About a-team Marketing Services
The knowledge platform for the financial technology industry

A-Team Insight Blogs

Beyond Vibe Coding: The Production Challenge for AI-Generated Trading Applications

Subscribe to our newsletter

Goldman Sachs’ recent decision to restrict access to advanced AI coding tools among most of its traders and bankers has highlighted growing concerns over the proliferation of business-developed software within financial institutions. Similar challenges are emerging at JPMorgan Chase, where developers have reportedly expressed frustration at being asked to productionise applications created by non-technical staff using AI. Meanwhile, other banks, including MUFG, have been exploring more tightly controlled approaches to AI-assisted development as they seek to balance productivity gains against security and governance requirements.

The developments expose a growing dilemma for financial institutions. Generative AI is making it easier for front-office staff to create their own applications, while the technology requirements of trading businesses continue to outstrip the capacity of central IT teams. Restricting access to AI development tools may contain immediate risks, but it does little to resolve the underlying demand for faster software delivery.

David Perkins, Executive Vice President for Sales & Strategic Growth at Genesis Global, the capital markets application development technology provider, sees parallels with an earlier generation of business-led technology initiatives.

“This isn’t a new problem,” he tells TradingTech Insight. “I’ve seen this happen in previous roles, where desks were very frustrated with where their technology demands sat on a priority list or a budget list. Influential front-office people would use their desk budgets to pay for a technology person to sit on the desk and start building pricing spreadsheets or small workflow applications. It was all well and good until those applications needed to be industrialised. Then the work would get handed over to technology, which suddenly had to support code it hadn’t written and whose quality it didn’t know.”

The challenge of production readiness

The speed of AI-assisted development potentially magnifies these longstanding problems, particularly where individual desks or business units pursue their own solutions without coordinating with central technology teams. Perkins points to the possibility of credit and equities desks independently developing similar workflows, creating multiple implementations of functionality that could otherwise be shared across the organisation.

There are also important differences between the risks associated with internal applications and those deployed into external markets. Some institutions are already using AI-generated software for internal tasks, such as aggregating prices from multiple brokers. Applications that interact directly with counterparties, however, face more demanding requirements around resilience, security, regulatory compliance and operational accountability.

An application may perform its intended function in a controlled environment without being equipped to handle unexpected inputs, integration failures, changing market conditions or the demands of continuous production operation. These shortcomings may not be immediately apparent to the business users who developed it.

“I think the initial reaction is often that what they’ve built looks really good,” notes Perkins. ‘People say, look at this, isn’t it impressive? But it’s really just the thin veneer of a solution. Anyone can build something. Not many people can build something and deploy it safely. The idea that building the application is the heavy lift is completely wrong. That’s the easy bit. The hard bit is deploying it.”

The growing availability of AI coding tools also raises questions about whether institutions are measuring development productivity appropriately. Faster initial delivery may offer limited economic benefit if the resulting software requires extensive engineering intervention before it can be supported, integrated and maintained. Without consistent architectural standards, an increase in application development capacity could ultimately add to the technical debt that financial institutions have spent years attempting to reduce.

Separating AI-generated functionality from enterprise controls

One emerging approach is to limit the scope of AI-generated code, using established software components for functions that require consistent, predictable behaviour. The distinction is particularly relevant in regulated environments, where security, auditability and access controls must meet defined requirements rather than depend on the variable output of a generative model.

Genesis has adopted this model through Genesis Create, an AI-assisted development offering built on the company’s existing capital markets application framework. Developed from its earlier low-code and no-code technology, Create combines AI-generated business functionality with reusable software components for underlying infrastructure and controls.

“Rather than having an AI large language model code an entire application from a prompt, we encourage the LLM to look first within our existing set of tools and applications and use as many of those as possible,” explains Perkins. “That means the amount of new code being written is very limited. There are smaller amounts of code to maintain, validate and check, and the client is left owning that code and the IP, and an understanding that it’s safe to deploy.”

Under this model, AI can translate business requirements into application-specific workflows, configurations and interfaces, while pre-engineered components provide capabilities such as entitlements, integrations and audit functions. Genesis argues that this reduces the amount of bespoke development required and allows technology teams to focus their scrutiny on the elements that are unique to each application.

However, the use of established components does not remove the need for independent validation. Errors in business logic, incorrect configurations or poorly designed integrations can still introduce significant operational risks, irrespective of whether the underlying security and audit functions have previously been tested. Institutions must also consider how changes to shared components affect applications that depend on them.

Balancing business autonomy and IT oversight

The organisational implications extend beyond the choice of development tools. Business users may be best placed to define application requirements, but technology teams remain responsible for architecture, deployment standards and the ongoing integrity of production systems. Finding an effective division of responsibilities becomes increasingly important as AI enables more employees to participate directly in software creation.

For Genesis, this has led to a development model in which clients can build applications themselves, with specialist engineering support available where required. Perkins describes a recent discussion with a financial institution considering a new trading application.

“We had a meeting with a client this week about building an RFQ platform. We could build that platform ourselves using Create, but as clients become more agile, technology teams increasingly want to own the IP of what they build. So in this case, we suggested that instead of us building the RFQ platform, the client should build it themselves. They could take a Create Builder licence, and we could provide a forward-deployed engineer to work alongside them, making sure they know how to use it properly. We’d provide guidance, helping their people build exactly what they want.”

The example illustrates a potential shift in the relationship between financial institutions and their technology suppliers, with firms taking greater responsibility for applications developed using third-party frameworks. It also introduces questions about the skills required internally, the boundaries of vendor support and the extent to which ownership of application code translates into operational independence.

Genesis says clients can maintain their own applications and retain ongoing access to the underlying components, although upgrades depend on their continuing arrangements with the vendor. Clients can also replace those components with internally developed alternatives if they wish to reduce their reliance on the platform. Such flexibility may be attractive to institutions seeking greater control over their technology estates, but maintaining a mixture of proprietary and vendor-supplied software introduces its own architectural and support considerations.

The broader challenge for financial institutions is how to accommodate a growing volume of business-generated software without undermining the standards that govern production technology. AI may help address longstanding development backlogs, but the benefits will depend on whether banks can establish processes that translate faster application creation into reliable, maintainable systems. Restricting access to individual tools may provide some immediate protection; it is unlikely to diminish the commercial pressure for more responsive technology delivery.

Subscribe to our newsletter

Related content

WEBINAR

Recorded Webinar: Generative and Agentic AI in Financial Markets: What the Data Really Shows

Artificial intelligence is reshaping financial markets – but the reality on the ground is more nuanced, more uneven, and more instructive than the headlines suggest. A new A-Team Insight research programme, drawing on responses from senior AI decision-makers at buy-side and sell-side institutions across the USA, UK and Canada, reveals a market in transition: from...

BLOG

TT Connects to Kalshi as the Institutional Case for Event Contracts Takes Shape

Trading Technologies has announced that it will support client execution on Kalshi, the largest US federally regulated prediction market, with trading expected to go live on the TT platform in the third quarter. It is the first regulated prediction market TT has connected to, and the firm frames it as the first of several regulated...

EVENT

Buy AND Build: The Future of Capital Markets Technology

Buy AND Build: The Future of Capital Markets Technology London examines the latest changes and innovations in trading technology and explores how technology is being deployed to create an edge in sell side and buy side capital markets financial institutions.

GUIDE

AI in Capital Markets Handbook 2026

AI adoption in capital markets has moved into a more disciplined phase. The priority is now controlled deployment: where AI can be used safely, where it can deliver measurable value, and how outputs can be governed, monitored and evidenced. The 2026 edition of the AI in Capital Markets Handbook examines how AI is being applied...