UK-based Starling Bank has deployed an AI agent that investigates transaction-monitoring alerts and reaches a decision within 17 seconds.
The bank’s agentic financial crime investigator receives an alert from Starling’s transaction-monitoring systems, carries out the investigation, produces the audit trail, and issues a decision. The agent is built natively into Starling Bank’s core banking engine instead of being connected through external integrations. Since the bank owns both the data and the model, the agent can directly access real-time telemetry, historical behaviour patterns, and transaction context. This avoids the delays that come with retrieving data from separate systems.
Targeting investigation and documentation
In AML operations, an alert typically moves through detection, triage, investigation, documentation, and a final decision. The outcome can be closure, escalation, or a suspicious activity report to the National Crime Agency. During the investigation, analysts gather KYC files, account history, counterparty, device, and login data, and previous alerts. They then record the reasoning behind each outcome. These two stages carry most of the cost, and backlogs build up there when alert volumes outgrow staffing. Starling’s agent manages all steps from investigation, documentation, to decision, in a single pass.
Bernadette Smith, Chief Customer and Banking Officer at Starling Bank, said the bank expects the future of financial crime compliance to depend on native AI agents operating in milliseconds, not on larger analyst teams or faster rule engines.
With the above in mind, several details have not been made public. These include whether the agent runs on a third-party large language model or an in-house system, the extent of human review of its decisions, and figures such as alert volumes, false-positive rates, or investigation times before the deployment.
Part of a wider AI programme
In October 2026, Starling announced it is working with NVIDIA and Civo to train a transactional foundation model on more than 15 billion anonymised Starling transactions, hosted on Civo’s UK sovereign cloud. The model is intended to power a unified Starling Intelligence layer, which would sustain the bank’s AI capabilities. Fraud detection is the first use case, and deployment is expected in early 2027.
Earlier in 2026, the bank introduced a Scam Intelligence agent within Starling Assistant. The agent runs on Google’s Gemini models and is available to around five million customers. With the new investigator, Starling now applies AI both before a payment is made and in investigations after the fact.
Regulatory background
The deployment follows regulatory action against the bank. In 2024, the Financial Conduct Authority (FCA) fined Starling around EUR 34 million over failings in its financial crime systems and controls between December 2019 and November 2023. The final notice identified several issues. Until early 2023, customers were screened against only part of the UK sanctions list. The bank had no management information on sanctions alert volumes and trends, and no formal methodology for testing and calibrating its screening systems. It also opened around 55,000 accounts for high- or higher-risk customers, despite agreeing with the FCA in 2021 not to onboard such customers. The regulator linked these failings to rapid customer growth outpacing controls.
