Live Nov 18 · Cohort 3Nov 18, 19:00 MSK — free hands-on AI agents live workshop · Live Cohort 3 starts Nov 23Join Nov 18 Live →Cohort 3 (Nov 23)
ReleaseInsight · On RadarOpenAI Blog→ original

Revolut Cuts Feature Development Time by 30–60% and Hands Payment Screening to an AI Agent

Revolut can't hire investigators as fast as payments and fraud attempts are growing. So the bank gave first-line screening of suspicious transactions to an AI agent built on OpenAI models, while its engineers run several Codex agents on parts of a single task. According to the company, the path from starting a task to merging code has shrunk by 30–60%. The figures are self-reported, but the "agent does the work, human stays accountable" model fits any team.

AI-processed from OpenAI Blog; edited by Hamidun News
Revolut Cuts Feature Development Time by 30–60% and Hands Payment Screening to an AI Agent
Source: OpenAI Blog. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceOpenAI Blog ↗
Editor-in-Chief
Key metric #1One engineer built a code-review service by handing tasks out to several Codex agents
Key metric #2Seven years of design documentation became context for AI in a single day
Key metric #3One major code security review with Codex took 1–2 days less

Revolut faces simple arithmetic: payments and fraud attempts are growing faster than the bank can hire investigators and developers. The solution the company described in a case study on OpenAI's website has two parts. An AI agent built on OpenAI models handles first-line review of suspicious transactions (so-called FinCrime, or the fight against financial crime).

And instead of a team of implementers, engineers run several Codex agents at once, each building its own part of a feature. The claimed result: time from starting a task to merge (folding finished code into the project's main branch) drops by 30–60%. One caveat: this is a customer case study on the vendor's own site, and the figures are self-reported and not independently verified.

Still, as a description of a working setup, it is useful to any team whose workload is growing faster than headcount.

An agent instead of a queue of investigators

Revolut operates in more than 40 countries and serves about 800 thousand businesses, and its number of retail customers grew by 23% over the year. According to one report, more than half of adults worldwide encountered fraud in 2025. Every transaction is a potential alarm signal, and no bank has enough human analysts to cover all such cases. David Abitbol, head of AI, says the main problem AI can solve is demand growing faster than the number of people.

That is why one of the company's largest uses of AI is the FinCrime agent. It studies the case data, applies Revolut's internal rules, and works through the prescribed procedure to the end without human involvement. Abitbol calls the release of GPT-5.5 a turning point: that was when models became smart enough and cheap enough to hand screening and first-line investigation over to them entirely. According to the description, the result is more transactions reviewed without a proportional increase in investigator hiring.

What the case study does not say: how many transactions the agent processes, what its false-positive rate is, and what the model queries cost. For a financial firm, that is a significant gap.

The engineer designs, the agents implement

Dmitry Pitik, a senior platform engineer, was building an internal code-review service. He designed the system and broke it into features, and at first assumed he would need a whole team of developers. Then he realized he could run several Codex agents and give each one its own task. He handled architecture, planning, and checking what the agents returned.

This is where he felt the 30–60% time savings. Abitbol points to a general pattern: the time freed up is spent by teams on discussing design and looking for a better solution, not on typing code. While Codex works on one task, the engineer moves on to another. That matters at Revolut: according to Kim Maniovich, the average employee there has "three different jobs."

Other examples from the case study:

  • a tool for building web interfaces for internal and external products;
  • seven years of design documentation turned into AI-readable context in a single day;
  • 1–2 days saved on one major code security review.

The documentation story is the most transferable: an agent needs clean, structured context, and the exercise also shows where an organization's own knowledge is poorly documented.

Accountability stays with the human

Mo Amiri, an information security engineer, describes how new AI applications get approved. He traces the whole chain: what data the system sees, where that data goes, which tools it can reach, what it can do with them, and what happens when it fails. Separately, he checks whether convincing-looking conclusions are backed by real evidence.

Everyone building anything with AI follows a single set of standards for development, A/B tests, and forecasting business metrics. Each developer is responsible for their own code, even if AI wrote it. If an agent proposes a change, the engineer must understand it well enough to vouch for it. In phishing investigations, AI assembles the email, links, attachments, and sender into a structured case, but leaves the check of primary records and the decision to the analyst.

What to take away

Three practical lessons. Start where workload is growing faster than headcount, the cost of error is high, and the process is already formalized in rules. Measure the savings in time from task to merge, and decide in advance what the freed-up time will go toward. And define the boundaries up front: what the agent sees, what it can do, and who signs off on the result.

Next, Revolut wants to build a conversational interface for customers: blocking a card, managing a subscription, or saving for a trip through dialogue. For now these are plans, with no timelines or metrics.

Story

Key Questions Answered

What are the key details on An agent instead of a queue of investigators?

Revolut operates in more than 40 countries and serves about 800 thousand businesses, and its number of retail customers grew by 23% over the year. According to one report, more than half of adults worldwide encountered fraud in 2025. Every transaction is a potential alarm signal, and no bank has enough human analysts to cover all such cases. David Abitbol, head of AI, says the main problem AI can solve is demand growing faster than the number of people.

What are the key details on The engineer designs, the agents implement?

Dmitry Pitik, a senior platform engineer, was building an internal code-review service. He designed the system and broke it into features, and at first assumed he would need a whole team of developers. Then he realized he could run several Codex agents and give each one its own task. He handled architecture, planning, and checking what the agents returned. This is where he felt the 30–60% time savings.

What are the key details on Accountability stays with the human?

Mo Amiri, an information security engineer, describes how new AI applications get approved. He traces the whole chain: what data the system sees, where that data goes, which tools it can reach, what it can do with them, and what happens when it fails. Separately, he checks whether convincing-looking conclusions are backed by real evidence. Everyone building anything with AI follows a single set of standards for development, A/B tests, and forecasting business metrics.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…