Financial crime investigators are employed to investigate. In practice, much of their working day goes on something else: trawling through systems, gathering data, compiling evidence, writing up findings and clearing large numbers of alerts that turn out to be harmless.
With alert volumes climbing, deadlines tightening and criminals becoming more sophisticated, already stretched teams in both Transaction Monitoring (TM) and Fraud are under growing pressure.
WorkFusion, a UiPath company, argues that AI can change this operating model, but not by handing everything over to Large Language Models (LLMs). Its AI Agent, Isaac, is built on what the firm calls a “Goldilocks” principle. It combines AI, deterministic automation and human judgement in the right proportions and orchestrates them across the investigation.
The aim is to shift where investigators spend their time, not to remove them from the process. After a TM alert fires, an analyst may need to pull customer records, review transactions, research counterparties, check case history, compare expected and actual behaviour, map relationships and document it all. Fraud cases are similar. An account takeover (ATO) alert can require evidence from fraud engines, authentication services, card networks, digital identity platforms, behavioural intelligence tools and third-party sources.
Isaac takes on much of this “hunter-gatherer” work. In Fraud Alert Review, it connects to existing fraud engines and data sources, pulls together relevant context, analyses transactions, drafts a narrative report and passes the finished work to investigators for human-in-the-loop review.
The firm stresses that not every task needs an LLM. Language models are well suited to unstructured information, such as synthesising historic SAR narratives, reconciling past dispositions against a current alert or interpreting messy counterparty data. Transaction arithmetic is a different matter.
Aggregation windows, velocity calculations, threshold proximity and peer comparisons should be calculated deterministically and fed to downstream AI as established facts. In short, LLMs handle language and context, rules handle consistency, and people handle judgement and accountability.
The result is decision-ready work. Investigators become reviewers and decision-makers rather than data collectors and report writers. This matters when regulatory timelines collide with unpredictable workloads, holidays and complex cases, producing the familiar “day-30 scramble.” Because Isaac can start working alerts as soon as they arrive, it adds capacity and enables earlier escalation of potentially suspicious activity.
The early results are notable. One financial institution using Isaac for first-party fraud and ATO reviews cut manual research time by more than 70%, and its analysts handled two to three times their previous case volume. A regional bank applying it to structuring alerts reported around 10% of alerts auto-grouped, roughly 60% auto-closed, and one to three hours saved on cases that still needed investigation.
The real question, WorkFusion suggests, is not whether AI can make the decision, but how much work can be intelligently automated before a person has to make it.
Read the full Workfusion post here.
Copyright © 2026 FinTech Global
