AI Got You to a Prototype. Here’s What It Takes to Get to Production.

A quiet software revolution is underway in capital markets. Traders, analysts and operations managers who once needed an engineer to translate their requirements into code can now describe what they want in plain language and watch a working prototype take shape in minutes. That’s not hype. It’s real and it’s spreading fast across banks, asset managers and vendors alike.

Give vibe coding its due. It is genuinely good at what it does. A business user can prompt an AI assistant like Claude, Gemini or Replit and generate a functioning prototype complete with a custom workflow and a usable interface, often faster than it takes to schedule the first requirements meeting with IT. For low-stakes tooling and early-stage proofs of concept, that speed is a real advantage and firms are right to want more of it.

However, trouble starts the moment that prototype needs to become production software.

The bottleneck between “it works” and “it’s ready”

We heard this dynamic described bluntly in a recent Waters Technology feature on vibe coding across capital markets. Engineers interviewed for the piece described real productivity gains alongside real exhaustion. One senior engineer at a data company put it simply: output has gone up, but so has the effort required to get there, because every prompt-generated script, workflow or interface still has to be reviewed for compliance, performance and security before it can be trusted. Another engineer at a global bank noted that the underlying complexity of financial software hasn’t gone away just because the code is easier to produce. It has simply moved downstream, into code review, into maintenance and into the teams responsible for keeping systems audit-ready.

That downstream set of bottlenecks are exactly why prototypes stall before they reach production. A vibe-coded application might work beautifully in a demo. But does it authenticate against the firm’s entitlements model? Can it withstand the transaction volumes a trading desk generates on a busy day? Does it produce the audit trail regulators expect under MiFID II, SEC or FINRA rules? Was it built on an architecture that a different engineer can maintain two years from now, or is it a one-off that only the original prompter understands? These aren’t edge cases. They’re the baseline requirements of financial markets software and vibe coding, by design, has no mechanism for enforcing any of them.

As a result, the age-old friction between business and IT is only growing. Businesspeople, for the first time, can generate facsimiles of what’s possible, but IT is forced to police a rapidly-growing backlog of non-compliant prototypes.

Gartner research puts a number on where this is heading. The firm’s 2026 Strategic Predictions warn that prompt-to-app development by citizen developers will drive a 2,500% increase in software defects by 2028, and the reason is instructive: AI-generated code is often syntactically correct but architecturally blind. It doesn’t understand the broader system it’s being dropped into or the business rules that govern how that system is supposed to behave. Those aren’t simple bugs needing a quick fix. They’re deep, contextual flaws that get more expensive to remediate the longer they sit in a codebase, and they consume the very budget firms had earmarked for innovation.

Two different kinds of intelligence

Here’s the tension underneath all of this. Vibe coding runs on probabilistic tools. An AI model generates the most likely correct answer based on patterns, and it does that extraordinarily well for well-understood, well-documented problems. Financial markets software, on the other hand, runs on deterministic requirements. A trade allocation either satisfies T+1 settlement rules or it doesn’t. An access control either enforces the right entitlement or it doesn’t. There’s no “probably compliant.”

Asking a probabilistic tool to single-handedly produce a deterministic outcome is where the bottlenecks come from. It’s not a knock on the tools. It’s a mismatch of what they were built to do and what capital markets demands.

Where Genesis fits

This is precisely the gap Genesis can close. Our framework doesn’t ask firms to choose between the speed of AI and the rigor financial software requires. It combines both.

At the foundation sits our library of pre-built, battle-tested software components: pieces of production logic already engineered, validated and hardened at some of the world’s largest financial institutions. These components carry the deterministic guardrails that AI-generated code lacks on its own, including security, entitlements management, deterministic processing paths, full observability and immutable audit trails.

Operating within that framework, AI can produce applications which inherit that governance automatically.

That’s the model behind Genesis Create, our application builder for capital markets. It pairs AI-powered assembly with our component library and keeps human oversight in the loop, so business users can do what they are best at, defining requirements, designing workflows and validating functionality.

With Create, AI isn’t writing applications from scratch. By design, we guardrail AI, focusing it on modifying specific components at specific extension points to meet business needs. Directing AI in this way has two distinct benefits: costs are lower, because our approach uses fewer tokens, and the output is standardized, production ready code, because AI is not deviating from the deterministic foundations in the Genesis framework.

Create will generate a specifications document and the code is exportable to any IDE, so that, for highly complex applications, engineers can take that same output and harden without starting over.

In this way, business and IT stop building past each other and start building toward the same result, on the same standards-based foundation.

The real promise of AI in capital markets

None of this is an argument against AI-assisted development. It’s an argument for constraining it properly. The firms that will win with AI in financial markets aren’t the ones generating the most code the fastest. They’re the ones that can take a business user’s prompt and turn it into something a compliance officer, a security team and an engineer three years from now can all still trust.

Vibe coding proved that describing software in plain language and watching it appear is possible. The next challenge for capital markets isn’t proving that AI can build applications. It’s finding a governed path from idea to production that complements the models, tools and infrastructure firms already use. That’s the challenge Genesis was built to solve.