[KongchangAI]
· 2 min read· 1,171 words

How Revolut Uses AI to Streamline Teams and Tackle Bigger Tasks

How Revolut Uses AI to Streamline Teams and Tackle Bigger Tasks

Revolut partners with OpenAI to achieve efficient delivery with a streamlined team using autonomous agents and Codex.

Revolut's collaboration with OpenAI shows how a streamlined team can redistribute work with AI agents. Its cybersecurity team delegates transaction screening and investigation to a financial crime agent, and a single engineer built an internal code review service with Codex and multiple agents—cutting coding time by 30-60% and reaching delivery speeds ~3x market standard. Success isn't just speed, but decision quality and system resilience.

Fintech company Revolut's collaboration with OpenAI offers a worthwhile case study of the "small but elite" team model. This segment from YouTube doesn't stop at the slogan of "AI boosts efficiency"—it provides frontline observations specific to roles, workflows, and decision quality.

Cybersecurity: From Manual Screening to Autonomous Agents

Revolut's cybersecurity team has a clear goal—protecting the company and its customers. After partnering with OpenAI, they gained first access to the latest frontier models, each generation bringing capabilities previously thought impossible.

The most representative implementation is the "Fin Crime Agent." This is an autonomous agent responsible for transaction screening, monitoring, and investigation. The team delegates initial investigation and screening work to the agent, equipping it with a "hard drive"-style set of tools and data permissions so it can independently handle most of the process.

Initial investigation and screening delegated to the agent

The turning point came during the GPT-5.5 era. The team began to feel that the model's intelligence level and cost had reached an acceptable threshold—enough to delegate preliminary investigations. Previously, the "dirty work" of cleaning up rules, sources, and tools often took 3 to 4 hours; now the agent can complete these investigation steps autonomously. For customers, this translates directly into stronger protection and more credible system resilience.

But the team's definition of success goes beyond "faster." They place greater emphasis on decision quality: identifying what truly matters earlier and continuously distilling lessons learned into a more secure platform.

Success isn't just about running faster

Autonomous Agent refers to an AI system capable of perceiving its environment, making decisions, and executing actions to achieve specific goals—different from chatbots that merely generate responses based on a single prompt. In financial crime scenarios, the agent needs to invoke multiple tools (such as database queries, rule engines, and risk scoring models) and autonomously plan its path through multi-step reasoning. The "hard drive"-style tools and data permissions Revolut configured for it effectively define the data boundaries the agent can access and the set of operations it can execute—ensuring the agent can independently complete most of the process while controlling risk through permission isolation. Agent autonomy typically falls into semi-automatic (humans approve key steps) and fully automatic (end-to-end execution); the financial crime agent in this article belongs to the latter in its application to the initial investigation phase.

Internal Tool Development: One Person Is a Whole Team

Codex's use within Revolut also deserves attention. An engineer building an internal code review service used Codex as his primary tool. His takeaway: you don't need an entire team of developers—just launch different agents, assign each a specific task, and wait for the implementation.

You don't need an entire team of developers

The change this working style brings is at the capacity level. As the team uses Codex more in daily work, coding time dropped by 30% to 60%, and time spent searching for information across different knowledge bases also fell significantly. The team thus believes delivery speed can reach roughly 3x the market's conventional standard—and this happens frequently.

Codex is a series of code generation models developed by OpenAI that can generate, complete, and modify code based on natural language descriptions or existing code context. In Revolut's case, the engineer used Codex as the primary tool to build an internal code review service. The working pattern isn't having Codex generate the entire service in one shot, but rather launching multiple agents and assigning each a specific task (such as generating a module, writing tests, or reviewing code style) to advance in parallel. This "agent orchestration" approach reduces dependence on team size, but requires engineers to have clear task decomposition skills and the ability to review generated results. The statistic of 30% to 60% reduction in coding time typically covers end-to-end time from requirement understanding to code commit—not merely time spent typing.

Goals Are Getting Higher—But Also More Achievable

One easily overlooked detail is the chain reaction at the management level. The team lead mentioned that month after month, the goals set for the team are becoming more aggressive—but also more achievable. Standards are rising, and AI makes that upward shift possible.

Time spent searching for information dropped significantly

Behind this is a simple but important judgment: AI changes the way everything is done. The statement may sound clichéd, but it's happening in every link of the chain. What's truly scarce is no longer stacked manpower, but the ability to decompose tasks clearly and orchestrate agents well.

Lessons from the Streamlined Team Model

Revolut's case shows that the path for streamlined teams to tackle bigger tasks isn't simply layoffs or squeezing efficiency—it's redistributing work: handing rule-clear, highly repetitive investigation and screening to autonomous agents, and reserving human energy for decision-making and system design. The cost is the upfront effort required for standardization and tool configuration; the reward is a dual improvement in response speed and decision quality.

For teams also exploring AI implementation, the value of this case lies in demonstrating a replicable approach—first identify the "time-consuming but logically clear" steps in the workflow, then assess whether model intelligence and cost have crossed the usable threshold, and finally use agents to take over execution while humans control the results.

The "small but elite" team model in the AI context means team size doesn't grow linearly with task volume; instead, by handing rule-clear, highly repetitive steps to AI agents, a small number of people can cover a broader scope of work. This model runs counter to the traditional "mythical man-month" logic—it doesn't rely on adding manpower to shorten timelines, but rather redraws the boundary between human and agent responsibilities. Prerequisites for replicability include: sufficient "time-consuming but logically clear" steps in the workflow, model intelligence and cost crossing the usable threshold, and the team's ability to decompose tasks into standard instructions. The cost is upfront effort in process standardization, tool configuration, and permission design; the reward shows up in response speed, decision quality, and system resilience. This model is especially valuable as a reference for rule-intensive industries such as finance and law.

One-line Summary

Revolut partners with OpenAI to achieve efficient delivery with a streamlined team using autonomous agents and Codex.

Paragraph Summary

Revolut's collaboration with OpenAI demonstrates how a "small but elite" team can redistribute work with AI agents. The cybersecurity team delegates transaction screening, monitoring, and initial investigation to a financial crime agent, crossing the intelligence and cost threshold during the GPT-5.5 era—work that previously took 3 to 4 hours of cleanup can now be done autonomously. An internal code review service was built by a single engineer using Codex and multiple agents, cutting coding time by 30% to 60%, significantly reducing information search time, and enabling delivery speed roughly 3x the market's conventional standard. The team lead notes that goals are getting higher but more achievable, and that success isn't just about speed but also decision quality and system resilience. The key to a streamlined team is delegating repetitive processes to agents while focusing human effort on decision-making and system design.

#Revolut#OpenAI#autonomous agents#Codex#streamlined teams#AI efficiency
Share:

Related articles