Goldman Sachs cuts off Hong Kong staff from Anthropic's Claude

Goldman Sachs has cut off its Hong Kong employees from Anthropic's Claude, an AI tool that helps write software code faster. According to Bloomberg, the restriction is already in effect, but the bank has not publicly explained why.

What happened in Hong Kong

This is not Goldman Sachs abandoning generative AI altogether, but a targeted restriction in one of its key Asian offices. Bloomberg, citing a source familiar with the situation, reports that employees in Hong Kong can no longer use Claude. It is still unclear whether the decision affects other divisions of the bank, whether the measure is temporary, and whether it is tied to an internal review, a security policy or local data-handling requirements.

The regional nature of the restriction matters as much as the shutdown itself. At large international banks, access to internal and external digital services is often configured not uniformly but by country, legal entity and team function. So even a narrow change in a single office can mean either a local adjustment of the rules or a broader rethink of how AI tools are admitted into engineering workflows across a global organization.

What may have played a role

The article does not name a reason, so any explanation here is only a cautious inference from context. For large banks, tools like Claude are not just chatbots but a potential access point to internal code, documentation, workflows and client data. If employees use AI to speed up development, a bank typically assesses separately which pieces of code can be sent to an external service, where requests are stored, how auditing is set up and who is responsible for compliance with local rules.

• sending internal code fragments and configurations to an external service
• regional restrictions on data processing and storage
• a review of AI pilot programs after an internal audit
• different sets of permitted tools for different offices and teams

The report describes Claude as an agent that speeds up software development. For a bank this is a useful but sensitive use case: the assistant can suggest code fragments, explain the logic of systems and help with refactoring, but any gain in speed has to be weighed against compliance risks and the risk of intellectual property leaks. This is especially true for international organizations, where the same platform has to pass review in several jurisdictions at once.

A signal for the market

The Hong Kong story matters beyond a single bank. Financial institutions are actively testing generative AI, but banks run into security, regulatory and governance requirements more often than most. Any tool that helps write code automatically falls under stricter controls than an ordinary corporate chat. That is because it involves not just text but access to the logic of internal systems, libraries, engineering practices and potentially data tied to the business's critical infrastructure.

It is also a telling episode for Anthropic. Even if the restriction covers only one region, it is a reminder that winning in the enterprise segment depends on more than model quality. Deployment terms, data routing, log transparency, admin-control flexibility and willingness to adapt to local rules matter just as much. In enterprise AI, buyers pay not for an impressive demo but for predictability, control and the ability to pass an internal risk review.

What this means

The Goldman Sachs news shows that the enterprise AI market is entering a phase of tougher filtering. For large companies it is no longer enough for a tool to be simply strong: it must fit regional policies, compliance and the access-management model. For vendors this means one thing: enterprise sales are increasingly decided not in the interface but in the security department.