The North Sea at night under a low sky, one distant light on the horizon

Field notes Governance

Who gets the tool is a control. It should not be the only one.

November 2026  ·  5 min read

Goldman Sachs has limited advanced AI coding tools, Anthropic's Claude Code and Microsoft's GitHub Copilot among them, to its engineers. Traders and bankers who had spent months using them to monitor trading flows and produce content no longer can. Bloomberg reported the change on 2 November 2026, citing internal guidelines issued in October, and The Decoder's account of that report sets out what is known.

The bank's explanation is short. A spokesperson cited a "variety of factors including productivity and risk management", said the experimentation phase was over, and said the tools were now going to the people best placed to use them for the long term. In July the bank's chief information officer had described a different model: a trader with an idea builds a quick prototype with AI and hands it to an engineer to take into production.

Be clear about what this is. It is not a ban on AI at Goldman, and the bank has not said it followed an incident or a regulator's request. Some coverage has claimed both. The reporting we could trace does not.

The control that is always available

What Goldman made is an allocation decision: who may use a capable tool. Allocation by role is the coarsest control there is. One line, drawn at a job title, applied to everyone on either side of it. It is also the one control every organisation has from the first day, because it needs nothing more than a record of who does what.

The finer controls need more. Scoping what a tool may read and change, requiring a named person to sign off before anything a non-engineer builds touches a live system, keeping a record of what was built, with which model and on whose approval: each of those is infrastructure, and it has to exist before the tool is in wide use, not after.

Without that infrastructure the choice narrows to two positions. Let everyone in, or draw the line at a role. A coding agent in a trader's hands is the case where the difference matters most, because the person building the tool is not the person who would normally review code, and what they build may inform decisions on live positions.

Restricting access is a legitimate first answer. The question is whether it is the only answer an organisation has.

The survey says this is common

In October, IBM published a study of 3,221 executives at organisations with 1,000 or more employees across eight countries in Europe and the Middle East. Different firms, different region, the same shape.

Thirty-five per cent of respondents say they run agentic AI in production today, and three-quarters expect agents to have some autonomy over decisions by 2027. "Only 37% say their current governance system is designed to scale with agentic AI deployment." When deployment outpaces governance, 56% press ahead within agreed risk limits and about a third delay deployment entirely until governance is ready. Eight in ten say an AI-related incident reached their board or C-suite in the past year, and fewer than half of those (40%) cite a change to governance policy in response.

Delay and restriction are the same move at different scales. When governance cannot carry the load, the remaining lever is to reduce the load: fewer systems, or fewer people using them.

What a finer line looks like

Scope by what the tool may touch, not only by who holds it. A coding agent that can read market data and write to a sandbox is a different risk from one that can write to a production system. Set the boundary on the system, and the job title matters less.

Make the hand-off a gate. The prototype-to-engineer model is sound. It holds up when the hand-off is a recorded step with a named reviewer, rather than a convention people follow when they remember to.

Keep the record. For anything an agent helped build, keep what was built, by whom, with which model, and who approved it for use. That record is what lets you widen access later and defend the decision if asked.

Widen in stages. Extend access one group at a time, against that record, rather than switching it on for everyone or off for everyone.

We are not neutral

ORCAHQ sells governance, and we build on several model suppliers, Anthropic among them. Read this with that in mind.

Goldman's decision may well be the right one for Goldman. Its own spokesperson framed it as productivity as much as risk, and a bank of that size can absorb the cost of a blunt line. Most organisations will meet the same choice with less room. The useful question is not whether to restrict. It is whether, when the time comes to widen access again, there will be anything finer to widen it with.

All field notes

Tell us what you are accountable for.

The useful first conversation is usually about the thing that worries you, not about our capabilities. If it turns out we are the wrong firm for it, we will say so.

One reply from a person who has read what you sent.

Colophon

  • Plate 06, the sea at night: generated, Azure OpenAI gpt-image-1.5, 3 November 2026.

Every plate on this page is a generated stand-in for commissioned photography. The prompt that made each one is embedded in its file. How this site was made.