Fraud Isn’t Artificial – So Why Would Our Judgement Be?

I recently published a short series of LinkedIn posts talking through some of the fraud detection tricks of the trade I’ve picked up over years spent investigating different forms of fraud. None were new. Most would be familiar to experienced counter-fraud practitioners. Some would probably surface if you asked generally AI-for-guidance. But you’d have to prompt carefully to get beyond generic answers and into the practical details: tests for identifying fraud in procurement, tips for spotting the deviation of payment controls, ways of finding loopholes in employee vetting, tricks for identifying patterns in data network analytics, or tried-and-tested solutions to tackle fraud on the online insurance channel.

The purpose of the posts wasn’t to promote these individual tips. It was intended to underline something more fundamental: that effectiveness in tackling fraud still comes from marrying technical capability with learned, human experience. Technology can surface signals, connect dots in data, and speed up our ability to triage large amounts of information. In my experience, though, it’s still human judgement that defines what is pursued, how approaches are deployed, and how decisions are ultimately reached in an investigation.

In practice, the balance between technology and people has been an insurance fraud issue first and foremost. Fraud is about understanding and managing behaviour: motivation, pressure, opportunity, and rationalisation. It can be highly organised and opportunistic, impulsive and complex, or a series of small, seemingly harmless actions that convert victims of genuine vulnerability into victims of fraud. When we get that balance wrong, the consequences go beyond missed fraud. Customer outcomes suffer, trust erodes, and confidence in decision-making is weakened.

I still remember my first investigative interviewing training course. I learned how to prepare for a suspect interview, how to subtly build pressure during a conversation, and developed observational skills on what signs of stress or emotional leakage to watch for. Most memorable, though, was being introduced to the concept of the empathetic “thug”: the moment in an interview where emotional connection is established and a suspected fraudster becomes more receptive to telling the full story. I’m so relieved to have been taught that skill. Even when dealing with fraudsters, emotional intelligence and empathy remain core investigative skills, and a clear expression of human judgement at work.

In insurance fraud, empathy matters beyond the interview room. Protecting the good and stopping the bad are flip sides of the same investigative coin. It’s necessary for us to combine the most human of skills with the technology that enables scale, speed, and automation. We also need to ensure vigilance keeps pace with the malicious use of AI by fraudsters themselves. And we need to preserve human traits, particularly empathy, if we’re to prevent fraud while still delivering good customer experiences.

Philip K. Dick’s dystopian science-fiction novel Do Androids Dream of Electric Sheep? (later adapted into the film Blade Runner) captures this inherent tension. Bounty hunter Deckard tracks down escaped androids — Replicants — using a test based on emotional response. Their failure to remain undetected isn’t about intelligence or capability. It’s their inability to feel empathy. The dividing line between human and machine is emotional understanding.

Our approach to innovation in counter fraud is hardly new. Fraud teams have consistently been early adopters of technology: from early data sharing and matching, through device recognition and risk assessment, to data enrichment, machine learning, and predictive analytics. More recently, we’ve seen the rise of voice, document, and image-risk capabilities designed to detect manipulation. In many respects, counter fraud has been at the forefront of applied AI in insurance — reducing loss while enabling more proportionate and less intrusive customer journeys.

We’re now seeing more mature adoption of agentic and generative AI: automated red-flag collection, faster risk assessment, clearer investigative outputs, and more targeted prompts back to human investigators. Much of this remains exciting and powerful. But some of it still raises as many questions as it answers. It’s easy to worry about technology replacing people. But the greater risk may be quieter: allowing tools to set the agenda, and losing the craft of human curiosity and investigation.

As insurance work becomes more technologically centric, the challenge is balancing deep technical capability with the industry’s long-standing strengths in human judgement and relationship-based work. Not every investigation needs to become a data scientist, but teams do need enough understanding to challenge what models show, where they don’t, and what they don’t see. Equally, we must protect the human parts of the job: customer conversations, business relationships, contextual understanding, and the application of empathy.

For me, the lesson is clear. In embedding more “human in the loop” AI, we should be asking: are these tools helping us amplify our traditional strengths, or are they diluting them? Will we dream of electric sheep? Like Deckard, this remains a perhaps.