AI: The Basics
While there is no uniform definition of artificial intelligence, the term generally refers to a broad field of computer science that deals with using advanced technology to simulate human intelligence and activities.
AI is a family of technology that includes a number of unique “members.” Let’s meet some of them:
- Natural language processing. Natural language processing is a “computer science field dedicated to enabling communication between humans and computers using human languages.” Computers are able “to break down written or spoken language into concepts and entities” and thus are able “to build relationships in order to analyze language like any other data type.”
- Machine learning. Machine learning is the approach of realizing AI by learning from, and making data-driven predictions based on, data and learning experiences. It is “the analysis of large data sets to find patterns; feedback and repeated iterations allow AI to learn, improve, and predict similar patterns in new data.”
- Robotic process automation. Robotic process automation encompasses “computerized workflows and applications that [perform] the mundane, repetitive—and often boring—aspects of various business processes” (such as data entry, data conversion, and forms population), reducing the amount of attention needed from people.
- Artificial neural networks. Artificial neural networks are “AI systems that are patterned after the structure of the human brain, with millions of interconnected, discrete nodes that . . . analyze external stimuli to arrive at an answer.” These “networks are learning systems that can be trained to improve over time.”
- Image recognition and computer vision. Image recognition and computer vision involve “[l]everaging machine learning, artificial neural networks, and other AI technologies to categorize and identify static or video images along with their . . . characteristics.”
- Chatbots. A chatbot is “[a] type of bot that can conduct a dialogue with a human user through written or spoken language,” such as an online customer service representative or a virtual agent.
AI Uses in Insurance: Real-World Examples
“AI is helping to solve real problems, improve the customer experience, automate workflows, and supercharge the analysis of massive amounts of data.” The availability of AI is changing how insurance companies operate and deliver their product and assess risk.
Insurers state that AI usage (in all of its iterations) falls into three main categories: (1) customer experience, (2) process improvement, and (3) process innovation. AI is modifying the customer experience to provide a faster and more personalized service by using “chatbots, customized pricing and coverage, and analytics to anticipate customer needs and drive [solutions].” Startup Lemonade, Geico, and Allstate already are using chatbots. “Process improvement” means streamlining external and internal processes, like claims management and call center operations, to increase efficiency. Lastly, “process innovation” describes the “[l]everaging of the Internet of Things [IoT] to develop usage-based products for customers. Examples include telematics and connected home and self.”
Underwriting. In the underwriting world, use of IoT devices has yielded an explosion of valuable data that can be utilized to regulate and simplify the process of determining insurance coverage and premiums. Insurers can automate the underwriting process, doing away with timely applications and tedious inspections. Indeed, McKinsey, a worldwide management-consulting firm, predicts that manual underwriting will “cease[] to exist” by 2030 for most property and casualty insurers. McKinsey also predicts that the underwriting process will become largely automated through use of machines and deep learning technology, reducing the process to seconds. It is anticipated that this change will empower consumers, who will soon be able to see (in real time) how their behavior will “influence coverage, insurability, and pricing.”
Consider examples of how the policy-purchasing world already has changed. Facebook Messenger’s Insurify, with the assistance of its 24/7 virtual agent “Evia,” “instantly verifies [consumer] data” using the potential insured’s online profile to “provide coverage recommendations” and quote comparisons in real time. Lemonade is providing renters and homeowners insurance to consumers by using chatbots and AI to create paperless policies “via desktop and mobile apps,” without use of human underwriters.
In addition, Flyreel
guides [users] through property inspections using the camera on their mobile device and powerful . . . computer vision. As users pan their mobile phone across a property’s interior and exterior, Flyreel’s . . . technology automatically documents property features, contents and policy-worthy details. For carriers, this enables personalized self-inspections at scale, real-time data intelligence, and comprehensive baseline records of properties to inform decisions across the value chain.
There is also Cape Analytics, a company that makes use of deep learning and data science, such as geospatial imagery, to provide property risk and value information to insurers. This system replaces the more labor-intensive and often more costly ways of obtaining risk and value information, such as on-site inspections. Many insurers are also partnering with Zesty.ai in an effort to improve the accuracy of the underwriting process. Zesty.ai, which recently won the silver award at Zurich’s global insurtech competition, uses computer vision paired with other “data points on residential and commercial properties to extract key building features to accurately model the potential impact of catastrophic . . . loss events.”
Claims processing. AI is equally useful in claims processing and already is transforming the claims world.
With the introduction of AI, human claims management will become more specialized, focusing on the “unusual” and/or “contested” claims. AI technology will take over the responsibility for routing initial claims and for most policyholder voice and text interactions. Touchless insurance claims-processing software “can remove excess human intervention and can report the claim, capture damage, update the system and communicate with the customer all by itself.” This is expected to save insureds time and hassle. McKinsey predicts that by 2030 “IoT sensors and an array of data-capture technologies, such as drones, largely [will] replace traditional, manual methods of first notice of loss.” Galacticar, a Galaxy product, already uses machine vision and deep learning to automate the claims process by letting insureds submit claims on their smartphone and to prioritize high-severity claims.
In the near future, the turnaround time for most claims will be “measured in minutes,” not days. Currently, AXA is saving about 17,000 man-hours per year by using AI bot “Harry,” which performs routine data entry and automatically attaches claims records to emails, allowing claims staff to focus more time on substantive, analytical tasks.
Xactimate and Colossus: Lessons from Litigation
The insurance industry has been using AI for several decades. “[T]he initial forms of AI, such as case-based reasoning and rules engines, were deployed for underwriting and claims fraud in th[e] early days” of insurance usage of AI. The AI was simpler then, and these early efforts led to mixed results.
Xactimate and Colossus are two examples of early AI software programs used in the insurance industry to calculate damages and quantify loss. Because these programs, which are still employed by the industry, have been around for a long time, there is no dearth of cases involving issues related to them. Looking back at this litigation history may assist insurers as they traverse the current terrain of AI and look forward to a world where AI surely will be the dominant landscape.
Xactimate. In 1986, Xactware Solutions opened and debuted Xactimate, an innovative computer program allowing contractors and insurance professionals to create estimates for building construction, remodeling, and repairs faster and more accurately. Today, Xactimate is touted as “the industry’s most powerful and comprehensive solution for property claims estimation.” Industry professionals have used Xactimate not only to generate loss estimates but also to inform their reserve-setting and claims-settlement offers. Twenty-two of the top 25 property insurers in the United States use Xactimate software.
Xactimate is referenced in nearly 200 published cases nationwide, the earliest in 1996. A few of these cases actually discuss the utility and appropriateness of using the AI software program. The vast majority of the cases have concluded that Xactimate is a common estimating program in the insurance industry and that opinions based on Xactimate calculations are sufficiently “sound” and “reliable” to withstand Daubert challenges. Consistently, courts have deemed the software program an accepted “industry standard” for adjusting claims or losses.
Courts routinely find that arguments regarding data input inaccuracies and application errors “go[] more to the weight of the evidence, rather than its admissibility.” Calculation and input errors, therefore, fail to justify the exclusion of an expert’s testimony regarding the utilization of Xactimate. As one court explained,
[w]here an expert applies a sound methodology but commits errors or relies on incomplete information in reaching his conclusions, [s]uch deficiencies impact the weight of the expert’s testimony rather than its admissibility.
As another court stated, such concerns “are best resolved by vigorous cross-examination and the presentation of contrary evidence.”
The use of the Xactimate software program also has been the basis of class actions. In McKinnie v. State Farm Fire & Casualty Co., for example, the plaintiffs brought a putative class action lawsuit against State Farm, contending that the insurer, through its manipulation of Xactimate, had a “pattern and practice” of failing to pay overhead and expenses mandated by Tennessee law. Plaintiffs claimed that
even though Xactimate allows an . . . adjuster to . . . provide full payment for a prime contractor’s overhead and profit when calculating replacement cost value, State Farm intentionally change[d] the settings to omit [this] overhead and profit.
However, the plaintiffs’ breach of contract claims were ultimately dismissed for failure to state a claim in light of prevailing case law pertaining to the calculation of overhead.
In addition to class actions, the Xactimate software program has generated bad faith claims. For example, in Sands v. State Farm Fire & Casualty Co., the plaintiff alleged that State Farm handled her claim in bad faith by relying on Xactimate to calculate depreciation without investigating the “assumption models” that the software relies on. The plaintiff claimed that because the State Farm adjuster did not know whether the models were accurate and did not investigate the models to verify their accuracy, State Farm’s “reliance on the calculation was ‘recklessly indifferent to the rights of Plaintiff.’” The argument was deemed unpersuasive. The court reasoned that “because Xactimate is a standard software in the insurance industry for estimating replacement costs,” State Farm had “a reasonable basis to believe that [it] could rely on Xactimate’s calculations.” In granting summary judgment to State Farm on the bad faith claim, the court stated that
[t]his Court cannot conclude that State Farm acted in bad faith by not second-guessing the depreciation calculations of an industry standard computer program whose sole purpose is to accurately make those calculations.
Colossus. Computer Services Corporation developed the computer program Colossus to calculate settlement offers for bodily injury claims using information provided by adjusters in response to questions generated by its software. Many major insurers have found Colossus useful “because it considers a great many factors” in quantifying a settlement range, which is then used as one piece of the valuation puzzle. It was first used by Allstate in the 1990s as a way to standardize its claims valuation and processing.
Colossus is mentioned in nearly 90 published cases, the earliest in 1998. Most of the cases, however, do not discuss the software program in much detail.
In one case, Brown v. Allstate Indemnity Co., Allstate’s claims-handling practices came under scrutiny, including its use of Colossus to determine the settlement value of claims. It was found that Allstate periodically updated Colossus with recent settlements, which impacted how the program valuated injuries. It also was found that Colossus does not factor in special damages (such as lost wages or other out-of-pocket expenses) on its own; rather, this figure is input manually by the adjuster. The software program also does not consider the percentage of liability owed and whether there is comparative negligence. Any percentage of comparative negligence must be input manually by the adjuster. The program then “reduces its recommendation by th[e] manually entered percentage.” Nor does the program “assess or revise offsets.” The examinations concluded with a strong recommendation that there be
enhanced management oversight to ensure adherence to established criteria for selection of claims to use in the tuning process and a uniform methodology for determining the number of claims to be used in each tuning region.
The examinations, however, revealed “no evidence of improprieties with respect to any particular claim selection or tuning.”
In other cases, too, like McLaughlin v. Nationwide Mutual Insurance Co., courts have focused on insurers’ use of the software program—namely, the adjusters’ ability to stray from the software’s recommended settlement value and bring their own judgment to the unique set of facts underlying a claim. In McLaughlin, the court explained that Colossus was just one of several components used to evaluate a claim. The court also emphasized that adjusters could “disagree[] with the range of damages suggested by Colossus” and that the program was “not used on all casualty claims.”
Liability Concerns
Despite the fact that courts have been generally supportive of AI insurance software programs, cautious commentators have warned against relying too heavily on AI. Their list of concerns is long. Companies using external data need to comply with a myriad of privacy and regulatory rules. Computer system malfunctions are common and may lead to serious consequences. What if a software glitch overlooks the need for certain property coverage? What if a system error results in inaccurate claims processing or damage assessment? How will insurers face the threat of social engineering and hackers?
The liability concerns are complex ones. How will society (and eventually the courts) effectively allocate risk and products liability when insurers rely so heavily on software programs purchased or used by outside vendors? How will bad faith claims unfold in this strange, new arena?
Many also worry that automation will reduce or eliminate jobs and that, therefore, the human element and application of judgment will be limited or lost. This is particularly troubling because, after all, it is this human element that becomes critical when insureds face devastating and catastrophic loss.
Conclusion
The story of AI’s fascinating, complicated, exciting, and worrisome uses in the insurance industry began a number of years ago, but the plot is becoming more intense and compelling.
The beginning of this tale involves Xactimate and Colossus, two of the insurance industry’s earliest AI software programs. A careful look back at the litigation history of issues involving these programs reveals important lessons that we should heed going forward. For example, there must be a partnership between humans and technology. As the Colossus and Xactimate cases make clear, the programs must be tuned and overseen continually, mandating human involvement and insight. Furthermore, the Colossus and Xactimate case law evidences that data advances can and should be harnessed by the industry for use with the more routine and tedious tasks, but certainly not all tasks.
Technology is advancing at a rapid pace, but advances in technology will not guarantee a perfect system. The potential for error and miscalculation will always exist. Thus, the soundness and use of AI insurance software programs most certainly will be vetted and challenged. And the thrill of innovation will be tempered by the realities of complicated questions regarding risk allocation and liability—questions that most certainly will find their way into our courtrooms before long.
Insurers should proceed with caution as they participate in the creation of AI systems and adopt these programs. Only with proper use, data vetting, and continual tuning will insurers be able to look at AI processes and software and quip, “It is a far, far better thing that I do, than I have ever done.”
