ING vibe coding pushes AI into trading systems at banks

ING Groep NV is using so-called vibe coding to build trading tools for currencies and credit, moving generative AI from productivity theatre into the market systems banks use to quote clients, price risk and defend share in electronic trading.

Bloomberg’s report on ING’s trading-system buildout points to a cleaner read. The loud phrase is “vibe coding”, shorthand for staff describing what they want and letting a model generate the first version of the code. More important is where that code is headed. Once AI-generated software touches foreign exchange, credit dashboards and pricing workflows, the discussion moves from office automation to market plumbing.

Two readings follow from the same evidence. Inside ING’s electronic-trading desk, AI is a way to compress build cycles and reduce dependence on large vendor stacks. Risk teams, supervisors and rival banks will hear a duller question with higher stakes: can faster code still be tested, governed and contained when markets are moving?

In The Full FX interview with Simon Bevan, that tension is hard to miss. ING’s global head of electronic trading framed the tool as an everyday desk capability, not a pilot parked in an innovation lab.

“At this point, our quant desk is pretty much using vibe coding on a day-to-day basis.”

Simon Bevan, The Full FX

For bank-technology watchers, the point is no longer that AI can draft code. That became obvious across software teams over the past year. A European lender with a relatively lean electronic-trading group is now willing to let the technique shape the systems behind currencies and credit. In a business where execution quality, client response time and product coverage turn into revenue, hours saved in development can become a strategic asset.

From demos to dealing rooms

Speed is ING’s first claim. Work that previously took weeks can now be done in hours, according to Bloomberg, and the projects include tools for foreign exchange and credit traders rather than generic back-office workflows. The distinction matters because trading technology is part of the product. A pricing model, dashboard or client-facing tool can change what a bank is able to quote, how quickly it responds and which flows it can profitably handle.

Bevan’s desk is not a huge engineering division. Yahoo Finance, summarising the same Bloomberg reporting and ING commentary, put the electronic-trading team at about 10 people and cited a 50 per cent increase in large-ticket trades after an in-house AI currency-pricing model. These are not bank-wide automation numbers. They are small-team leverage numbers.

Competitive stakes show up there. Large dealers usually enjoy the advantage of scale: bigger technology budgets, deeper quant benches, more capacity to absorb failed experiments. A smaller team that can turn a trader’s specification into a working tool in a day does not erase the scale gap. It chips at one of the gap’s foundations.

Vendors will notice first. Bevan told The Full FX that in-house AI-assisted development could pressure external providers if banks can assemble more of the stack themselves.

“I think this is a big risk to providers.”

Simon Bevan, The Full FX

No one should expect a sudden bank exodus from vendor systems. Financial institutions rarely rip out core platforms because a model can produce a demo. Marginal work is the real battleground: dashboards, workflow tools, internal pricing aids and desk-specific utilities, the kind of glue software vendors charge for and bank teams complain about waiting for.

The build-versus-buy line shifts

Across Wall Street and European banking, talk of agentic AI, copilots and productivity has often sounded like procurement language: licenses bought, training sessions scheduled, consultants hired. Bloomberg separately reported that AI trainers are charging banks as much as $25,000 a day to push employees toward agentic workflows. ING’s case is narrower and more useful because it shows where the budget argument lands.

Using a commercial model such as Anthropic’s Claude to speed development changes the internal debate. Previously the question was whether to buy a tool, build it in-house or leave the workflow alone. Now a small desk may be able to build enough of the first version itself that vendors are needed only for the hardest parts: connectivity, resilience, compliance, scale and long-term maintenance.

Banks do not become software companies overnight. More desks, though, start to behave like product teams. A trader who can describe a workflow, a quant who can check the logic and an engineer who can harden the output become a faster unit than a traditional request queue. Operational risk sits in the same compression. Requirements, code generation, review and deployment move closer together. If one link is weak, the error travels faster too.

The consumer version of the AI-agent story can mislead here. Robinhood’s move to let AI agents trade stocks on users’ behalf is about delegated retail action. ING’s use is about institutional infrastructure. One sits at the edge of the customer account. The other sits inside the bank’s machinery.

Control becomes the bottleneck

Skeptics are not arguing that AI-generated code is unusable. Their case is that the bottleneck moves from writing code to proving it should be trusted. In low-latency markets, a badly tested tool can create bad quotes, wrong hedges or broken workflows before a committee has finished describing the issue.

Bevan’s comments suggest ING understands that. The Full FX described the deployment as supervised, and Bevan said early results surprised him.

“I was surprised by how good the results were.”

Simon Bevan, The Full FX

Early results are still not production proof. Bank supervisors will want to know who approved the generated code, how it was tested, what logs exist, whether model prompts are retained, and whether a human can explain the system after a market incident. Those questions are not glamorous, but they decide how far this moves beyond desk tools.

Recent security reporting gives the caution a practical edge. The Register described patched bypass bugs in Claude Code’s sandbox, a reminder that coding agents are software supply-chain participants, not neutral clerks. A vulnerability in an AI coding environment is not the same as a vulnerability in ING’s trading systems. Still, it illustrates why banks cannot treat the model as a black-box intern with unlimited permissions.

A new hierarchy follows inside the bank. The fastest teams will not simply be the teams with the best prompts. They will be the teams with testing harnesses, permission controls, audit trails and senior engineers who know when generated code is good enough for a prototype and when it is unsafe for a market-facing workflow.

Staffing signal follows the software

Labour is harder to separate from the technology. Bloomberg said ING is cutting 1,250 operations roles in 2026 as part of a wider AI and digitalisation push. Standard Chartered has also targeted more than 7,000 jobs as banks automate more of their core systems. Those cuts are not evidence that vibe coding caused job losses. They do show that banks are already translating AI adoption into operating-model decisions.

Trading desks may feel the staffing shift more subtly. A bank may need fewer people to maintain routine internal tools, but more people who can validate model outputs, manage entitlements and understand both market microstructure and software risk. That is not a clean substitution. It is a change in the kind of labour banks prize.

Supervisors will follow the risk, not the marketing label. They are unlikely to care whether ING calls the practice vibe coding, AI-assisted development or something more sober. Accountability is the test. Who owns the tool? Who signs off changes? How quickly can the bank disable a flawed model-generated component? What happens when two AI-assisted systems interact in stressed markets?

That final question makes this a market story rather than only an ING story. One bank using AI to speed a dashboard is operationally interesting. Many banks using similar models to generate similar tools, test similar strategies and respond to similar data could make trading infrastructure more correlated. The efficiency gain is real. So is the possibility that everyone becomes faster in the same direction.

ING’s experiment is best read as a live case study in the next bank-technology race. The winner is not the institution that lets AI write the most code. It is the institution that turns shorter build cycles into better tools without letting the control function become an afterthought.

Vibe coding sounds flashy. Banking tends to reward something plainer: a faster system that survives contact with clients, regulators and a bad day in the market.