Nigel Cannings, CTO at Intelligent Voice will be speaking on this topic at NVIDIA GTC, taking place online April 12 – 16, 2021. Register free and join us for his session ‘Detecting Deception and Tackling Insurance Fraud using GPU-powered Conversational AI’ plus so much more. GTC offers free access to 1400+ sessions covering the latest breakthroughs in AI, data center, accelerated computing, healthcare, intelligent networking, and more. Click here to learn more.
Ongoing investment suggests that the perception amongst insurance professionals is yes. According to a 2021 survey, approximately 31 percent of insurance CIOs reported that they had already deployed AI and an additional 23 percent will deploy in the next 12 months. Focussing on Fraud, a recent survey by the Coalition Against Insurance Fraud showed that 21 percent of respondents are planning on investing in AI for fraud detection over the next two years. Yet against this backdrop of adoption, we see that most insurance companies process only 10-15% of the data they hold. The majority of the data currently analysed by insurers is structured, highly organized and formatted in a way that makes it easily searchable, housed in relational databases. Circa 10 percent of data is structured so whilst we can see that Insurers are making a dent on unstructured data there remains valuable insights left hidden in vast volumes of unstructured data. Effective utilisation of this unstructured data can help insurers identify potential fraud but also shorten the customer journey by validating information more efficiently. One form of unstructured data that remains significantly underutilised is voice.
AI has an enormous remit, but it is Conversational AI, natural language processing (NLP) and automatic speech recognition (ASR) that have the potential to change the game across fraud and the contact center. Call centers are still seen as the soft target for fraud, with agents being faced with a constant barrage of sophisticated social engineering techniques, designed to manipulate and bypass scrutiny.
Contact center agents handle scores of calls a day, each one expected by business and customer alike to process queries as quickly as possible and today agents are working remotely with the added pressures that that brings. These factors combined mean the call centre presents a vulnerability that even inexperienced fraudsters can exploit. Unless a caller has been previously flagged for suspicious activity or is particularly unskilled at the social engineering requisite to scamming, their fraudulent first notification of loss (FNOL) may be processed like any other. This vulnerability is further exposed during times such as these where economic downturn correlates with an uptick in organised and opportunistic fraud. An example of this uptick is the spike in motor vehicle theft since the start of the pandemic. The NYPD reported a 57 percent increase in thefts of cars, motorcycles and mopeds between March 16th and April 14th, 2020. Yes, the requirement to shelter means that vehicles were left more vulnerable but many of those owners were also financially impacted, loans soon turned upside down and negative equity took hold because of unemployment, a terrible situation, and a challenge for insurers to unpick with care.
Contact center agents must strike the balance between serving the needs of their genuine customers and staying alert to fraudulent behaviour. With estimates putting fraudulent claim volume at around 10 percent then one might expect that an agent’s attention is best focussed on the 90 percent of genuine claims. Where this balance is not well understood or implemented it is common for well-intentioned anti-fraud measures to put obstacles in the way of genuine customers. AI has the power to prevent fraud at its source without impacting genuine customers.
Voice authentication has been used for some time at the front end of contact centers in the form of Interactive voice response to route calls more effectively. With the addition of sentiment analysis to determine emotion and tone automatically and accurately in the customer’s voice, this triage can be further enhanced, calls are routed faster to the appropriate team and less time switching between teams means more time helping and pleasing customers.
Insurance fraud across all lines of business in the US is estimated to be $100B per year and with the pandemic creating hard times for all, current data indicates that 2020’s figures could be as much as 21 percent higher. Reducing these figures is in the interest of everyone and artificial intelligence (AI) potentially holds at least part of the answer.
Did you know that fraudsters tend to share a common set of speech characteristics? Negation, hedging, latency, and emotion are all frequent features of the fraudster’s interaction. They will pause more often, as they think through their answers, trying to avoid inconsistencies. They will react more emotionally if they believe that they are being caught out, and they will negate their sentences. These things sound relatively easy to detect, but if you’re busy, just trying to do your job, focussing on genuine customer needs, they can easily be missed, even by highly trained staff. With conversational AI, semantic parsing allows the meaning to be extracted from each and every utterance. This isn’t a one-sided event either, by understanding both the questions posed by an agent and the responses received, false positives are significantly reduced as context is understood. Let us think about an example from the motor vehicle insurance claim and the exhibiting of hedging language in a claimant’s verbal account, hedging is a form of non-committal language, an individual is withholding. One particularly tricky hedging term is ‘around about’. If a claimant is asked how much their, now missing (reported as stolen) gold chain cost and they respond, “it was around about $3000, but I’ve got the receipt to prove it” then we would be concerned. How much was it exactly? If you have the receipt then why do you not have this figure to hand? Here we flag ‘around about’ as a genuine event of Hedging. Now let’s imagine the claim is for a motor vehicle collision and the claimant was approaching ‘a round about’ when they were hit from behind. Without context we would have flagged a risk that doesn’t exist. Only a truly intelligent voice analytics solution can exploit value, anything else is a risk. Intelligent Voice have spent the last 5 years developing their LexiQal Conversational AI solution, utilising Nvidia’s cutting edge Triton Inference framework to surface only those true indicators of risk.
Another of the call centre’s unavoidable Achilles heel is the number of people employed. Callers will rarely talk to the same person twice and this is very much in the fraudster’s favour. This issue is most prevalent across banking, the UK Finance IVR’s report stated that a fraudster will make 26 calls to the contact centre during the execution of a given fraud. When the fraudster is lucky enough to speak to a different person with each call the chance of any inconsistencies being picked up is dramatically reduced. Intelligent Voice witnessed the power of pattern recognition across a piece of work for a UK Insurer where analysis of the insurers fraud calls against their believed genuine calls showed a marked increase in the use of the phonetic alphabet by fraudsters reporting cash for crash claims. With ASR and NLP, every call is monitored, every irregularity noted, and every behavioural pattern identified.
Medicare fraud losses are estimated at $60B per year in the US, the biggest single source of fraud. In addition to those losses the insurance market devotes an estimated $3 billion in work hours to collating and summarizing medical records in support of claims. Intelligent Voice has developed the ability to automatically generate ICD10 and CPT codes by analysing the voice channel during Telehealth consultations. Not only does this offer efficiency gains but simultaneously shines a light on any overbilling through enhanced scrutiny and the ability to audit.
Ninety-seven percent of US Insurers contribute data relating to fraudulent claims that is available to all – Voice Biometrics, voice recognition or voiceprint is a form of identification unique to an individual. This is currently one of the most common forms of AI and has long been used to validate customers into a contact center or populate watch lists within an organisation. The ability for insurers to begin collaborating over this rich data source could be transformational but the appropriate data security procedures must be in place.
Employed together, Conversational AI, NLP and ASR have the potential to stop insurance fraud before it has even started. Of course, they cannot stand alone. They cannot be onboarded and abandoned. Every ASR system needs calibrating for different use cases. And even then, the overall engine cannot make decisions (yet). But together, these systems can provide decision-makers with the real-time, actionable, analytics-led insights that are simply not available from any other source right now. And, excitingly, these technologies have further applications.
Sentiment analysis is now a fairly staple feature of customer analytics but with the evolution of semantic understanding comes the ability to alert agents to emotional features in real time that are deserving of consideration. When responded to, these alerts help steer an interaction in a positive direction for both the customer and the organization.
Understanding customer needs better offers the insurer the chance to not only delight their customers through enhanced customer experience but in parallel ensure that needs are identified and met appropriately every time. Intelligent Voice were recently approached by an innovative UK health insurer looking to identify vulnerability at underwriting. Does the insured understand what they need from their policy? Have the appropriate questions been asked by the agent to ensure that effective cover is in place? Not only does this safeguarding ensure that should a claim be forthcoming then the insured is properly covered but for the insurer it avoids policy leakage.
Understanding and automatically recording the reason for every call allows organisations to identify failings sooner and structure resourcing to meet changes in demand across their contact center. An early famous example of the value to be gleaned here was the work of Expedia where they identified $100M in costs associated with calls from 20M customers a year, all looking to obtain a copy of their itinerary. A seemingly trivial issue with a huge cost.
Working with a Police Force in the UK, Intelligent Voice have successfully addressed the issue of unlogged demand through the analysis of each and every call into the emergency and non-emergency contact centre to surface real time trends across the organisation.
Despite enormous recent leaps, AI is still in its early days. In most instances, it cannot replace the human element, and in most cases nor would we want it to, but with conversational analytics, AI has the potential to provide vital information that could not otherwise be accessed. This has potential in a wide variety of field, but in insurance it could prevent billions of dollars of losses through fraud, waste and missed opportunity, that has to be worth investigating.
