Goldman Sachs has restricted access to AI coding tools largely to engineers and quantitative specialists, reversing an earlier push for more employees to build applications with them. The shift puts a practical question at the centre of enterprise AI: who gets to use these tools, and what systems can they reach?
Watch Desk analysis
What happened
Goldman had encouraged bankers and traders to build prototypes using AI, with the bank’s engineers expected to refine them. PYMNTS reports, citing Bloomberg, that Goldman has now limited access to Anthropic’s Claude Code and Microsoft’s GitHub coding tools largely to technical specialists. A Goldman spokesperson told Bloomberg the decision followed an internal review of productivity and risk management.
The account describes other banks taking different approaches. BNY lets employees outside engineering build within its Eliza platform, which PYMNTS says has more than 400 AI-enabled solutions in production. Deutsche Bank gives more than 15,000 developers coding assistants and says developers remain responsible for code they commit. MUFG built an in-house alternative to Claude Code after finding it could work around information-access limits and compliance safeguards, according to a July report by WatersTechnology cited by PYMNTS.
Why it matters
Making code easier to produce does not make it safe to deploy. An AI-generated application could use a wrong pricing assumption, expose confidential information or bypass approval controls. For a bank, those are not merely untidy software bugs.
PYMNTS also cites a survey of 60 US CFOs at companies with at least $1 billion in revenue: among those at firms using or piloting AI agents, 43% would allow moderate system access with human approval, and none said they would grant full access. That is a small, specific survey, not a census of finance leaders, but it points to the same tension: interest in automation, paired with a firm hand on permissions.
Our read
This is not evidence that Goldman’s wider experiment failed, and PYMNTS says the bank has not disclosed a security incident or quantified productivity findings behind the restriction. It is evidence that the case for putting coding tools in more employees’ hands depends on the controls around them, not just how quickly they can produce a prototype.
Banks are trying several models: restrict access to specialists, build inside a governed platform, or develop an in-house alternative. The useful test is what each approach lets employees do, what data and systems remain off limits, and who is accountable when generated code misbehaves. “The AI wrote it” is unlikely to impress a compliance team, or a customer whose account is wrong.
What to watch
- Whether Goldman explains what its internal review found and whether access rules change again.
- How banks limit coding tools’ access to customer data and internal systems.
- Whether firms publish evidence of productivity gains alongside information about errors and human review.
Discussion spark: Should banks let more employees build with AI coding tools inside tightly controlled platforms, or keep access mainly with technical specialists?
Sources and evidence
- Source update (3 November 2026, 20:52 UTC)
- Releases · unslothai/unsloth · GitHub (3 November 2026, 20:52 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.