A 30–60% reduction in coding time from task start to feature merge outlines the operational gains Revolut is extracting from OpenAI models and Codex. The digital bank operates across more than 40 markets, serving 800,000 businesses alongside a consumer base that grew by 23% over the last 12 months. This rapid expansion brings an inevitable operational bottleneck: as transaction volumes surge, staffing human compliance teams to review every suspicious event becomes financially unviable.
Transaction monitoring and anti-fraud operations face severe structural strain. According to one report, over half of adults worldwide reported a scam encounter in 2025. The volume of incoming transactions far outstrips the human capacity available to any banking institution, creating an urgent operational challenge: can autonomous agents handle complex investigations without inflating headcount or missing critical compliance failures?
Integration of OpenAI models inside Revolut demonstrates a transition from experimental tooling to structural cost control in financial crime operations. Delegating entire investigation workflows to autonomous agents provides a functional template for running regulated financial services without expanding investigator payroll in lockstep with client growth.
Delegating the Investigative Pipeline
To decouple transaction volume from compliance headcount, Revolut deployed a FinCrime agent driven by OpenAI models. Rather than requiring human staff to review every single flagged transaction manually, the system evaluates incoming data, executes internal rulebooks, and concludes designated case processes autonomously.
"The biggest problem we can tackle with AI today is the scaling of humans versus the demand that we get," says AI Lead David Abitbol.
This architecture changes the unit economics of fraud operations. David Abitbol explains that the release of GPT‑5.5 was a turning point, providing the technical competence and cost profile needed to hand over transaction screening and complete investigations entirely to the agent. Information Security Engineer Mo Amiri notes that this automation directly strengthens protections and produces a resilient operating model.
How does an engineering organization maintain consistency across disparate internal applications while delegating authority to automated agents?
By enforcing uniform technical constraints across all development pipelines. David Abitbol points out that Revolut maintains a single framework applying identical development, A/B testing, and business metric prediction standards to every employee shipping internal tools with artificial intelligence.
Refactoring Engineering Overhead
Operational capacity is shifting in core software delivery as well. Codex cut coding time by 30% to 60% and saved 1–2 days on security code reviews. It also took Revolut just one day to convert seven years of accumulated design documentation into structured context readable by artificial intelligence.
Senior Platform Engineer Dmitry Pitik observed these mechanics firsthand when architecting an internal code-review service. He originally planned to allocate an entire development team to build the components. Instead, Dmitry Pitik drafted the architecture and assigned individual feature implementations to different autonomous agents acting as a virtual development group.
"I realised I don’t need a whole team of developers to implement this. I can just launch different agents and assign each of them a particular task," says Senior Platform Engineer Dmitry Pitik.
Engineering leadership emphasizes that automated generation does not eliminate rigorous oversight. Dmitry Pitik points out that developers remain strictly accountable for their code and do not blindly trust model outputs. Kim Maniovich, Revolut’s Global Head of Business Marketing, notes that partnering with OpenAI enables the company to construct solutions in a lean, cost-efficient manner aligned with small operating teams.
Audit internal developer workflows and transaction queues to identify tasks consuming 1–2 days of routine manual review. Establish a unified baseline for testing and metric evaluation, then test automated model delegation on closed, repetitive screening stages before altering investigator team budgets.
How the picture shifted
While by mid-June Codex reached 5 million weekly active users, the digital bank deployed this mechanics directly across its footprint of over 40 markets. Converting seven years of internal documentation into structured context for AI took the team exactly one day, according to the company, breaking the dependency between surging transaction volume and compliance headcount.
Our stories: OpenAI Codex: A New Era of Business Automation (June 13).