Goldman Sachs cuts off most traders' access to Claude Code
Building a script to analyze trades is easier with AI, but who is responsible for its calculations? Goldman Sachs restricted most traders' access to Claude Code and GitHub tools following a performance and risk assessment. According to Bloomberg, rules updated in October retain access for engineers and specialists whose work bridges software development and trading. For teams, it is a reason to establish in advance who reviews and maintains these programs.
AI-processed from Habr AI; edited by Hamidun News
An employee spends an evening using AI to build a script that previously would have required a request to the development team. By the next morning, colleagues are already using its numbers. Convenient—until someone asks who checked the calculations and who will fix the program when the data changes. This is where the Goldman Sachs story becomes relevant to any founder who has allowed their team to build internal tools with AI.
According to a Bloomberg report, the bank restricted most traders' and bankers' access to Claude Code and GitHub tools. The decision followed a performance and risk assessment. It is a telling case: being able to write a program quickly does not answer the question of how to integrate it into a company's operations.
What changed at Goldman Sachs
For several months, bank employees could experiment with these tools. Then internal rules updated in October limited access to engineering roles. According to the same report, an exception remained for strats—specialists who combine trading, quantitative research and software development.
The tools already had clear uses. Traders used them to monitor client trading flows, model market changes and positions, and produce text and graphics. The restrictions therefore affected everyday tasks where employees had already found AI useful.
A Goldman spokesperson described the bank's approach as experimentation followed by evaluation: access is allocated based on effectiveness and risk management. However, the published report contains no measurements of time saved, code review costs or potential losses. It provides no basis for concluding that vibe coding proved unprofitable.
The script runs—but can you trust the numbers?
Consider a hypothetical example that illustrates the problem. A trader asks AI to build a spreadsheet: load the day's trades, group them by client and show how positions would change if the market fell. The program runs, generates a chart and produces convincing numbers.
But a successful run checks only part of the task. Are currencies matched correctly? How are canceled trades handled? Are the units consistent? What happens if the source supplies incomplete data? An error in a script like this can look like perfectly polished analysis. This illustrates an engineering risk; it does not describe a confirmed incident at Goldman.
A domain expert can spot a strange result. An engineer can review error handling, dependencies and access permissions. A reliable tool requires both sets of skills, as well as someone who will still be responsible for it a month later. The presence of AI does not assign those responsibilities by itself.
That raises a useful question for the team: how long does the entire journey from request to verified result take? The speed of writing the first version is easy to see. The time needed to check formulas, make corrections and provide support often becomes apparent later, when colleagues are already using the tool.
Takeaways for developers and founders
There is no need to copy the bank's restrictions wholesale. A small team and a large bank have different processes. A practical step is to define when a personal experiment becomes an operational service: colleagues start using it, it gains access to data or its output begins to influence decisions.
At that point, assign an owner, check calculations against known examples and agree on who approves changes. For the hypothetical trading script, a useful test would be a set of previously analyzed trades with expected results. An attractive chart is no substitute for that check.
Product managers should measure the time to a usable result, including review and corrections. Founders should allocate engineering time in advance to maintain internal tools. Otherwise, the developers' task backlog may turn into a queue of other people's prototypes that urgently need support.
The Goldman story offers a concrete reason to revisit this process. Public data is insufficient to assess the economic impact of the bank's decision. But the question of responsibility already applies to every company today: who will confirm that the program calculates correctly, and who will stay with it after its first successful run?
- Nov 12, 2026Jensen Huang steps down as Nvidia CEO: company loses $1.4 trillion in two weeks
- Nov 12, 2026Claude in Chrome requires approval for every click on category 3 sites
- Nov 12, 2026AWS Quick Gets Six Agents for Product Managers: From Customer Complaints to a Spec
- Nov 12, 2026Anthropic to Cut Off Claude for Russian Companies From November 12, Even Those With Foreign Legal Entities
Key Questions Answered
The script runs—but can you trust the numbers?
Consider a hypothetical example that illustrates the problem. A trader asks AI to build a spreadsheet: load the day's trades, group them by client and show how positions would change if the market fell. The program runs, generates a chart and produces convincing numbers. But a successful run checks only part of the task. Are currencies matched correctly? How are canceled trades handled? Are the units consistent? What happens if the source supplies incomplete data?
What changed at Goldman Sachs?
For several months, bank employees could experiment with these tools. Then internal rules updated in October limited access to engineering roles. According to the same report, an exception remained for strats—specialists who combine trading, quantitative research and software development. The tools already had clear uses. Traders used them to monitor client trading flows, model market changes and positions, and produce text and graphics.
What are the key details on Takeaways for developers and founders?
There is no need to copy the bank's restrictions wholesale. A small team and a large bank have different processes. A practical step is to define when a personal experiment becomes an operational service: colleagues start using it, it gains access to data or its output begins to influence decisions. At that point, assign an owner, check calculations against known examples and agree on who approves changes.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.