Goldman Sachs has restricted its use of advanced AI development tools, including Anthropic's Claude Code and Microsoft-owned GitHub, to engineers. Bloomberg reported on November 2, having reviewed internal guidance updated in October. Until a few months earlier, bankers could also use these tools in their work, but the change removed them from the list of eligible users. Notably, Goldman Sachs has not banned AI coding itself; it has narrowed who may use the tools. Engineers can still use them, so the move looks like a review of the scope of use rather than a full retreat.
In April the restriction was by region; in October it was by job role
According to Bloomberg, Goldman Sachs weighed concerns about productivity and risk management and then barred most employees from using advanced AI coding tools. Bankers and traders had previously used Claude Code and GitHub for tasks such as monitoring trading flows and creating content. Reports say the October change, which limited use to engineers, has drawn complaints internally.
This is not the first time Goldman Sachs has restricted access to AI tools in 2026. On April 29, Bloomberg reported that employees in Hong Kong had lost access to Claude. That restriction was based on location rather than job role, and it also applied to staff visiting Hong Kong from overseas. At the time, a person familiar with the matter said Claude's main users were software engineers.
In other words, access was restricted on different grounds: region in April, job role in October. The two measures were not necessarily based on the same policy, however. On the April case, Finextra reported that it was linked to a review of the scope of use under the contract with Anthropic and to the fact that Hong Kong is not among the regions where the company offers its services. The April restriction may therefore have stemmed from circumstances separate from any risk assessment of the AI tools themselves.
The October measure, meanwhile, keeps engineers' access in place, which suggests that, at least for now, the firm does not intend to remove AI coding tools from the organization. However, the reports do not make clear how many employees were affected or whether alternative tools have been provided for non-engineers.
MUFG built its own version to address Claude Code's control challenges
Japanese financial institutions face similar challenges. According to a July 8 report by Risk.net, Mitsubishi UFJ Financial Group (MUFG) viewed Claude Code's capabilities as so powerful that it could bypass restrictions on information access and the bank's security and compliance constraints, and so developed its own version.
Stu Taylor, who oversees AI for global markets outside Japan at MUFG, explained that all of the bank's AI tools incorporate a "human-in-the-loop" approach, in which a person reviews and approves actions. He also indicated that general-purpose agentic tools such as Claude Code, Cursor and GitHub Copilot Workspace make it difficult to maintain the controls the bank requires.
With conventional conversational AI, it was relatively easy to operate on the basis that a person reviews the generated text or code and uses it if there are no problems. Agentic coding tools, however, can carry out a series of tasks autonomously, including searching for information, manipulating files and executing code. If processing proceeds without a person checking each individual operation, risks related to access permissions and information management may increase. The more autonomous a tool becomes, the more situations arise that traditional approval procedures alone cannot manage.
That said, there is no material showing that Goldman Sachs restricted use for the same reasons as MUFG. What the two firms' responses have in common is that, rather than opening general-purpose AI development tools to all employees as they are, they are trying to keep use within a manageable scope. MUFG developed its own version, while Goldman Sachs limited users by job role. This shows there is more than one way for financial institutions to manage the adoption of AI coding tools.
Productivity or risk management: which carried more weight?
Goldman Sachs's decision cannot be explained by risk management alone. Nick Carcaterra, a spokesperson for the firm, said the measure took into account "a variety of factors, including productivity and risk management." The idea is to end the trial phase and narrow access to the employees best suited to long-term use. It is therefore possible that use by non-engineers did not deliver the productivity gains that had been expected.
The decision also partly diverges from the direction of the AI adoption Goldman Sachs had pursued until then. In a July interview, Chief Information Officer Marco Argenti praised a development approach in which traders and asset management staff use AI to build prototypes, which engineers then turn into systems suitable for production. The idea is to improve operations by having frontline staff shape their own ideas and hand them off to specialist engineers.
If non-engineers can no longer use the same tools, this development process could be affected. Frontline staff may have fewer chances to build working prototypes, and there may be more cases of a return to the traditional approach of conveying requirements to engineers. On the other hand, limiting the tools to specialists could make it easier to standardize code quality and information management. From the information currently public, it is impossible to tell whether productivity or risk management most influenced this decision.
Still, there is a basis for seeing the measure as a limit on the scope of use rather than a retreat from AI coding. The spokesperson described a review of who may use the tools rather than a blanket ban, and engineers are still permitted to use them. This fact alone does not establish that risk management was the main driver, though, and it is also consistent with the interpretation that this was a selection driven by productivity.
The key question going forward is whether Goldman Sachs will narrow the scope further or allow non-engineers to use the tools again in a controlled environment. If even engineers' use were scaled back, the likelihood of a broader policy shift on AI coding tools would grow. Conversely, if a controlled in-house environment like MUFG's or alternative tools were introduced, the core issue would appear to have been not job role itself but how tool permissions and operations are managed.
For financial institutions, adopting AI coding tools is becoming more than a productivity measure; it is an issue that shapes how information management and internal controls are handled. The responses of Goldman Sachs and MUFG show that, when bringing AI autonomy into operations, there are two approaches: restricting who can use the tools, and reworking the tools themselves into a form that can be managed. Future updates to internal guidance and the rollout of alternative tools will offer clues to how far AI agents will be used in the financial industry.