Goldman Sachs restricts AI coding tools to engineers

The move to limit access

Goldman Sachs has updated its internal guidance to restrict the use of AI coding tools to its engineering departments. According to Bloomberg, the bank is barring most traders and bankers from using popular platforms such as Microsoft's GitHub and Anthropic's Claude Code.

This shift comes after the firm allowed bankers to experiment with these tools for several months. During that time, the bank monitored how the technology was being integrated into daily workflows and evaluated the potential consequences of its use.

The bank's spokesperson, Nick Carcaterra, explained that the firm is trying to identify the best use cases for different parts of the organization. He noted that after a period of experimentation, the bank decided to put the tools in the hands of those best suited to use them for the long term.

How bankers used the tools

Before the restrictions were put in place, many traders and bankers utilized AI coding tools to streamline their daily tasks. These employees used the software to follow how clients were trading and to simulate various shifts in market positions.

In addition to complex financial modeling, staff members used the tools for more general content creation. This included drafting various types of text and generating visuals to support their work. These tasks were part of a broader effort to see how AI could improve efficiency across the firm.

The bank's Chief Information Officer, Marco Argenti, previously spoke favorably about a practice called 'vibe coding.' This involved employees outside of technical roles creating their own solutions with AI and then handing the results to a pool of 12,000 engineers for polishing.

Risk and skill concerns

The decision to limit access was driven by concerns regarding risk management and overall productivity. While AI can speed up work, the bank is wary of the implications of non-technical staff generating code or data without proper oversight. The firm is looking for a balance between innovation and safety.

Some internal voices have warned about the long-term effects of these tools on the workforce. Goldman partner Chris Churchman suggested that the widespread use of AI across Wall Street could risk degrading the analytical abilities of the next generation of bankers. He noted that the bank has not yet determined how to manage that specific transition.

These concerns reflect a broader debate in the financial industry about the quality of human output. If automated tools handle the heavy lifting of analysis, there is a fear that junior employees may not develop the foundational skills necessary for high-level financial work.

Exceptions and internal tools

Not every employee at Goldman Sachs is being affected by the new restrictions. The directive does not cover a group of employees known as 'strats.' These individuals perform jobs that combine software engineering, quantitative research, and trading, making them more suitable for coding tools.

While the bank is limiting external tools, it is still investing in its own proprietary technology. Goldman has created its own AI assistant that draws on models from three major providers: OpenAI, Anthropic, and Alphabet. This allows the firm to maintain a level of control over the underlying technology.

The bank has a history of being cautious with AI deployment. Earlier this year, the firm restricted staff in Hong Kong from using Anthropic's Claude. These actions suggest a pattern of careful, localized adjustments as the bank navigates the integration of artificial intelligence.

The broader industry trend

Goldman Sachs is not alone in trying to figure out the proper boundaries for AI usage. Other major financial institutions are also developing their own internal guidelines to ensure that staff use the technology responsibly. The goal is to maximize benefits while minimizing errors.

For example, JPMorgan Chase recently stated that it wants its staff to match specific tasks to suitable models. The bank wants to ensure that its most expensive AI options are reserved for difficult work rather than simple tasks like condensing an analyst report.

As these firms continue to experiment, the industry is moving toward a more nuanced approach to AI. Instead of a total ban or total freedom, banks are creating specific rules based on the complexity of the task and the technical expertise of the user.

What comes next

The restriction at Goldman Sachs marks a significant turning point in how the bank views the role of AI in non-technical departments. The firm will likely continue to monitor the productivity of its engineers who have full access to these tools to see if they can provide a blueprint for the rest of the company.

Traders who were frustrated by the loss of access may have to find alternative ways to simulate markets and draft content. They will need to rely on the bank's internal AI assistant or traditional methods until the firm decides to expand access again.

The ultimate goal for the bank appears to be a controlled environment where AI enhances human capability without creating systemic risks. The industry will be watching to see if these restrictions lead to better risk management or if they simply slow down the pace of innovation on Wall Street.

The role of 'strats'

The exclusion of 'strats' is a key detail in the bank's new policy. Because these employees work at the intersection of software engineering and trading, they are viewed as the appropriate bridge between the two worlds. They possess the technical literacy that the general banking population may lack.

By allowing strats to keep using coding tools, Goldman is ensuring that the people most capable of handling the code still have the resources they need. This creates a tiered system where the most complex, code-heavy tasks remain with specialists while general staff use more restricted interfaces.

This distinction highlights the bank's preference for 'human-in-the-loop' systems. The bank wants to ensure that whenever code is generated or executed, there is a high level of technical competence involved in the process to prevent errors in market positions or client data.

The cost of AI models

The mention of JPMorgan's strategy regarding model costs points to an emerging economic reality in AI. Not all AI models are created equal, and using a high-powered, expensive model for a simple summary is seen as inefficient. Banks are looking to optimize their spending on API calls and computing power.

Goldman's own AI assistant, which uses models from OpenAI, Anthropic, and Alphabet, likely allows them to switch between different providers depending on the task. This multi-model approach gives the bank flexibility to choose the best tool for a specific job while keeping costs under control.

As these tools become more standard, the focus for large institutions will shift from 'can we use it' to 'how do we use it most efficiently.' The goal is to find the sweet spot where the cost of the AI tool is justified by the value it adds to the final financial product.

Impact on junior staff

The warning from Chris Churchman about the 'next generation of bankers' is a significant concern for the industry. If junior analysts rely on AI to perform the basic tasks they used to do manually, there is a fear they won't learn the underlying logic of finance. This could create a skills gap as they move into more senior roles.

Goldman's decision to limit coding tools to engineers may be an attempt to protect this learning process. By restricting the ability to 'vibe code' solutions, the bank may be forcing non-technical staff to engage more deeply with the underlying data and logic of their work.

The bank is currently in a period of assessment to determine how to manage this transition. They are looking for ways to integrate AI into the workflow without sacrificing the analytical rigor that defines the firm's reputation and the success of its clients.

Regional variations in AI

The fact that Goldman Sachs restricted Anthropic's Claude specifically for its Hong Kong staff shows that AI policy can be geographic. Different regions may have different regulatory requirements or different levels of risk tolerance regarding specific AI models.

This suggests that the bank's approach is not a one-size-fits-all solution. They are likely adjusting their AI permissions based on local laws, the specific needs of regional offices, and the types of data being handled in different markets.

For the reader, this means that the rules for AI in finance are still being written. As the technology evolves, we can expect to see more specific, localized, and task-based permissions rather than broad permissions for all employees.

The future of 'vibe coding'

The concept of 'vibe coding'—where non-technical users describe a desired outcome and the AI generates the code—is a major trend. While it allows for rapid prototyping, it also introduces a layer of abstraction that can be dangerous in a regulated environment like a bank.

Goldman's move to limit these tools suggests they are not yet comfortable with the 'vibe' being the primary driver of production code. They prefer the 'polishing' step by a human engineer to ensure that the final product is accurate, secure, and compliant with financial regulations.

Moving forward, the bank will likely refine its internal AI assistant to provide a safer version of this experience. They want to give employees the ability to create solutions without the risk of deploying unverified or poorly structured code into the firm's systems.

Productivity vs. Risk

The core of Goldman's decision rests on the balance between productivity and risk management. AI tools can significantly speed up the drafting of visuals and the simulation of market shifts, which are key goals for any investment bank.

However, the risk of an AI-generated error in a trading simulation or a client report could be catastrophic. By limiting access to engineers, the bank ensures that those with the highest level of technical training are the ones responsible for the output of these tools.

This reflects a broader corporate trend where the 'move fast and break things' mentality of the tech world is being tempered by the 'do no harm' requirement of the financial world. Goldman is choosing a slower, more controlled path to AI adoption to protect its core business.

The role of the AI assistant

Goldman's internal AI assistant is designed to be the primary interface for the bank's employees. By drawing on models from OpenAI, Anthropic, and Alphabet, the bank can leverage the strengths of different AI architectures while maintaining a single, controlled entry point for its staff.

This internal tool likely serves as a 'walled garden' where the bank can monitor every prompt and every output. It allows the firm to provide the benefits of AI—such as summarizing reports or drafting text—without giving employees direct access to the raw coding tools that caused the recent restrictions.

As the bank continues to develop this assistant, it will become the primary way that bankers and traders interact with AI. The success of this tool will determine whether the bank can eventually expand AI access beyond the engineering department.

Conclusion of the review

The review of productivity and risk concerns led to the current state of restrictions at Goldman Sachs. The bank is currently in a phase of 'experimentation' to see how these changes affect the bottom line and the safety of their operations.

While some traders are frustrated by the loss of tools, the bank's leadership seems committed to a cautious rollout. They want to ensure that the technology serves the firm's long-term goals rather than providing a short-term boost that could lead to long-term risks.

For the curious reader, this story illustrates the growing pains of the AI era. It shows that even the most advanced firms are still figuring out how to balance the incredible power of new technology with the need for human oversight and traditional expertise.

Summary of the shift

In summary, Goldman Sachs is moving toward a model where AI coding tools are a specialized instrument rather than a general-purpose utility. This move protects the bank's analytical integrity and ensures that high-risk tasks are handled by qualified personnel.

The bank's strategy involves a mix of internal tool development, selective access for specialized roles like 'strats,' and a cautious approach to external models. This multi-layered strategy is designed to navigate the complexities of AI in a highly regulated industry.

As other banks like JPMorgan also refine their AI policies, we can expect a standardized set of 'best practices' to emerge. These will likely prioritize accuracy, cost-effectiveness, and the preservation of human skills in the face of rapid technological change.

The impact on traders

The traders who are most frustrated by these changes are those who saw immediate benefits from the tools. They were using them to stay ahead of client trends and to visualize complex market shifts, tasks that are central to their competitive advantage.

By removing these tools, the bank is essentially asking traders to return to more traditional methods or to use the bank's approved internal assistant. This may slow down their workflow in the short term but is intended to prevent errors that could arise from using unvetted code.

The bank's goal is to ensure that the 'speed' provided by AI doesn't come at the cost of 'accuracy.' For traders, this means a shift in how they interact with technology, moving from direct coding to guided AI assistance.

The 'strats' exception

The inclusion of 'strats' as an exception is a notable detail in the bank's policy. These employees are uniquely positioned because their roles require both a deep understanding of the markets and the ability to write and manage software.

Because strats are already tasked with quantitative research and software engineering, they are the natural fit for AI coding tools. They have the training to understand the risks of the code the AI generates, making them safer users of the technology than a general banker might be.

This distinction suggests that Goldman Sachs views AI as a tool for 'power users.' The bank wants to empower those who have the skills to use it correctly while providing a safer, more restricted experience for the rest of the organization.

The future of banking AI

The story of Goldman Sachs is a microcosm of the larger trend in the financial sector. Banks are moving away from the 'wild west' of AI experimentation toward a more structured, governed approach to technology integration.

We can expect to see more banks following Goldman's lead by creating internal AI assistants and restricting access to raw coding tools. The focus will be on 'safe' AI that can perform routine tasks while leaving the high-stakes decisions to human experts.

For the reader, this means that the future of AI in the workplace will likely be defined by these boundaries. The goal isn't to replace humans with AI, but to create a system where humans use AI as a tool that is both powerful and predictable.

Final thoughts on risk

At its core, the Goldman Sachs story is about the management of risk. In finance, a single error in a piece of code can lead to millions of dollars in losses or regulatory fines, making the bank extremely cautious about who can generate that code.

By limiting tools to engineers, the bank ensures that there is a human 'gatekeeper' for every line of code that influences the markets. This is a deliberate choice to prioritize stability over the speed of individual employees.

As AI continues to evolve, the tension between the desire for efficiency and the need for safety will remain a central theme. Goldman's current policy is a practical response to that tension, reflecting the reality of working in a high-stakes industry.

The role of external models

The fact that Goldman is using models from OpenAI, Anthropic, and Alphabet shows that they are not trying to build everything from scratch. They are leveraging the best available technology while wrapping it in their own security and governance layers.

This approach allows the bank to benefit from the rapid improvements made by the leading AI labs. However, by controlling the interface through their own assistant, they can ensure that the data used by these models remains private and that the outputs meet the bank's standards.

This 'hybrid' model of using external intelligence within an internal framework is likely to become the standard for many large corporations. It provides the best of both worlds: the power of cutting-edge AI and the security of a private, controlled environment.

The impact of 'vibe coding'

The mention of 'vibe coding' by Marco Argenti highlights a shift in how people think about programming. It suggests that the barrier to creating software is falling,