JPMorgan Chase is tightening the usage of its internal AI programming tools. According to internal sources, the bank has set a monthly expenditure limit of $2,000 for some employees who use Anthropic Claude Code, and is also implementing a new isolated development environment to reduce the security risks associated with AI access to internal systems.

Some engineers are encountering budget reporting errors.

It was reported that last month, within a Teams group for certain engineers, someone asked about a newly appearing error code “ExceededBudget” and noted that “Budget = 2000.0”. Several employees replied that the usage quota for some engineers in this group is reset monthly, and if necessary, they can apply to increase the limit.

This is the first time that JPMorgan Chase has been exposed to having set clear expenditure limits or token limits for internal AI tools. Previously, although the bank had been continuously tracking engineers' use of AI, it was not clear to the outside world whether there was any specific budget control in place.

Wall Street has begun to pay attention to the costs of token.

As large models are being rapidly implemented in investment banking and trading operations, token expenses are becoming a new cost item on Wall Street. Zachery Anderson, the head of payment business data and analysis at JPMorgan Chase, stated in June this year that some employees' expenses on AI and token even exceeded their salaries.

However, JPMorgan Chase's management was previously cautious about this expense. The bank's Chief Financial Officer Jeremy Barnum stated during the second-quarter earnings call in July that at that time, the token costs were still "on a relatively small scale," but it was expected to increase significantly in the second half of the year.

JPMorgan Chase declined to comment on the specific quota arrangements. A spokesperson for the bank stated that the bank is implementing AI in a more stringent manner and linking costs with indicators such as business value, adoption rate, efficiency improvement, quality, speed, capacity, and risk control.

Devspace restricts AI from accessing internal systems

In tandem with the advancement towards the budget cap, there is also a new architecture named Devspace. Reports indicate that JPMorgan Chase previously allowed some developers to test Claude directly on their personal desktops, but the Devspace plan is to migrate the AI tool to a more isolated environment, restricting its access to employee credentials and internal systems.

According to the internal technology blog, Devspace is a containerized sandbox environment deployed on Amazon Web Services (AWS). It acts as an independent coding space where software can run without directly affecting other systems within the enterprise.

This design is particularly suitable for AI agents. Such tools are not only capable of generating code but may also execute tasks autonomously. In April of this year, Pat Opet, JPMorgan Chase's Global Chief Information Security Officer, stated that placing powerful AI tools directly on employees' desks where emails, sensitive documents, and personal information are stored poses obvious risks.

At that time, he proposed that in an ideal state, the AI agent should have an independent identity, but by default, it should not possess system permissions; when it needs to access resources, the enterprise should be able to clearly understand the context of the access and grant authorization accordingly. Following this logic, even if the AI agent breaks through the test environment, it would not be able to directly invoke internal credentials or abuse the enterprise system.

The coverage range is still limited.

Currently, the coverage of Claude within JPMorgan Chase is not yet extensive. Reports mention that the bank's global technology department has approximately 65,000 employees, and in August of this year, internal dashboards indicated that around 8,000 people had Claude permissions. At the beginning of September, there were about 1,900 members in the discussion group related to Devspace and Teams.

This means that JPMorgan Chase is still in the stage of gradually expanding the usage scale of the AI programming assistant. On one hand, the bank hopes to leverage AI to improve engineering efficiency; on the other hand, as concerns about out-of-control AI agents increase from the outside world, financial institutions are also simultaneously strengthening their isolation, authorization, and auditing capabilities.

For large banks, customer identity information, credit card data, and internal business systems are all considered highly sensitive assets. AI To deploy this tool on a larger scale, cost control and secure isolation have become equally important.