← Back to feed
WritingArticle

Removing a coding tool does not remove the workflow

A reported restriction at Goldman Sachs raises a narrower product question: can domain experts shape useful software without also carrying authority to run it against sensitive systems?

Sourcethestateofai.comthestateofai.com ↗

A trader’s need to understand client flows does not disappear when a coding tool is withdrawn. That is the useful tension in State of AI’s November 2 report that Goldman Sachs restricted Claude Code and GitHub access largely to engineers. The account attributes the change to Bloomberg reporting; the supplied evidence does not include Goldman’s guidance or the underlying Bloomberg article. It is a reported restriction, not an independently established account of the bank’s policy.

The same article reports that traders and bankers had built tools and passed them to engineers for improvement. Those are two different activities: expressing a workflow in software, and accepting responsibility for operating that software. A role-based restriction can reduce access to a coding agent, but by itself it does not provide another route for the employee’s working knowledge to reach engineering.

Consider a hypothetical trader who wants to flag unusual changes in client activity. A prototype using synthetic records could demonstrate which changes deserve attention, which apparent anomalies are routine, and how the result should be displayed. An engineer could then turn those distinctions into tests and implement approved data access. The prototype would be a proposal for behavior, not permission to connect an improvised tool to client records. Whether Goldman offers such a route is not established by the supplied reporting.

That distinction suggests a product opportunity narrower than giving everyone a general-purpose coding agent: an environment where practitioners can construct examples and challenge outputs without receiving production execution authority. If it preserves representative cases and expected results, the handoff could give engineers something more precise than a feature request. The potential saving is in translating the workflow, not a demonstrated reduction in staffing or software costs.

A reasonable objection is that another prototyping environment simply creates another review queue. That would weaken the opportunity substantially, especially if synthetic examples miss the exceptions that matter in real records. The design earns its place only if engineers can reuse the resulting cases. Restricting execution and preserving expert participation need not be opposing choices—but connecting them requires a usable handoff, not merely a different access policy.

Grey Haven

Video summaries are narrated with an AI-generated voice.

Grey HavenApplied AI Venture Studio