Goldman’s AI pullback isn’t about AI. It’s about production.

Reading reports on Goldman this week, my first reaction wasn’t surprise. It was recognition. Goldman has pulled AI coding tools such as Claude Code and GitHub back from its bankers and traders. Internal guidance issued last month reserves them for engineers. Traders had been using them to monitor client flows and put materials together, and some of them are frustrated. I would be frustrated too.

But I don’t think Goldman made the wrong decision, at this time. As recently as July, its CIO was publicly discussing non-technical staff building with AI and handing the work to engineers to finish. They tried it, looked hard at what came out, weighed it against the risk and made a call. That’s what a responsible bank does. What concerns me is what comes next. If the conclusion is that all AI-built software is only ever safe in the hands of engineers, the trading desk that needed that tool, the business that needed that capability is back in the queue, and the need that sent them to it compounds.

AI made software easier to start, not easier to finish

AI has made it remarkably easy to start building software. A trader who can describe a workflow can have working code by the end of the afternoon. That’s genuinely exciting, and we shouldn’t lose sight of it.

What they can’t have by the end of the afternoon is something IT will put into production. Anything that touches client flows at a regulated institution has to be secure, resilient and auditable, integrated with core systems and signed off by compliance. None of that comes for free, and general-purpose AI doesn’t bring it on its own. Every time you ask, you get new code nobody has reviewed, and it’s different each time.

So the handoff becomes the bottleneck. The business produces prototypes, and engineers inherit the job of making them production-ready. Spread that across the institution and the backlog grows faster than any team could clear. Most of those applications never reach production at all. We’ve been here before, with spreadsheets, with macros, with Python dashboards. Every time, the tool got restricted and the need stayed exactly where it was or operational risk increased with side-of-desk coding.

What would have to be true

I don’t think the question is whether business users should build some class of applications. AI is already here and the pressure to build, one way or another will continue to grow. The question is what they build on. I believe there are four necessary conditions before business-built software comes anywhere near production:

• It starts from proven components that already meet production standards, so security, resilience, audit, integration and compliance are there from day one, not regenerated from scratch every time.
• The AI works inside well defined guardrails, so what comes out is controlled and repeatable.
• IT sets the standards and decides what goes live.
• IT can see all of it: the specifications, tests, data model and source code, in technologies they already run.

This is why we built Genesis Create

That list is the brief we gave ourselves when we built Genesis Create. Create is software for building production-ready applications for financial markets. Applications start from proven Genesis components (UI components, business components, services, workflows and application patterns) that already handle security, resilience, audit, integration and compliance. Those components carry more than a decade of what we’ve learned building for financial markets. The AI works inside strict guardrails and only creates the parts that are genuinely unique to the workflow. Reuse what already works. Create only what’s unique.

The business describes what it needs. IT sets the standards and approves what goes into production. And the firm owns the application that comes out: the source code, the prompts, specifications, data model and tests, in standard Java Spring and TypeScript. There’s no black box.

And when you generate less, you have less to look after. In our Bond RFQ benchmark, using the same underlying AI, Genesis Create needed 60% less application code and 77% lower AI generation cost than a general-purpose, AI-only approach. Every line you don’t generate is a line nobody has to test, govern or support.

Where this goes

Goldman won’t be the last bank to face this challenge. Every institution I speak with is wrestling with the same trade-off between speed for the business and risk for the firm. I don’t believe AI-built software has to stay exclusively in the hands of engineers to be safe, understanding that some things are so business critical that they will always stay with IT. But to truly break the backlog, I believe it has to be built on foundations IT already trusts. The banks that pull ahead won’t be the ones that hand their traders the most powerful tools, or the ones that take those tools away. They’ll be the ones that give the business a safe way to build.

That’s the problem we’ve chosen to spend our time on, and I’d genuinely like to hear how your firm is thinking about it.