The Center for Data Innovation spoke with Niels Thoné, chief executive officer and founder of Sprout.ai, a company based in London that uses AI for claims automation and fraud detection. Thoné discussed how the combination of technologies such as natural language processing and image recognition can accelerate tedious insurance processes such as damage claims.
Eline Chivot: What is the problem that Sprout.ai set out to solve? How did your activities evolve from deploying activities in the insurance sector and motor insurance, to working on health claims and other lines of businesses?
Niels Thoné: Claims settlements in insurance are long and tedious for customers and insurers alike. Customers are unhappy with the service provided when settling claims, which is when they are most vulnerable and need help. Insurers have a lot of manual touch points, high operational costs, and a less than ideal reputation. The problem lies in the limited data capture when a claim is made, which is called first notice of loss. This initial problem pervades the whole workflow and slows everything down. We solve this problem by bringing in external data sources in real time to enrich the limited claims data and validate automatically if the claim is safe from the get-go—which allows insurers to settle claims immediately. Our combination of data enrichment and predictive analytics allows insurers to settle claims in record time, save money, and provide unprecedented customer service.
Until recently, our company was known as BlockClaim. We had a hybrid solution of a private blockchain for data gathering across legacy IT systems (which are everywhere in insurance) and AI for predictive analytics. Throughout our journey and by working with global tier one carriers, we quickly realized that procurement is very rigorous and time consuming, and they did not embrace a new technology such as blockchain—so implementation cycles become very lengthy. Therefore, we developed alternative solutions to pair with legacy systems as a plug in and focused solely on our world-class deep learning predictive analytics. Hence the name change to Sprout.ai.
We originally started in property and casualty (P&C) insurance but have always been opportunity-driven. So, when the chance came to work with a very reputable health insurer on their claims challenges, we leapt at the opportunity—knowing that we could provide a transformational solution with our data enrichment. The results were extremely positive, and our health claims solution is now one a highlight in their product suite. For now, we will continue to focus on health and P&C insurance. We get a lot of requests from other types of insurance but for the time being we are holding them off. We are keen to stay focused and continue working on what we do best.
Sprout.ai uses a mixture of AI and machine learning techniques, such as natural language processing, image recognition, and optical character recognition, but no longer uses blockchain. While theoretically it makes a lot of sense to leverage private blockchain technology for data gathering and scrubbing across legacy systems, in practice it is challenging to get this across to compliance and procurement departments in large financial institutions. Instead we now implement our plug-and-play solution via different methods, such as cloud application programming interface (API), virtual private cloud (via docker installation), and on-premise integration.
Chivot: How do you collect, use, and leverage data to enrich claim processes? Why is capturing these various technologies in a “hybrid way” particularly efficient, and more so than other approaches?
Thoné: Part of our secret sauce is our data engineering capacity and data enrichment network. We currently have a network of over 50 data categories, ranging from weather, financial information, business information, car values, international medication and treatment information. The data is gathered via two mediums. Fifty percent via API, where we work with global data providers specialized in one specific information stream (e.g., credit risk score). We integrate the other 50 percent ourselves by building crawlers to scrape databases and websites (all of it GDPR-compliant of course). We have a special legal team that validates every database before it gets aggregated. This approach is very powerful as it gives us access to data that otherwise wouldn’t be leveraged as it’s not available via straightforward APIs. It really brings unparalleled information to the claim at the earliest possible moment, which solves the entire bottleneck and streamlines the process both for insurers and their customers. Win-win!
Chivot: How is Sprout.ai changing claims processes to the benefit of companies and assessment experts, and concretely improving outcomes for customers?
Thoné: Our mission is to enable any insurance company to settle claims in less than 24 hours (instead of the 25 day average!). We improve customers’ outcomes by ensuring less hassle, a faster and less biased service, and more accurate payouts. We improve insurers’ outcomes by leveraging data enrichment and proprietary deep learning algorithms. Our process is first to automatically check for coverage by reading the policy (via NLP) as soon as the claim comes in. Second, we create a fraud referral stream at the start of the claim, informed by the relevant external data (deep learning on macro and micro data). This allows insurers to catch more fraud, save money, and free up precious claim handler time to deal with the actual valid claims and customers because it’s always the honest customers that suffer from the actions of fraudsters. Third, we provide cognitive claims recommendations. Using deep learning, we can predict that the claim is a valid one, days before all the claim information is even available (a first in the industry)—therefore speeding up payments to customers without making a trade-off in accuracy.
In a nutshell: Using Sprout.ai allows insurers to fast-track the majority of claims, immediately spot the complex claims that require extra assistance, and double down on fraudulent claims.
Chivot: How has the insurance industry been looking at emerging technologies, and do you see an evolution?
Thoné: We believe that the insurance industry is starting to embrace emerging technologies. There’s a growing awareness that to survive, they need to adopt new technologies otherwise they will be disrupted—e.g., by using Lemonade for homeowners insurance.
In terms of evolution, eight years ago the focus was on digital claims experience (chatbots, website forms). Five years ago, it was about increasing fraud detection via rules-based systems or traditional machine learning. Three years ago, it was about the underwriting side of things. And since the last couple of years, there has been growing awareness that the most efficiency gains and customer satisfaction increase lies in the claims workflow, which explains this increased focus on claims innovation.
Chivot: Fraudulent claims remain often associated with insurance. Can you give an example of how your solutions can go beyond fraud detection and prevention, serving and expanding to other sectors such as healthcare?
Thoné: Fraud has been a “sexy” topic in insurance for the past five years—and it’s absolutely necessary to do if you want to pay claims quickly (you cannot settle a claim fast without first checking for fraud, it’s one of the “hard values”). Although, this only represents five percent of the total amount of claims. Therefore, this means that 95 percent of claims are overlooked in terms of innovation and improvement. That is really where Sprout.ai operates.
We enable insurers to expedite the bulk of safe and relatively simple claims and only focus on the claims where actual claim handler expertise is needed. Insurance customers should not have to suffer for the insurers’ inability to triage their claims accordingly.
What excites me the most in what we’re doing is the application of deep learning (versus traditional machine learning) to these datasets via our proprietary Super Resolution software. For one, it provides much stronger predictive capacity (95 percent or, as we call it, a “human+” level of performance versus 70 percent hit rate for traditional machine learning) but when applied to the cognitive claims use case it will enable us to predict what will happen to claims even when the information hasn’t come in or hasn’t even been created. It really starts going into the realm of predicting the future (in a claims context). Sort of a Minority Report for claims—without the negative connotations, and with a human in the loop to verify. This is the holy grail for claims, and I’m extremely excited that we are cracking it, with the help of our insurance partners globally. It will redefine the insurance and claims space and solve the entire bottleneck that has been there for nigh on centuries—at least, in personal lines insurance, to start with.