AI in Insurance Fraud

Is AI’s Disruptive Power Set to Unleash a new Wave of Insurance Fraud?

AI is revolutionising how insurers do business. But will they soon be playing catch-up as professional and opportunistic fraudsters exploit the technology? To evaluate the emerging threat Avoira’s Head of CX Ian Taylor sat down with former esure and Go Compare executive Adrian Webb and former insurer Head of Enterprise Fraud, Matt Gilham.

This may seem counterintuitive in a digital first age, but voice is set for an insurance industry comeback.

I’m not, of course, suggesting that insurers and brokers will abandon their online channels. Far from it. The efficiencies and service benefits delivered are embedded,
the customer journey refined through digital tools to optimise the experience and promote sales, retention and satisfaction.

But digital has its limitations and, with the democratisation of AI, these mean very real threats are on the horizon.

How did I come to this conclusion? Avoira is active in the insurance industry through provision of both multi-channel CX analytics and unified communications and call centre technologies.

We have a good grasp of what’s going on in the insurance world and, better still, have established relationships with highly experienced executives who offer invaluable insights into not just the industry in the here and now, but what lies ahead. To echo the old Allied Dunbar ad, ‘there may be trouble ahead’.

Keen to map out future threats and opportunities – so Avoira as a technology company can respond accordingly – I had a roundtable with Adrian Webb and Matt Gilham.
These guys know their stuff.

Adrian spent 20 years in the upper echelons of insurer and aggregator worlds, having been former Chief Marketing Officer and a non-exec at Go Compare parent GoCo Group, not forgetting his many years at esure, Direct Line and Virgin Money.

He introduced voice stress analysis to the UK insurance industry. Meanwhile, Matt, a Certified Fraud Examiner, was Head of Enterprise Fraud & Financial Crime at esure and sat on the ABI’s General Insurance Fraud Committee, with involvement in oversight of the Insurance Fraud and the Insurance Fraud Enforcement Department.
Our conversation was equal parts illuminating and, if you’re an insurer, exciting but in equal measure alarming.

So, the return of voice?

“There’s a battle of technologies,” Adrian explained, noting that the technological cycle means that insurers’ increasing reliance on AI presents a future threat. Given we’re counting in months rather than years, the future is, though, pretty much now.

The attractions of generative AI for a business are obvious. Complex tasks can be automated, data interrogated and structured to assess risk – underwriting and fraud – AI agents deployed to front end customer interactions. Many insurers also see benefits in claims processing.

But – and it’s a big but – AI deployments also bring risks, including misidentification of potential fraud and overstepping compliance and regulatory marks through the limitations of, and over-reliance on, algorithms.

Let’s return to that later.

In the rush to secure the benefits AI offers, have insurers ignored or, at the very least underestimated, the risk it presents?

Generative AI is, of course, far from exclusive to the corporate world. Everyone, including the bad guys, has some kind of access.

Adrian sums it up succinctly: “Publicly accessible generative AI is able to produce keyboard inputs, photographs, fill forms and so on. A script can do it. So, suddenly there’s an asymmetry between the power of the fraudster and the power of the insurer to detect their fraud.”

This, Matt pointed out, could see insurers re-evaluating the power of voice in the transactional mix as the voice channel builds greater rapport with insurer customers and supports the faster capture of greater data, supporting both enhanced CX and data decisioning. “We’ve had this headlong rush into online data capture but now, with modern AI technology unleashed, are insurers going to once again utilise voice?”

As Matt noted, voice by its very nature also allows for the capture of greater and richer data. “Voice is 140 to 150 words per minute whilst typing is maybe 40 to 50. Voice delivers more data and depth to assist decisioning.”

And this is where voice will both enable the journey of the majority, genuine customer and provide opportunity for scale identification of fraud risk.

There’s a strong logic for this. When we speak, we reveal much more about ourselves than if we are interacting online. Telling the truth is much easier than lying, for which there are a number of linguistic tells.

Adrian expanded on this “The human brain can only process language at a certain speed. When you’re lying you’ve got four processes going on as opposed to one when you’re telling the truth. There’s no cognitive load in truth-telling.

“People telling the truth never stress the truth of what they’re saying. When lying, all sorts of linguistic things come in to play as people try and buy time to think.”

These include a loss of information density– the use of words that fill rather than enlighten – and disassociation. “People unconsciously use a passive voice that distances them from their deception. The phenomenon of verum focus also comes in. That’s when people swear on their lives and their children’s lives that they’re telling the truth. People genuinely telling the truth rarely try to stress it.”

Identifying such verbal tics is crucial in combatting claims fraud which depends on fabrication, creating an event or a version of an event that did not happen.

Claims is where identification of truth matters most to insurers. Claims is where the money goes.

Matt referenced the 10-80-10 rule and how the availability of AI could change its dynamics. The rule states that there are 10% of people who will always try to defraud you, 10% who will never and 80% who, if the conditions are right, might try.

Anyone can now use AI to drop in an expensive item of jewellery, a watch, a camera or whatever into a photo they submit to evidence a claim. How many more might be tempted into the opportunistic fold to bump a claim? Does 80% become 85% or more?

What’s worrying for insurers is that AI generated fraud may already have negatively impacted claims ratios as Adrian explained. “AI is a discordant technology, a technology that  loves faster than regulation, laws or societal norms.

He pointed to historical precedents, such as the invention of the Gutenberg press which, through faster, mass printing, undercut clerical censorship of texts. Previously the spread of ideas deemed dangerous could be achieved simply by destroying the author’s notes.

“In AI’s case, the massive speed of development means that during its most critical development phase it is effectively unregulated because the regulations in place never foresaw
LLMs and the like.”

That means there’s a lag between the actual impact of a discordant technology and discovery of that impact. “Claims statistics are historic, usually a year, two years behind. That’s a window of opportunity for a fraudster who can and will make use of key technologies. Because a discordant technology is at work, the problem and its scale will not emerge in
longitudinal data but only at the end of a reporting period.”

If insurers thought they already had an issue with latent fraud – the losses they don’t detect – AI’s disruptive power is going to present quite a shock.

Aviva have identified, in a recent article, that a “growing number of claims have been supported by AI-generated images and manipulated documents, particularly in motor insurance. Fraudsters are using these tools to fabricate accident scenes and damage imagery to support false or exaggerated claims.”

Crucially the insurer notes that its fight back embraces not simply advanced analytics and its own AI tools, but human oversight.

So, AI is the problem but equally, in terms of digital interactions, not the solution. Potentially, voice is.

However, at present there’s no effective way of evaluating voice communications in order to flag potential frauds.

Indeed some of the technologies deployed to tackle this issue, could in fact be opening insurers to compliance, regulatory and even civil legal liabilities.

Their methodology is simplistic. “Some analytics offers are based on very limited questions and are equivalent to a basic polygraph,” Adrian advised, pointing out that relying on verbal cues of stress alone is no longer recommended as it can easily pull vulnerable customers into the suspicious camp.

Without a broader intelligence underpinning their analysis they could flag false positives. Their bluntness could negatively impact vulnerable customers withdraw genuine claims. “They make people know they’re being inspected which could make those who are vulnerable nervous to the point that they might not actually continue with a claim, even though
it’s genuine.

“Similarly, someone who speaks English as a second language may be interpreted as pushy because they’re a non-native speaker and translating directly to English.”

Both these scenarios would see red flags waving under the FCA’s Consumer Duty regime.

Matt expanded on this point: “If you’ve simultaneously got fraud risk indicators and a customer demonstrating vulnerability or a complaint, insurers need to be empowered to make the right decision, ideally in real time. The information and the linguistic tics Adrian mentioned are in the call, but a more powerful analytics solution is required to
extract them.”

Adrian further pointed out that the FCA’s recently reopened AI Input Zone (AIIZ) throws another regulatory spanner in the works with reference to both Consumer Duty and the Senior Managers & Certification Regime (SM&CR).

The AIIZ has been collating examples of good and bad AI practice to inform an evidence-based standard. Rather than looking at principles or policy statements, the FCAS called for specific examples spanning governance, resilience, oversight, assurance, deployment controls and consumer outcomes.

“The regulator sought evidence of actual outcomes when policy documents and board-approved framework have embedded AI governance. They want to know how such deployments are governed,tested and monitored.

“That has further implications for the accountable senior manager under SM&CR.”

He also foresees a time when the use of AI tools – such as large language models and automated scanning platforms – to review documents for compliance, falls foul of the regulator. “Relying on a third-party AI tool to identify a compliance issue, without documented human oversight and an accountable executive, will become untenable. The tool is not accountable.”

Clearly faster and more accurate identification of fraud, coupled with enhanced regulatory compliance, would offer significant financial – and potentially reputational – benefits.

Equally, a solution that could deliver this could likely be deployed to deliver wider operational benefits to enhance performance across sales and service channels.

Whilst the speech analytic tools which could calm the incoming fraud storm exist, they’ve yet to be modelled to handle the task. This, perhaps, is because the headwinds have not yet shown on insurers’ radars?

Those that recognise how sophisticated voice capture and analytics technology can be deployed to nullify the threat, will clearly benefit from early adopter advantage.

This begs the question; who will take the lead?