Every alert that lands with a fraud or SIU team arrives with dozens of fields, flags and linked records, but no story. Before an investigator can judge the case, they have to read across scattered data points, chase further evidence and piece the narrative together by hand.
The decision itself is often quick once the full picture is in front of them. Getting to that picture is the real bottleneck.
That is the gap we explored in our recent webinar with Insurance Times, Fraud or Claim? Making Agentic AI Work for the Investigator. Our own George Robbins, Head of UK, Ireland and Nordics, was joined by Nick Rawlings, Director of Transformation at Covéa Insurance. The aim was to look at what agentic AI actually means for fraud investigation. It is not a buzzword, and it is not a replacement for human judgement. It is a way to shorten the distance between an alert landing and a decision being made.
For years, investigators have worked across multiple systems, datasets and workflows that were never designed to work together. George calls it "swivel chairing", and it consumes time before any real analysis begins.
Agentic AI changes this. A plain-language, domain-aware summary of an alert, paired with an agent that gathers and cross-references supporting evidence on its own, turns minutes of reading into seconds of orientation. Investigators spend their time on judgement rather than data-hunting.
It also opens up new territory. Commercial and health insurance, with their large volumes of unstructured documents, were long out of reach for automation. That barrier is coming down.
Every SIU team has the same nagging suspicion: we think there's a blind spot here. Traditionally, turning that suspicion into a working detection scenario meant waiting for a data scientist or raising a vendor request.
Agentic AI shortens that path. An agent that helps teams explore their own claims data and tune detection scenarios directly lets the people closest to the problem close the gap themselves.
This is where the throughput point matters. As George put it, fraud is a combination of how much you detect, how accurately you detect it and how fast you can process it. Change the time an investigator needs per case, or the fidelity with which they find what matters, and you change the output of the whole system: more fraud prevented, not just less effort spent.
Nick Rawlings added the customer perspective. At Covéa, the business case rests on cost avoidance, which far outweighs efficiency savings. Efficiency is welcome, but it is not the principal driver of value.
This is the question every insurer, regulator and consumer is asking: does "agentic" mean losing human control?
Our answer is no. We believe agents should be case-preparers, not deciders, with an accountable human and a clear audit trail at the centre of every decision. As Nick put it, you can delegate tasks to AI, but you can never delegate accountability.
Where the line sits in practice is still being worked out across the industry. If an agent checks in at every step, it creates friction. If it runs the whole process unchecked, the risks multiply. In a regulated industry built on looking after customers in their time of need, someone has to understand the technology's decisions, challenge them and spot when it moves outside the right guardrails.
Watch the full webinar: Fraud or Claim? Making Agentic AI Work for the Investigator
Want to see agentic AI in action for fraud investigation? Talk to our team