and cannot be opened to all employees like an ordinary chatbot.
The AI Paradox in Finance: Wanting Efficiency, but Fearing Loss of Control
Goldman Sachs is not an institution that resists AI. On the contrary, it has long been one of Wall Street's pioneers in exploring generative AI, and internally has promoted the deployment of various AI-assisted tools in research and development scenarios. Pulling back Claude Code is less a rejection of AI capabilities than a recalibration between "efficiency gains" and "risk exposure."
The core work of bankers and traders is capital allocation, client relationships, and risk pricing—not writing executable code. Putting a developer-oriented agentic coding tool in their hands offers limited potential benefits, but once an AI agent misoperates, accesses beyond its authority, or generates unauditable changes, the cost may far exceed the efficiency gains. This asymmetric risk-reward structure makes the pullback a rational choice.
Agentic AI and Auditability
Agentic AI refers to AI systems that can autonomously plan steps, invoke tools, and execute multi-step operations, as distinct from conversational AI that only generates text. Its core feature is "action capability"—it can read and write files, access APIs, and run commands. For enterprises, this brings a fundamental challenge: the output of traditional AI is suggestions that humans can review; the output of agentic AI is operations that have already occurred, and post-hoc tracing depends on logs and sandboxes. The financial industry is subject to regulatory constraints such as those from the SEC and FINRA, requiring complete audit trails for operations that affect trading, client data, and accounting—exactly what general-purpose agent tools struggle to satisfy directly.
Three Voices in the Hacker News Discussion
Around the 37 comments on this event, three typical viewpoints can be summarized. The first holds that this exposes the fundamental difficulty of deploying agentic AI tools in enterprises: the stronger the capability, the higher the governance cost. The second points out that the problem is not the tool itself, but that enterprises have not designed differentiated AI usage policies for different roles. The third voice is more sober—Goldman Sachs' approach precisely shows that its AI governance system is functioning and can identify and correct inappropriate tool proliferation.
The common thread in these discussions is that the dividing line for enterprise AI adoption has shifted from "can it be used" to "who should use it, how, and how to audit it." Agentic tools with execution capabilities like Claude Code have pushed this question to the forefront.
Implications for AI Tool Vendors: Role-Based Permissions Will Become Standard
Goldman Sachs' pullback is a clear signal to Anthropic and similar AI coding tool vendors. Enterprise customers are no longer satisfied with a single general-purpose version of a tool; they need fine-grained role permissions, operation audit logs, sandboxed execution environments, and restricted modes for non-engineering roles.
It is foreseeable that the product roadmaps of agentic AI tools will accelerate toward being "enterprise governance-friendly": distinguishing read-only suggestion modes from executable modes, presetting permission templates for different functions, and incorporating every agent operation into traceable records. Whoever can turn governance capabilities into product capabilities earlier will find it easier to open the doors to highly regulated industries such as finance, healthcare, and legal.
The Significance of RBAC and Sandboxed Execution in Enterprise AI
Role-based access control (RBAC) is a mature paradigm for enterprise software governance and is now being transplanted into the AI tool domain: presetting capability boundaries for different functions, for example, analysts read-only, engineers writable, managers approvable. Sandboxed execution refers to running AI-generated code or commands in an isolated environment to avoid direct impact on production systems. Audit logs record the inputs, outputs, and side effects of every agent invocation and are the foundation of compliance review. Together, the three form the "governance triad" for agentic AI to enter highly regulated industries, and also a new dimension of differentiated competition for tool vendors.
Conclusion
Goldman Sachs taking Claude Code away from its bankers and traders is not a story about AI failure, but a story about AI maturity. When enterprises begin to seriously distinguish "who should have which AI capability," agentic AI truly enters the stage of scalable deployment. For tool vendors, governance is not a drag on adoption but the ticket to the next stage of competition.
Related articles
Hermes Agent Beginner's Guide: Hands-On with a Self-Evolving AI Agent
A beginner's guide to Hermes Agent, the open-source self-evolving AI agent that gained 41.7K GitHub stars in two months—covering Harness Engineering, the learning loop, three-layer memory, WSL installation, and hands-on configuration.
AI Agent in Action: Building a Smart Office Assistant with WorkBuddy
A practical case study shows how Tencent WorkBuddy batch-converts and archives files, compares chat LLMs with AI Agents, and analyzes WorkBuddy's credit system, use cases, and positioning for office automation selection.
From Single-GPU Bottlenecks to Hundreds of Billions of Parameters: 6 Key Questions in Large Model Distributed Parallelism
Based on a Bilibili tutorial, this article organizes core knowledge of large model distributed parallelism: starting from decoder-only architecture's token-by-token generation and the essence of X times W computation, it analyzes six key questions including data parallelism, tensor parallelism, and pipeline parallelism under single-GPU bottlenecks. Suitable for learners in private deployment and model training scenarios to build systematic understanding.