Within a few days, two decisions have become known that point in the same direction: AI coding tools are being given tighter limits. According to a report, Goldman Sachs is restricting access to tools such as Claude Code and GitHub Copilot to engineers; traders and bankers are excluded. Almost simultaneously, Anthropic prohibits its coding assistant Claude Code from recreating existing video games.
Both cases concern different levels — one the use within the company, the other the provider's rules. What they have in common is that the initial openness towards so-called vibe coding, i.e. programming largely generated by AI without deep development knowledge, is being readjusted.
From praise to restriction
The turnaround at Goldman Sachs is remarkable. As recently as July, CIO Marco Argenti had publicly praised vibe coding by non-technical staff. Now the bank cites "productivity and risk management" as the reason for the restriction.
The source does not provide more details. It therefore remains open whether specific incidents tipped the scales or whether an internal assessment did not confirm the hoped-for effects. The wording, however, suggests that both were considered together: benefits and risks were evidently weighed against each other — with the result of narrowing the circle of users.
For a financial institution, this is understandable. Code created in regulated environments is subject to review, documentation and approval obligations. When business departments independently create analyses or small tools, artefacts emerge that run outside established review processes. Those who have no control mechanisms for this tend to draw the line restrictively.
The provider draws its own limits
The second report concerns Anthropic itself. The company is reacting to a wave of game recreations and prohibits Claude Code from cloning existing video games. Here too, the published information is sparse; further details on the technical implementation or the scope of the rule are not evident from the source.
The point is nevertheless relevant: the terms of use of an AI service can change, and without any action on the part of customers. What was possible yesterday may be blocked tomorrow. For companies that firmly integrate AI assistants into their development processes, this is a planning factor.
Two types of risk that are often conflated
The two reports concern different risk categories that are frequently mixed up in internal discussions:
- Organisational risk: Who may use which tools, and what happens to the generated code? This is about responsibilities, quality assurance and operations.
- Legal and compliance risk: What may be created with a tool? This is about third-party rights, terms of use and the question of whether a result may be used at all.
A policy that covers only one of the two levels leaves a gap. The Goldman Sachs case shows the first level, the Anthropic case the second.
Roles instead of blanket bans
A complete exclusion of entire user groups is the simplest solution, but not necessarily the most effective. A graduated approach based on the purpose and the impact of the generated code is usually more effective.
A distinction into three zones is practicable:
- Experimental zone: Prototypes, data analyses and scripts without production integration. Here a broader circle of users is justifiable, as long as the results do not flow into decisions or systems without being reviewed.
- Transition zone: Code that is to go into production. From this point on, review by developers, version control and tests apply — regardless of who created the draft.
- Core zone: Production systems, customer data, regulated processes. Here the existing development standards apply without exception.
The decisive difference lies not in the job title, but in where the generated code flows. Those who separate this cleanly do not need to lock out departments across the board.
What belongs in an AI coding policy
Several points can be derived from the two cases that are often missing from internal rules:
- Approved tools: A binding list of which assistants may be used — and via which company access points.
- Purpose limitation: For which tasks use is intended and which are expressly excluded.
- Review obligation: From what point human review is mandatory, and who is responsible for it.
- Origin and rights: Clarification that recreations of protected works and the adoption of third-party content are not permitted — regardless of whether the tool blocks them technically.
- Resilience to change: A procedure for the case that a provider adjusts its terms of use or restricts functions.
- Traceability: Documentation of which parts of an application were created with AI support.
The productivity question remains open
Goldman Sachs expressly cites productivity as a reason — not only risk. This is an indication that the expected efficiency gain among non-technical staff evidently did not materialise as hoped. There are no reliable figures on this, and no general rule can be derived from a single decision.
What can be noted, however, is that the period between the public praise in July and the restriction was short.
Anyone introducing AI tools should measure the contribution they actually make from the outset — instead of basing the decision on expectations alone.
Context for web and software projects
For projects based on TYPO3 or custom web applications, for example, this means: AI assistants speed up drafts, but they do not replace quality assurance.What matters is a policy that defines roles, areas of use and review obligations and at the same time clarifies how third-party rights and changing provider rules are handled.Anyone who sets these guardrails early will not have to lock out entire user groups later and retains control over what reaches production operation.