Goldman Sachs has limited access to AI-powered coding tools from Anthropic and Microsoft to engineering personnel, reversing months of broader employee use as Wall Street banks reassess the security, cost and productivity implications of artificial intelligence.
By yourNEWS Media Newsroom
NEW YORK — Goldman Sachs has prohibited most bankers and traders from using certain advanced artificial intelligence coding tools following an internal review of productivity and risk management, reversing an earlier policy that allowed employees outside its technology division to experiment with the software.
The investment bank updated its internal guidance in October to restrict access to Anthropic’s Claude Code and Microsoft’s GitHub coding tools primarily to engineering staff, according to a Bloomberg News report published Monday, Nov. 2.
The restrictions have frustrated some employees who had incorporated AI tools into their daily work, including monitoring client trading activity, analyzing market movements, modeling financial positions and creating written or visual materials.
Goldman confirmed that the changes followed an experimental period during which the bank evaluated the tools’ usefulness and potential risks.
“We make certain AI tools available to our people to determine the best use cases across different parts of the firm,” Goldman spokesperson Nick Carcaterra told Bloomberg.
“After a period of experimentation where we assess a variety of factors including productivity and risk management, we put them in the hands of those who are best suited to leverage them for the long term.”
The revised policy does not constitute a companywide prohibition on artificial intelligence. Rather, it narrows access to particular coding tools while Goldman continues developing and deploying AI technology throughout its operations.
Employees known as “strats,” whose positions combine trading, quantitative research and software engineering, are exempt from the latest restrictions, according to people familiar with the decision.
The policy represents a notable change from Goldman’s approach earlier this year, when executives promoted the use of AI-assisted programming among employees without traditional software development backgrounds.
In July, Chief Information Officer Marco Argenti praised the practice known as “vibe coding,” in which employees use natural-language instructions to generate computer programs without writing all the underlying code themselves.
Argenti described bankers and traders developing preliminary software solutions with AI before passing them to the company’s engineering teams for refinement.
Goldman has approximately 12,000 engineers, reflecting the substantial technology infrastructure required to support its global investment banking, trading and asset management businesses.
The earlier approach allowed employees closest to business problems to experiment with software designed to address those needs.
A trader, for example, could use an AI coding assistant to create a prototype for analyzing market positions or monitoring trading activity before requesting assistance from software engineers.
The October restrictions place greater responsibility for using those tools with personnel whose primary duties include software development.
The change illustrates the difficulties financial institutions face when introducing increasingly capable AI systems into businesses that handle sensitive customer information and conduct complex financial transactions.
AI coding assistants can generate, modify and execute computer code, making them potentially useful for tasks beyond traditional programming.
However, their capabilities also raise questions about oversight, information security, reliability and whether the software they produce meets an institution’s technical standards.
Financial firms must consider those issues alongside the potential productivity gains from allowing employees to automate routine work.
The tools can accelerate development, but AI-generated code may require careful testing and review before being used in systems involving financial data or trading activity.
Goldman’s spokesperson did not identify a specific security breach or software failure as the reason for the restrictions.
The company instead characterized the decision as the result of a broader evaluation involving productivity and risk management.
The distinction is important because the policy does not establish that the restricted tools caused a security incident or produced harmful results.
It demonstrates that Goldman has determined narrower access is appropriate for the products following its internal assessment.
The development comes as major Wall Street institutions invest heavily in artificial intelligence while attempting to establish controls over how employees use the technology.
Banks are introducing AI systems to assist with research, document preparation, customer service, regulatory compliance, fraud detection and software development.
At the same time, executives are examining whether the cost of increasingly advanced models can be justified by measurable improvements in efficiency.
Different AI models may be suitable for different tasks, and more powerful systems can require substantially greater computing resources.
JPMorgan Chase has adopted an approach that encourages employees to select AI models according to the complexity of the work being performed.
The bank has emphasized that routine tasks, such as summarizing an analyst report, do not necessarily require the most expensive or computationally intensive available model.
That strategy reflects growing interest among financial institutions in managing AI expenses as adoption expands.
Goldman has also developed its own internal AI assistant, which incorporates models from OpenAI, Anthropic and Alphabet’s Google.
The bank’s continued investment in that platform shows that its restrictions on specialized coding products do not represent a withdrawal from AI development.
Instead, the company appears to be distinguishing between general-purpose AI assistance and tools capable of creating or modifying software.
That distinction can affect how employees perform research, develop financial models and automate business processes.
Goldman is also experimenting with agentic AI, a category of technology designed to perform multistep tasks with varying degrees of autonomy.
Such systems can potentially coordinate information retrieval, analysis and software operations with less direct human intervention.
Their development has created additional questions about accountability and the controls needed when AI performs tasks involving financial systems.
Goldman’s leadership has repeatedly described artificial intelligence as an important opportunity to transform its operations.
In May, President and Chief Operating Officer John Waldron said the bank’s processes contained substantial opportunities for automation.
He characterized the organization as a “human assembly line” whose operations could become more efficient through technology.
The company’s latest policy therefore represents an adjustment in how AI tools are distributed rather than a rejection of automation.
Goldman has already imposed other restrictions on employee access to certain Anthropic products.
In April, the bank removed access to Claude for employees working in Hong Kong following a review of contractual and geographic restrictions.
According to Reuters’ April 29 report, Goldman employees in Hong Kong lost access to Anthropic’s models through the bank’s internal AI platform.
Other AI systems, including Google’s Gemini and OpenAI’s ChatGPT, remained available.
The earlier restriction concerned Anthropic’s supported markets and the bank’s interpretation of its contractual obligations.
The October decision differs because it primarily distinguishes employees by professional responsibilities rather than geographic location.
Together, the changes illustrate the increasingly selective approach financial institutions are taking toward access to advanced AI technology.
Goldman’s relationship with Anthropic also extends beyond its use of the company’s software.
The investment bank has been selected as a lead underwriter for Anthropic’s planned initial public offering, reflecting the growing financial opportunities associated with artificial intelligence companies.
Goldman has also worked with semiconductor manufacturer Nvidia on major debt financing arrangements related to AI infrastructure.
Those activities place the bank on both sides of the AI investment market, as a financial adviser helping technology companies raise capital and as a business adopting the same technological developments.
The competing considerations include the potential financial returns from AI investment, the productivity benefits of internal adoption and the operational risks of deploying increasingly capable systems.
Anthropic’s Claude Code is designed to assist with software engineering tasks, including writing, examining and modifying code.
Microsoft’s GitHub platform provides development tools, including AI-assisted coding capabilities, that can help users generate software and troubleshoot programming problems.
The products have attracted interest beyond the technology industry because they make certain programming activities accessible to employees with limited coding experience.
That accessibility has encouraged experimentation in businesses seeking to automate processes without waiting for dedicated engineering resources.
However, organizations must determine when such experimentation is appropriate and when software development should remain within established technical teams.
Goldman’s revised policy places greater emphasis on that distinction.
The change could also affect how the bank develops new internal applications.
Employees who previously used coding assistants directly may now need to collaborate more closely with engineering teams when creating software-based solutions.
The bank has not publicly disclosed how many employees have lost access or whether it intends to revisit the restrictions.
It also has not announced an equivalent prohibition on the broader AI assistant tools available to its workforce.
The decision highlights a wider challenge facing employers as increasingly powerful AI systems move into everyday business operations.
Allowing employees to experiment can reveal useful applications that might otherwise go undiscovered, but broad access can also create difficulties involving quality control, security and the consistent application of internal standards.
Financial institutions face particular scrutiny because software errors or inappropriate handling of sensitive information can affect customers, transactions and regulated operations.
Goldman’s latest guidance reflects its attempt to balance those concerns while continuing to invest in artificial intelligence.
For employees who had begun relying on AI coding tools, the restrictions mark a reduction in the autonomy previously available to experiment with software.
For the bank, they represent a decision to concentrate access among engineering and quantitative personnel after assessing the results of wider deployment.
As Wall Street continues expanding its use of artificial intelligence, Goldman’s policy illustrates how financial institutions are moving from broad experimentation toward more selective decisions about which employees should have access to particular AI capabilities.
Original article
Comments