Voice channels: the hidden front door to fraud

Criminals are increasingly targeting telephony channels as a gateway to wider fraud activity. But how can insurers use this intelligence to stay one step ahead? Philip Harding reports.
 

Voice channels, synthetic identities and AI technology are creating new fault lines in insurers’ existing fraud controls.

A recent Insurance Post webinar, in association with Smartnumbers, discussed the growing problem of customer data captured across apps, telephony and chat stitched together via accessible generative tools, making previously specialist techniques available to a wider range of attackers.

Insurers face practical questions about where to apply checks, how to govern fast‑moving tooling and how to make sure frontline teams are protected as well as empowered.

The evolving fraud landscape

The discussion opened with an assessment of the size and scope of the fraud challenge. “Fraud is the most prevalent crime type in the UK… accounting for more than 40% of all crimes reported,” said Ben Fletcher, director of fraud and financial crime at Allianz. What insurers call fraud is no longer limited to “spilling paint on carpets and dodgy theft” but appears across product lines and throughout the policy life cycle, he explained.

Fletcher described insurance fraud as “a systemic issue”. He distinguished opportunistic behaviour from organised crime, noting both co‑exist in the market and require different responses: “At one end of the spectrum you’ve got people that are seeing a quick opportunity to add extra money to a claim. They may not think it's that harmful but, if everybody does the same, it has a pretty big net impact. At the opposite end we are seeing organised criminality – people using insurance to launder money and to help further criminal enterprise.”

James Baverstock, lead insurance and fraud specialist at Cuvva, described how those broader trends meet the particular risk profile of app‑first, short‑term insurance. Cuvva’s focus on rapid onboarding and in‑device signals has forced the business to experiment: “We can look at behavioural biometrics… we can look at how customers interact,” he explained.

That same emphasis on device and session data, he said, has surfaced new types of synthetic and hybrid identity fraud: “We’re starting to see AI generated fraud… synthetic or hybrid synthetic IDs, synthetic images.”

Why voice has moved up the agenda

In the US, they’ve seen a 475% increase in synthetic voice attacks in the last year.

James Baverstock, Cuvva

Smartnumbers’s chief product and success officer, Tim Burton, pointed to the often-underestimated role of telephony channels and voice as part of the fraudster’s toolkit. “Telephony has definitely become a shadow area,” he said, and he described an important structural change that matters to fraud teams: conversational channels are being unified.

“The conversations that are held… are being converged and consolidated into one place,” he warned. In practice, that converged record can be mined and repurposed; a voice snippet, a chat transcript and a device fingerprint can all be stitched into a profile used elsewhere.

Baverstock described how he’d conducted his own experiments in this area. Using publicly available services, he was able to “…clone my voice to the point that my own family couldn’t tell the difference.”

He also pointed to trends in other jurisdictions: “In the US, they’ve seen a 475% increase in synthetic voice attacks in the last year.” This has implications for where and how a firm should apply checks. “If it’s coming through a phone, what’s the chances of voice analytics and biometrics picking that up?” he asked.

A multi-faceted problem

In seeking solutions, Fletcher pointed out analysis of data or other signals in isolation is not enough. “There’s no silver bullet. The idea of having a single tool to be able to stop fraud is not relevant,” he said, arguing instead for an environment made up of complementary controls.

Operationally, voice features should sit with signalling and network metadata, with behavioural markers gathered in the app, with document and image analysis and with cross‑sector intelligence. “You have to use it in collaboration with other things,” Fletcher said.

Burton described this task as particularly challenging in the context of live sessions: “It’s more of an arms race for how you can use and compare data that you already have in your estate, but also data that’s happening live in a session and the way in which you draw conclusions between those different pieces of information.”

He also highlighted a concrete and worrying example of scale – a criminal platform he referred to as “fraud GPT” that had been discovered running large‑scale contact‑centre operations with thousands of short‑term ‘voice over internet protocol’ (VoIP) numbers and multi‑language call variants.

“This contact centre had 25,000 different variations of global languages. Calls were spun up with VoIP – phone numbers, so they’re short, short shelf life,” he said. The discovery, he implied, showed attackers testing defences in bulk rather than mounting single, bespoke efforts.

There’s no silver bullet. The idea of having a single tool to be able to stop fraud is not relevant.

Ben Fletcher, Allianz

Human judgement and vulnerability

With voice channels increasingly targeted by fraudsters, discussion turned to the evolving responsibility and pressures on frontline staff and where AI can play a role.

Baverstock’s view was that insurers’ approach should be “human‑led AI, AI‑enabled.” But the operational reality is that many insurer interactions are low‑frequency and often occur at a stressful moment for the customer – so staff cannot be expected to have extensive prior relationship data to lean on.

He also noted that contact‑centre hiring profiles – young, empathetic people whose priority is helping others – can increase the risk of social engineering: “That also makes them susceptible to passing out information that we might not want them to.”

Fletcher highlighted some of the non‑transactional harm that phone calls can enable. He described pathways in which stolen direct‑debit details surfaced in contact‑centre interactions are later used as authentication material in other sectors, enabling number porting and account takeover.

“All of that is possible in the voice channel,” he said, highlighting a conundrum whereby firms need operating models that allow staff to serve customers while not exposing sensitive data.

Platform governance and vendor assessments

As use cases for AI and telephony tooling multiply, governance and oversight will need to evolve alongside. Fletcher spoke about the need to move from ad‑hoc approvals to platform approaches that can scale, and he emphasised the importance of contractual and technical vendor safeguards. “We are already looking at platform‑based governance,” he said, noting that many firms are assembling central repositories of approved patterns and pre‑allowed use cases.

Burton noted a related commercial trade‑off: data sharing is no longer just an ethical or legal question but a set of operational costs and benefits that boards now interrogate. “It’s about understanding the appropriate value it brings to your business and the trade‑off of signing up to more and more data‑sharing agreements,” he said.

Cross‑sector intelligence

Cross‑sector sharing of fraud intelligence was seen as a strategic imperative. Fraud rarely respects industry boundaries: stolen direct‑debit or authentication data can travel from retail to telecoms to banking to insurance, creating multi‑step chains that are invisible to any single sector.

Noting that the UK had made more progress than other jurisdictions, Fletcher was pragmatic about the legal and commercial barriers that shape what firms are willing to share.

“There’s always going to be a friction point in data sharing,” he said. “Firms are terrified of competition law and therefore there are limitations.” He framed the practical dilemma as a question of how sharing genuinely adds value versus where it exposes intellectual property or competitive risk.

AI is really exciting... If you stick to old ways of working, it’s going to become increasingly stressful and painful.

Tim Burton, Smartnumbers

Fletcher was nevertheless upbeat about the prospect of technical advances that create new, less sensitive signals for detection. He noted that richer telemetry “…gives us more opportunities to test and validate – so you can test the metadata in an image. You can test whether a voice is genuine or not,” and suggested those developments could change attitudes over time: “There’s an opportunity to shift minds”.

Baverstock emphasised that vendors and new entrants must be part of the solution, but smaller firms often need to be “considerate of where we spend our money,” focusing on in‑house experimentation with device and behavioural signals before layering in external feeds.

Reasons to be cheerful

Fast-evolving fraud tactics and methodologies will continue to test insurers, but are they winning the race?

Baverstock noted how insurers are adopting a more proactive approach. He said insurers were often “very reactive to fraud” in the past, responding to typologies only once they had appeared, but that the current cycle feels different because defenders and attackers are “in more of a high, fast‑paced race where we’re both learning at the same time.”

Fletcher highlighted two operational risks to watch: “We need to make sure we’re not complacent. And there is a real risk that some of the inherent knowledge that sits within businesses gets lost, gets put into a machine. It is a skilled area,” he said, urging continued investment in people and succession as models and automation are adopted.

Burton said he felt “very optimistic,” about the future in tackling fraud, and that the pace of operational experimentation with AI made the moment energising: “AI is really exciting… it’s fun to play with, and it’s enjoyable to see your colleagues find new ways of identifying bad actors.

“If you stick to old ways of working, I think it’s going to become increasingly stressful and painful.”

Sponsored content