Artificial intelligence (AI) has enormous potential to disrupt across the data- and process-heavy insurance value chain.
The growth of AI startups and adoption of (AI) by Australian insurers began in earnest over the last two years.
The approach taken so far is one of collaboration rather than direct competition between incumbents and disruptors.
Global investment in AI-related InsurTechs will reach at least $US1.42 billion in 2018, spread over 50 deals. Total in-house investment by incumbents building AI capabilities was projected to be US$450mn in 2018, and grow to US$1.44bn by 2021
About 9% of Australian InsurTech start-ups, representing 8–10 players, claim AI as their core technological innovation. Noteworthy players include Huddle, Flamingo AI and Audeamus Risk
Of the publicly disclosed raises, at least A$38mn has been raised by Australian AI-focused startups targeting the insurance industry across both B2B and B2C segments
The deployment of AI will be shaped by the problem of algorithmic auditability and the regulatory response. If Australia’s Consumer Data Right mirrors the EU’s GDPR regime, it will impose serious restrictions on financial institutions’ ability to use unsupervised machine learning and deep learning
Introduction to Australia’s Insurance Industry
Australia's insurance market can be divided into roughly three segments:
General insurance: Coverage for loss or damage to everyday items such as cars, property and belongings, but also extending to specialty areas such as business and travel.
Life Insurance: Coverage in case of death or incapacity.
Health Insurance: Coverage to pay for medical bills in case of illness or injury.
In 2017, the Australian insurance market was worth A$100bn in premiums and A$7.2bn in profits. Australians spent over A$100bn on insurance premiums in 2017, making the insurance market a huge target for the tech industry. Life insurance is the largest category, followed by General and Health insurance. Both the General and Health insurance sectors are seeing steady growth in net premiums, while Life insurance is in decline with major providers either being sold or exiting the sector.
Figure 1. Australian insurance premiums, 2014-17
SOURCE: APRA QUARTERLY STATISTICS
Traditionally, insurers sold most of their business through third-party brokers. More recently, many insurance companies have expanded their direct sales channels, developing digital sales channels alongside traditional telemarketing and advertising. Many InsurTechs are seeking to disintermediate either brokers, insurers, or both, by providing services directly to customers.
Figure 2. Insurance industry structure
SOURCE: VENTURE INSIGHTS
Market leaders vary between the three segments of the industry (see Figure 3 below).
Figure 3. Market position across major insurance markets in Australia, 2017
Major use cases for AI in insurance value chain
AI is affecting the whole of the insurance value chain, whether its marketing and distribution, pricing and underwriting, back office efficiency, or remediation and loss management.
SOURCE: Simon O’Dell, CEO, InsurTech Australia 
For an introduction to AI, please see Appendix 1.
The key elements of the insurance value chain are:
Marketing and sales distribution
Servicing and policy administration
Recruitment and internal management
Typically, there is the greatest scope for AI to achieve higher performance in areas that are data or people-heavy or those involving relatively repetitive tasks and rule application. This creates opportunities for insight to be extracted and processes/interactions to be automated. Thus, in insurance, major use cases for AI include (in a rough order of sophistication, moving from existing ‘narrow intelligence’ applications such as automation to more complex techniques such as unsupervised machine learning):
Lowering operating costs - for example, backend automation of claims processing and transactions
Improving customer experience across the customer journey - for example, rapid quoting and underwriting (driving higher completion rates), personalised policy recommendations, and conversational AI for queries and complaints
Delivering strategic business insights about internal operations, risk pricing, fraud detection and claims assessment - for example, advanced analytics of internal data sets by human-compatible AI (employees ask natural language queries, that are processed by an algorithm into structured queries asked of internal data sets)
Shifting policy holder risk profiles - for example, using AI to reduce claims using a ‘predict and prevent’ method, advising high-risk policy holders on how to shift their behaviour
Allowing dynamic pricing of risk - for example, combining machine learning with streams of data from connected devices (such wearables, IoT home devices and autonomous vehicles) to assess risk profiles more accurately in real-time, then price coverage dynamically
For further examples of AI use cases in insurance, see Figure 4 below:
Figure 4. Use cases for AI across the insurance value chain
As a result, most insurers are pursuing a mixed strategy of internal investment (building AI-related capabilities) and external investment (VCs, M&As, incubators, etc.) to reap the benefits of AI for their industry while protecting against market share loss. This spending is occurring across technologies, with AI representing a significant priority for most incumbents. Given the overlap between AI, robotic process automation (RPA) and the field of data analytics, Figure 7 likely understates the importance of AI as a broad domain of investment for insurers.
Figure 7. Incumbent insurers’ ratings of technological investment priorities for 2018
Internal spending by incumbent players in the market on AI technologies has been projected to grow decisively over 2016-21. Overall spending is projected to rise at a CAGR of 48% from US$205mn to US$1.4bn in 2021, with total spending in 2018 projected to be ~US$450mn. AI services and software are the biggest and fastest growing portions of this spend, followed by spending on hardware:
Spending on services is expected to rise from US$99mn in 2016 to US$752mn in 2021 at a CAGR of 50.0%
Spending on software is expected to rise from US$76mn in 2016 to US$561mn in 2021at a CAGR of 49.2%
Spending on hardware is expected to rise from US$29mn in 2016 to US$119mn in 2021 at a CAGR of 32.7%
Figure 8. Global insurance IT spending on AI technologies, US$Mn
SOURCE: Deloitte, 2018
In 2017, 45% of global incumbents reported partnering with InsurTechs. High barriers to entry mean that strategic investment by incumbents (as opposed to by industry-agnostic private capital) is a disproportionately large part of the InsurTech startup funding landscape.For this reason, trends in in-house investment by insurers and investments in the InsurTech landscape track each other quite closely (see next section InsurTech startups for more detail).
Investments by incumbents in AI-focused InsurTechs are on track to hit record levels in 2018, with at least 18 deals globally in the first half of the year, a more than 2.5x increase in the first half of 2017. The dominant focus of these deals so far has been on improving underwriting (48%), claims management (33%) and product development (24%), followed by sales and distribution (18%) and policy administration (18%).
Figure 9. Global investment patterns in AI-focused InsurTechs
InsurTech start-ups were expected to receive ~US$3.2bn in funding in 2018, the largest single year on record, of which at least 44% will go to AI-related startups
Funding in the global general InsurTech landscape has been growing slowly over the past four years (~3.3% CAGR) after spiking sharply upwards in 2015. 2018 looks to be the largest year yet; in fact, global InsurTech investment in the first six months of the year stood at ~US$1.6bn over 97 deals, putting it on track to surpass the record highs from 2017.
Figure 10. Investment in InsurTechs globally, 2010-18, US$mn
Globally, InsurTechs are active in all of the major insurance products and business lines, with concentrations in the P&C business (known as general insurance in Australia) and in the marketing and distribution areas of the value chain (see Figure 11 below).
Figure 11. Distribution of InsurTechs across sectors and value chain
Note: 1.~500 commercially most well-known cases registered in the database (excluding wealth management-related innovations). 2. Includes Sales. 3. Includes underwriting and policy issuance
The US has been the leading market for InsurTech startups, with North America and EMEA accounting for ~86% of the global market in 2017, but APAC is projected to be the fastest growing region in the coming years. The distribution of InsurTechs by geography as of 2017 was:
46% of InsurTechs headquartered in the Americas
40% of InsurTechs headquartered in EMEA
14% of InsurTechs headquartered in APAC
AI-related startups were estimated to represent a quarter of deals and ~44% of the total financing flowing to InsurTechs globally in 2016; this represented a rapid acceleration compared to 10% of the deal value they constituted in 2015. Assuming these proportions have held, AI- and IoT-related InsurTechs will have received at least US$1.42bn in funding from over 50 deals in 2018. The real figure may be substantially higher, given the growth of interest in AI within the industry over the last two years.
Australian investment landscape
Incumbents are beginning to make internal investments in building AI capabilities
Australian insurers are still in the early stages of capturing the potential of AI for their businesses; for example, ANZIIF records that just 20% of Australian companies have deployed AI chatbots, a relatively common business application of AI, with the number predicted to rise to 57% by 2021. EY and InsurTech Australia’s 2018 survey of the Australian insurance landscape reports that most insurance executives have one or more of the ‘Big 5’ technologies ‘on their radar’ - connected devices, AI and machine learning, blockchain, data analytics and platforms - but are finding it challenging to incorporate InsurTech into their innovation strategy while also serving existing customers and sustaining value for shareholders.
Some incumbents have announced efforts to incorporate AI into various parts of their value chain, but AI is a much less prominent theme in the public-facing communications of Australian insurers than their counterparts in the banking sector at present. Examples of AI usage announced by incumbents tend to be about scaling up pilots of customer-facing AI applications (e.g. chat bots) and beginning early-stage research partnerships for more complex applications of AI (e.g. improving risk profiling and underwriting):
NIB’s chatbot nibby, launched in December 2017, has reportedly handled more than 21,500 member interactions, with a 70% success rate, saving 535 hours of consultant handling time. It is now being expanded across group operations.
CUA’s virtual sales assistant Sam (trialled for health insurance and developed with Flamingo AI) will be expanded around the business.
Swiss Re Life and Health Australia have teamed up with mobile-led financial-services business Raiz Invest (formerly known as Acorns) to jointly research and develop tailored and personalised super insurance products using machine learning algorithms based on technology developed by Swiss Re Life. The goal is to deliver a simpler underwriting experience for customers, which will be used for a Raiz Invest Super product.
ANZ Wealth is collaborating with researchers at the University Of Technology Sydney (UTS) - Advanced Analytics Institute to explore how machine learning models can improve the insurance underwriting process. Specifically, they are investigating how client behaviour modelling (which is concerned with analysing personal statements and understanding the relationship between certain questions and outcomes), text mining and natural language processing, along with social and predictive analytics, can add value to the insurance sector.
While industry-wide numbers are not available for levels of AI-related internal investment by Australian insurers, estimating Australian investment levels by proportionately adjusting global investment trends would imply total in-house spending of ~A$10.8mn on AI software, hardware and services in 2018, rising to ~A$33.2mn by 2021. Pricing and underwriting models are reportedly the most popular use cases currently for AI among Australian insurers, particularly in the areas of General and Life insurance.
Insurtechs are looking for more strategic engagement from incumbents
In a survey of Australian InsurTechs, most startups expressed a desire for broader and deeper engagement from incumbents with technological innovation. About 65% classified themselves as aiming to add value to the existing value chain (rather than disrupt it), but less than 20% believed incumbent players were doing enough to collaborate with InsurTechs in driving innovation . As Aleksandar Kovacevic, Managing Director of Audeamus Risk, says, the “corporate sector appears somewhat ignorant and very slow in adopting new technologies and data-sets that already exist on the market”.
That said, Australian insurers have begun engaging much more closely with InsurTechs over the last 18 months, especially since the creation of a peak body - InsurTech Australia - which represents both incumbents and startups, deliberately calling for collaboration rather than direct competition. With founding partners from across the existing financial services and insurance value chains, including EY, Munich RE, DLA Piper, AUB Group, Macquarie, Stone & Chalk, TAS, QBE and dxc technology, its membership includes 45 InsurTech startups (which represents over half the players in the ecosystem). Its goal is ‘to make Australia a world leader in InsurTech by fostering a diverse community of insurance innovation and collaboration’. The key findings of its landmark report with EY, InsurTech: Enabler or disruptor? include:
Greater levels of collaboration are needed between InsurTechs and incumbents
InsurTechs and incumbents need to identify and leverage their strengths when seeking partnerships
InsurTechs are bootstrapping their ventures, with short runways. This puts them at increased solvency risk, exacerbated by prolonged and drawn-out sales cycles with B2C and B2B players
The Australian ecosystem is young, growing very fast and represents a great launchpad into other regions
As Simon O’Dell, CEO of InsurTech Australia, told Venture Insights, there are strong incentives for collaboration over competition in the insurance industry:
“We recently launched a report with EY on the topic ‘InsurTech - Enabler or Disruptor?’ , but in more developed markets, the conversation actually moved on from that question years ago. Now, it’s about how partnering works and what an InsurTech strategy looks like for an incumbent - for example, the merits of an InsurTech strategy based on VC versus incubators versus outsourcing…
Two years ago, InsurTech in Australia was very underdeveloped, but it has matured quite quickly since; our organisation, InsurTech Australia was established as an ecosystem builder and nurterer to enable that. I would say two-thirds of Australian startups seek to enable incumbents, and those who don’t are proportionally not that significant at this stage…
Most incumbents, including Australian firms, are beginning to pursue multi-pronged InsurTech strategies, balancing venture capital, in-house incubation and outsourcing to contractors… The barriers to entry are high in the insurance industry, and a lot of Australian InsurTech startups are after smart money - investors who can do more than just invest the capital to grow. That includes a lot of the other assets required to achieve a step-change in operating capacity; startups want to be able to use an existing licence, to get help in meeting capital requirements, to be introduced to other incumbents in the value chain and get access to data sets. Given the complexity of insurance products, there is a high ratio of strategic investors (versus VC firms) compared to other value chains - VCs are typically quite ‘standoffish’ towards InsurTech in Australia.”
AI-focused start-ups are still a relatively small portion of the Australian InsurTech landscape, but several high performers stand out. For case studies on how Australian startups are using AI, see Appendix 4, and for maps of the Australian Insurtech landscape, see Appendix 5.
A relatively small percentage of Australian InsurTech startups describe their primary technological innovation being in the field of AI and machine learning (9%, which would equate to around 8-10 players in the landscape); they are over-represented in certain segments of Insurtech, such as loss prevention and mediation segment (30%).
Figure 12. Predominant technological innovations among Australian InsurTechs, 2018
Of these AI-focused start-ups, three are particularly prominent; Flamingo Ai, Huddle and Audeamus Risk (profiled in the Local case studies section below). A vast majority (81%) of Australian InsurTechs are bootstrapped by their founders , and of the discoverable AI startups in the Australian ecosystem, disclosed funds raised total ~$38m to date.
Managing the risks of AI
AI may be a Double-Edged Sword
AI may be a double-edged sword for the insurance industry, bringing a set of commercial and repetitional risks that need to be managed carefully
The transformative changes made possible by AI in an industry such as insurance come with a set of downside risks, some of which could incur serious legal, reputational and/or commercial costs. The two major classes of risks are algorithmic transparency and data stewardship.
The set of risks regarding data stewardship are relatively easy to understand the consequences of; they involve handling issues such as privacy and data ownership in full compliance with the law. Compliance costs will likely rise as insurers are required to build out an IT infrastructure capable of protecting consumer privacy while also allowing consumers to transfer their data to other financial institutions easily.
Algorithmic transparency is a little more complicated. In essence, it is about being able to understand the logic used by an algorithm to make a decision, and explain that decision to regulators, customers and internal stakeholders. With more sophisticated forms of AI, such as unsupervised machine learning and deep learning, this is frequently extremely difficult or impossible (a limitation known as the ‘black box problem’). These algorithms are designed to operate and learn without supervision, and to make decisions based on complex patterns identified in untagged data
If algorithms cannot be audited, they may be effectively un-usable; European regulators have, as part of their General Data Protection Regulation (GDPR) reforms, already established a standard that requires financial institutions be able to explain in easily comprehensible terms any pricing and servicing decision for which a consumer requests an explanation. This aspect of the GDPR is described by UK Information Commissioner Elizabeth Denham below:
“...where a decision has been made by a machine that has significant impact on an individual, the GDPR requires that they have the right to challenge the decision and a right to have it explained to them… We may need, as a regulator, to look under the hood or behind the curtain to see what data were used, what training data were used, what factors were programmed into the system and what question the AI system was trained to answer.”
The US’ Fair Credit Reporting Act also requires that companies notify a consumer if consumer report information is used to deny credit; if similar principles are applied to the use of consumer data in insurance, it may be difficult for firms using AI to make risk pricing decisions to provide notifications and rationales for such decisions. In the medium-term, this may retard or substantially slow the adoption of more complex forms of AI as insurers grapple with the audibility issue.
The other risk with algorithmic transparency is obviously the inadvertent mispricing of risk by an algorithm; without the ability to understand what pattern is the premise of a recommendation, such errors may not be picked up or cannot be corrected if they are discovered. The effect of this would be to increase claim costs above their theoretical minimum. Figure 13 below lays out a comprehensive set of risks that enterprises in financial services are likely to encounter, and a set of possible responses:
Figure 13. Risk management needs for enterprises applying AI to financial services
European and British regulators have led the response to the risks of AI, while Australia is beginning to construct a framework for AI ethics to accompany its Open Banking regulations and Consumer Data Right (CDR).
“In some cases, AI might make decisions and we will never understand why – what was behind its ‘thinking’?... Yet the potential for improving the lives of people and communities is truly immense…”
SOURCE: Shayne Elliot, ANZ CEO, October 2018
“[The] regulatory environment has not been sufficiently shaped [for AI in insurance], especially in the domain of personal information and decisions that involve various ethical issues… we also see weaponisation (inappropriate use ) of an AI as a serious future threat that requires well-orchestrated response from the entire industry.”
SOURCE: Aleksandar Kovacevic, Managing Director, Audeamus Risk
“Regulations governing the privacy and portability of data will shape the relative ability of financial and non-financial institutions to deploy AI, thus becoming as important as traditional regulations to the competitive positioning of firms.”
SOURCE: World Economic Forum report on AI in Financial Services, 2018
Australia’s overall policy response to the challenge of AI is currently being formulated; while the country does not currently have a formal AI strategy, the 2018–19 Federal Budget promised a four-year, $29.9 million investment to support the development of AI in Australia. In addition to investments in R&D, this package also included the development of a Technology Roadmap, a Standards Framework, and a national AI Ethics Framework to support the responsible development of AI.
Limited insurance-specific regulations currently exist regarding the use of AI anywhere in the world, and none appear to exist in Australia. In Australia, the most likely route to regulation of the industry would be through replicating principles that were developed to regulate how banking institutions handle consumer data - namely, Open Banking and CDR. In essence, Open Banking enables detailed personal financial data to be shared between organisations, with explicit consent and privacy safeguards for consumers built in through the CDR, which is based on the premise that consumers have ownership of their data and a continuing right to control its usage. Open Banking is designed to end the competitive advantage of traditional financial institutions, which they derive from privileged access to comprehensive customer data, and enable new entrants to compete on an equal playing field. On the other hand, the CDR is designed to protect consumers in a world where data exploitation is an enormous commercial imperative.
European and British regulators are currently leading in this area. In 2015, the European Parliament adopted the Revised Payment Services Directive (PSD2), which went into force in 2018. PSD2 requires banks to share customer data and allow customers to use third-party providers for payment services. The UK’s Open Banking regime was created in response to findings by the competition regulator that there was a lack of competition for established banks; since 2017, the UK’s nine largest banks have been obligated to give regulated providers access to customers’ banking data. The European GDPR, involving a set of rights including a right to have credit and risk decisions explained), came into force in mid-2018 after being agreed upon in 2016.
The 2018 Farrell Review recommended implementation of an Open Banking framework in Australia, along the following timeline:
By July 2019, all major banks must make credit /debit card, deposit and transaction account data available.
By February 2020, all major banks must make mortgage data available.
By July 2020, all major banks will need to make data on all products available.
By July 2021, all remaining banks will be required to adopt Open Banking.
This will require Australian financial institutions to ensure they have effective application programming interfaces (APIs) to enable data sharing, and data governance infrastructures that are compliant. Draft legislation introducing the CDR was due to be introduced in Parliament in December 2018.
Deep dive: The talent shortage in Australian AI
"It’s very difficult to get machine learning or artificial intelligence talent in Australia, especially at the PhD level; anyone worth their salt will be head hunted or taken overseas, and I understand why, because far more advanced things happening there. The skill entry level is much higher; I meet young data scientists in the US who can’t get jobs because they need at least a graduate degree.”
SOURCE: Dr Catriona Wallace, CEO, Flamingo AI
“[Australian] fintechs are drawing on a limited start-up talent pool and particularly struggling to find engineering / software expertise. This issue will only become more pronounced as artificial intelligence (AI) and machine learning become increasingly integral to fintech offerings. Answers may include tempting ex- pat Australians back to the local industry, and recruiting for cultural fit and attitude (rather than skills), and upskilling in-house.”
SOURCE: EY Fintech Australia Census, 2018
“Unfortunately, Australia falls behind major global economies in accessing the skills and capability to effectively implement and manage AI across the enterprise… [and] the new visa restrictions mean Australian organisations will find it harder to borrow workers from overseas as it will significantly impact the free-flow of talent coming to Australia.”
SOURCE: Aaron McEwan, HR Advisory Leader at Gartner
“There are half as many AI pioneers in Australia as there are in US and Europe… The quality of the fundamental research on AI in Australia is very high, it’s some of the best in the world but there’s a big need for translators… so far adoption is not happening in the critical areas of business.”
SOURCE: Richard Kimber, former Google Australia MD & CEO of AI startup Daisee
Australia is more talent-constrained than funding-constrained when it comes to AI:
71% of Australian fintechs report talent shortages in engineering and software, with experience in design / UX (36%) and sales (33%) lagging far behind
Other countries have much larger and more skilled talent pools in AI:
Australia ~3,000 (mostly based in Sydney)
Appendix 1: Primer on ‘artificial intelligence’
“AI… is more profound than electricity or fire.”
SOURCE: Sundar Pichai, CEO, Google
“Rapid technological disruption will break apart the Australian banking oligopoly and overturn the universal banking model that has underpinned big bank profits for decades…”
SOURCE: Australian Financial Review, October 2018
“I think we’ll see 50% or even greater reduction in people employed in the industry [due to AI]”
SOURCE: Antony Jenkins, former Barclays Chief Executive, 2018
“It’s absolutely important to be slow [in AI adoption], especially in financial services.”
SOURCE: Ramneek Gupta, MD, Citi Ventures, 2017
“...where a decision has been made by a machine that has significant impact on an individual, the GDPR requires that they have the right to challenge the decision and a right to have it explained to them… We may need, as a regulator, to look under the hood or behind the curtain to see what data were used, what training data were used, what factors were programmed into the system and what question the AI system was trained to answer.”
SOURCE: Elizabeth Denham, UK Information Commissioner, 2018
“There is, clearly, an AI bubble at present; the question is whether this bubble will burst (cf. the dot com boom of 1996-2001), or gently deflate; and when this happens, what will be left behind? My great fear is that we will see another AI winter, prompted by disillusion following the massive speculative investment that we are witnessing right now. There are plenty of charlatans and snake oil salesmen out there, who are quite happy to sell whatever they happen to be doing as AI…. However, while I think some deflation of the current bubble is inevitable within the next few years, I think there is cause for hope that it will be a dignified and gentle deflation, rather than a spectacular bust. The main reason for this is that AI is delivering on competence. Across a wide spectrum of tasks, AI systems are showing steadily (sometimes rapidly) increasing performance, and these capabilities are being deployed with great success in many different application areas. To put it another way, I think there is substance underneath the current AI bubble, and major companies now understand how to use AI techniques productively.”
SOURCE: Michael Wooldridge, Director of the Oxford University Department of Computer Science, 2018
What do we mean by ‘Artificial Intelligence’?
Figure 14. Timeline of AI reaching human-equivalent performance across narrow tasks
AI is inherently an ambiguous term, but in a general sense, it refers to the idea of machines that can mimic or exceed what we understand to be ‘intelligence’ in humans. Though the concept of AI was first posited in theoretical terms during the 1950s, and experienced stop-start periods of practical advancement in the latter decades of the twentieth century, academia has advanced rapidly in the post-2000 era (see Figure 10 below) due to increases in affordable computing power, rise in the volume and variety of data, creation of large data sets, speed of access to data and the emergence of new techniques able to analyse the flood of data.
Australia produces a substantial amount of fundamental research in AI per capita, but globally, the field is dominated by the US and China (see Figure 16 below). China takes the lead from the US in around 2013 and has been continuing to invest heavily in becoming the world’s leading AI power (see the deep dive on US-China competition for AI leadership below for more detail).
Figure 16. Comparative number of research papers published on deep learning, by country
Computer scientists, AI safety experts and AI ethicists frequently refer to the distinction between artificial narrow intelligence (ANI) and artificial general intelligence (AGI). ANI means a machine has the capacity to perform discrete, programmed tasks in narrow domains and produce known outputs; machines have already reached this milestone in a series of domains, and are predicted to conquer a series of increasingly complex domains over the upcoming decades (see Figure 12, on next page). All of the innovations discussed for commercial purposes currently fall under the category of ANI, and while they may pose challenges for principles of transparency and fairness (see section on Regulatory Outlook), they may not pose significant existential risks to humans.
Figure 17. AI expert predictions on timeline for matching human performance
Note: Median expert predictions are represented by dots; minimum and maximum are represented by lines.
AGI refers to a machine’s ability to achieve human-level intelligence which is not programmed into it or restricted to a specific domain; the ability to reason, plan, solve problems, think abstractly, comprehend complex ideas, learn quickly, and learn from experience. Experts disagree on the timeline for the development of an AGI, but the majority expects it within the next 100 years, with the median prediction being less than 50 years.
AI: Sorting the hype from the reality
The term AI can refer to applications of varying levels of sophistication and commercial usefulness; as with other technological innovations, the uncertainty around the potential transformative value of AI and its market effects incentivises many players in the public and private markets to ‘jump on the bandwagon’, both in terms of investments and branding.
Figure 18. Perceptions of emerging technology in banking among CIOs
… when business people talk about AI, they typically are not talking about a particular technical approach or a well-defined school of computer science, rather they are talking about a set of capabilities that allows them to run their business in a new way. At their core, these capabilities are almost always a suite of technologies, enabled by adaptive predictive power and exhibiting some degree of autonomous learning, that have made dramatic advances in our ability to use machines to automate and enhance pattern detection, foresight (determining the probability of future events), customisation, decision-making and human-machine interaction.
The key techniques and technologies often clustered under the banner of ‘AI’ include:
Machine learning (ML): Most recent advances in AI have been achieved by applying ML to very large data sets. ML algorithms detect patterns and learn how to make predictions and recommendations by processing data and experiences, rather than by receiving explicit programming instructions. The algorithms learn from new data and experiences to become more effective over time. Machine learning is broadly divided into two learning methods. The most prevalent method at work today is supervised learning, with unsupervised learning showing great promise for broader applications.
Supervised learning: An algorithm uses a known dataset to make inferences based on labelled input and output data; for example, using population demographics and season to predict clothing demand. Typically used when the input data is pre-classified and the desired output is known, but it needs to be calculated using new data. The algorithm is trained on a sub-set of data, and once it is sufficiently accurate in connecting inputs to outputs, applied to the larger data set.
Unsupervised learning: An algorithm analyses datasets containing unlabelled data (for example, a dataset describing omnichannel customer journeys), infers a structure from the data, and identifies patterns (such as customers who exhibit similar preferences). Best used when the data has not been classified, and the algorithm can identify patterns and perform classification.
Reinforcement learning: Another important sub-part of ML. An algorithm learns to perform a task simply by trying to maximise rewards it receives for its actions (for example, maximising returns by trying different timing on stock orders), and optimises for the most effective approach over time. Best for cases with limited training data, where it is difficult to define the ideal end state, or the only way to learn about the environment is to interact with it.
Deep learning: A form of ML, deep learning algorithms can process a wider range of data, requires less pre-processing and can often produce more accurate results than traditional ML approaches. Deep learning algorithms are becoming more and more popular because of their effectiveness in tasks related to speech and computer vision. They are complex techniques where it is hard to decipher exactly how each input drives model outcomes, often resulting in them being characterised as “black boxes”.
Neural networks: In deep learning, interconnected layers of software-based calculators known as “neurons” form a neural network, modelled on the human brain. It can process vast amounts of input data and process them through multiple layers that learn increasingly complex features of the data at each layer. The network can then make a determination about the data, learn if its determination is correct, and use what is has learned to make determinations about new data. For example, once it learns what an object looks like, it can recognise the object in a new image.
Robotic process automation (RPA)
While not considered to be sophisticated AI, or even true AI by some, RPA is widely used and often described under the banner of AI in non-expert commercial contexts. RPA involves software that mimics the activity of a human being in carrying out a task within a process
“RPA deals with simpler types of task. It takes away mainly physical tasks that don’t need knowledge, understanding, or insight—the tasks that can be done by codifying rules and instructing the computer or the software to act. With cognitive automation, you impinge upon the knowledge base that a human being has and on other human attributes beyond the physical ability to do something. Cognitive automation can deal with natural language, reasoning, and judgment, with establishing context, with establishing the meaning of things and providing insights”
Beyond analysing traditional numerical data sets, key emerging applications of AI include:
Speech recognition and Natural Language Processing (NLP): The ability to understand and generate human speech the way humans do by, for instance, extracting meaning from text or even generating text that is readable, stylistically natural, and grammatically correct. Natural language processing (NLP) is the ability of computers to analyse, understand and generate human language, including speech
The next stage of NLP is natural language interaction, which allows humans to communicate with computers using normal, everyday language to perform tasks. This will form part of the process of cognitive computing
Cognitive computing: A subfield of AI that strives for a natural, human-like interaction with machines. Using AI and cognitive computing, the ultimate goal is for a machine to simulate human processes through the ability to interpret images and speech – and then speak coherently in response
Visual recognition: the ability to identify objects, scenes, and activities in images. Computer vision technology uses sequences of imaging-processing operations and techniques to decompose the task of analysing images into manageable pieces. Computer vision relies on pattern recognition and deep learning to recognise what’s in a picture or video. When machines can process, analyse and understand images, they can capture images or videos in real time and interpret their surroundings
These capabilities have advanced rapidly in the last several years, with some nearing and even exceeding human equivalents.
Appendix 2: Global case studies (startups)
Using AI and behavioural economics to change the insurance industry for good
Funding raised: US$180mn over 5 rounds, from investors including Sequoia Capital Israel, Allianz, Softbank and Tusk Ventures
P&C (property and casualty) insurer based in New York, specialising in home and renters insurance for urban areas
Valuation: December 2017’s raise put Lemonade’s pre-money valuation at US$500mn
The company uses artificial intelligence and chatbots to deliver insurance policies and handle claims for its users on desktop and mobile (through its iOS and Android applications) without employing insurance brokers
In 2016, a customer filed a claim for a stolen coat; after answering a few questions on the app and recording a report on the iPhone, Lemonade’s claims bot set a world record by reviewing, approving and paying the claim in three seconds.
Lemonade keeps a flat 20% fee of a customer’s premium while setting aside the remaining 80% to pay claims and purchase reinsurance.
Lemonade has B-Corp status and donates leftover unclaimed premiums to non-profits of its customers’ choice in an annual ‘Giveback’
Other key points:
Duke University professor, author and economist, Dan Ariely, is the Chief Behavioural Officer at Lemonade.
Much of his research on behavioural economics integrated into the DNA of the company, in pursuit of transforming the adversarial relationship between the client and company. Ariely is tasked, in part, with helping design systems and processes that ensure the interests of the insurer and insured are aligned
AI and geospatial imagery to improve property risk assessment
Funding raised: US$31mn, over 2 rounds
Investors: Insurers’ VC arms and pure VC funds
Product: AI and geospatial imagery are used to provide instant property intelligence so insurers can more accurately assess a property's risk and value.
Today, trillions of dollars are tied to real estate assets, but property-related financial decisions are often based on inaccurate or out-of-date information.
Cape Analytics used computer vision and machine learning to turn current geospatial imagery into the world’s most accurate structured database of property information in the US.
This data includes critical information such as building footprints, roof condition, and nearby hazards. With Cape Analytics, insurers can access data on over 70 million buildings across the country in milliseconds, allowing them to instantly pull property information at the time of a quote, choose better risks, and price policies more accurately.
Data from Cape Analytics accelerates the home insurance application process for consumers, powering accurate online quotes with fewer time-consuming questions.
Using AI to make travelling safer
Funding raised: Raised seed round (value undisclosed) from Techstars in 2018.
It uses machine learning to monitor traditional news and social sources for travel disruptions, such as a disease outbreak, violent protest, or transit strike, and deliver this critical information to travellers on the ground.
It targets the leisure travel sector (with subscription for individuals and companies, and a mobile app), with a white-label offering distributed by other brands, such as travel insurers, credit card providers, airlines, and travel management companies.
Other key points:
The company was founded by Dr. Ronald St. John, a global leader in the field of public health and emergency preparedness. Ronald spent more than 35 years in public health, consulted for various international governments and the World Health Organization and retired as Canada's first Director of Emergency Preparedness for the country. During his tenure at Health Canada, Ron led a team to create a little-known program called the Global Public Health Intelligence Network (GPHIN). GPHIN revolutionised the public health sector as it was the world's first automated, digital disease surveillance system. Sitata is based on the same concept.
Combining device data with machine learning to dynamically price insurance
Funding raised: 40k (EUR) over 2 rounds
The company offers white-label products for health and life insurers, allowing companies to translate wearable and digital app data into insights which improve the selection and pricing of risk.
Platform aggregates health data across multiple wearable touch points and health tracking applications, and normalises and combines them to create unique scores for users. These scores are then used to develop insurance policies to suit individual needs in real time. This in turn helps insurance companies to reduce claims and accelerate revenue streams.
CoverMore Group, a global specialist travel insurance and medical assistance provider, acquired a majority stake in December 2017.
It was originally a spin-off from National University of Singapore.
Appendix 3: Global case studies (incumbents)
Using computer vision and machine learning to coach and price drivers better
It is one of the largest North American insurers
It conducted an online competition to analyse data on driver behaviour (to find distracted drivers) and used the solution to develop the company’s Drive Safe and Save program, as well as make a patent application
The competition resulted in 1,440 participants and the company offered US$65,000, divided into 3 prize levels.
The dataset provided by State Farm comprised of the photos of drivers and were described as 2D dashboard camera images. Participants were challenged with classifying the perceived behaviour of each driver using a list of ten categories, including safe driving, texting, operating the radio and talking on the phone.
Competition scores were calculated using a log loss metric ranging from a minimum value of 0 to a maximum value of 1. The goal of a machine learning model is to achieve a score as close to zero as possible, which indicates the level of accuracy of a given model.
The first place application that achieved a score of 0.08739 utilised two neural network models and focused on image classification on two main photo regions: the head region and the bottom-right quarter, where the driver’s hand normally appears.
Drive Safe and Save gathers data from a smartphone app and gives discounts based on driving quality.
Improving customer service through AI Virtual Assistant
It is a large US insurer
It partnered with data science agency EIS to develop a virtual assistant called ABle, which assists Allstate agents with queries.
ABle reduced the issue of flooded sales support lines and low customer conversion.
ABle appears as an avatar and provides agents with step-by-step guidance for ‘quoting and issuing ABI products’ using natural language.
EIS claims that ABle processes 25,000 inquiries per month.
In-sourcing machine learning
It is a large US insurer.
It partnered with start-up H20.ai to use its open source machine learning platform H2O Driverless AI.
H20.ai claims its software is in use by 9,000 organisations and over 80,000 data scientists.
It is used by nearly half the Fortune 500, 14,000 organisations and hundreds of thousands of data scientists.
Platform uses AI to do AI in order to make it easier, faster and cheaper to deliver expert data science as a force multiplier for every enterprise. We want everyone to explore, learn, dream and imagine a new future.
Using AI across the entire business
One of the largest global re-insurers, it integrates AI across its business model, through a range of approaches, including:
Creating a Chief Data Officer (CDO) responsible for data engineering, analytics and AI
Launching an Advanced Analytics team to create a central hub of data and analytics skills and capabilities
Building in-house tools, especially around text analysis and image classification - for example, using AI-enabled image classification algorithms and remote sensing devices to automate large parts of disaster assessment and payout processes
Partnering with German-Israeli start-up Fraugster to allow instant payment verification using AI
Fraugster’s algorithm allows online sellers to approve transactions faster and more reliably, which lowers costs and increases profits since it reduces the number of legitimate transactions that are refused.
Fraudulent online sales typically cost online sellers 1.5% of annual revenues on average, but are currently checked almost exclusively by hand or using rule-based solutions, which ends up being slow, expensive and ineffective.
Munich Re is both using and insuring the algorithm for performance, allowing a significantly greater expansion.
Appendix 4: Local Case Studies
AI-enabled general insurance with a social good focus
Funding raised: >US$25mn in 2 rounds (2018 Series A led by Airtree Ventures raised $19.3m)
Product: It offers general insurance (car, home and travel) for retail market
To reduce costs and increase customer satisfaction, the company uses AI and chatbots to offer instant online claims processing for its customers online.
It is pitched as an insurance company that “could be trusted to pay out without hassle, and that has values which are aligned with those of its clients”
As a registered B-Corp, customers are allowed to have a say in where to donate leftover premiums among a range of non-profit social causes.
700% revenue growth over 2017-18
>$1 billion in policies written
Human-compatible AI for sales, service and knowledge management
For AI, the key metrics are how unique are the data sets you are creating, and what is the quality and accuracy of the algorithms being created? AI is more analogous to IP-heavy fields like tech, biotech and medtech - it tends to suffer with a short-term investment focus.
SOURCE: Dr Catriona Wallace, CEO
Based in Sydney and New York
Funding raised: $12.7m over 3 rounds
Product: The company offers AI-powered virtual assistant platforms for enterprises, with a focus on financial services and insurance industries
Rosie - cognitive virtual sales assistant (Conversational AI Product for Research, Quote, Conversion & Payment)
Riley - cognitive virtual service assistant ( Conversational Ai Product to guide customers through Service, Enquiries, Claims and Support)
Maggie - virtual enquiry assistant, ingests company knowledge and allows internal employees to make queries, taking “subject matter expertise out of the heads of humans and building a brain for financial services companies” - able to be stood up in 2-3 weeks
Libby - unsupervised knowledge engine designed to act as a queryable knowledge library or ‘brain’ to help users navigate organisational data
Contemplating an entry into the RegTech space
It focuses on unsupervised machine learning products and advocates ‘White box AI’ (see Regulatory section below for an explanation on the Black Box AI problem)
As of Oct 18, 10 enterprise clients are in paid trials and 9 enterprise clients are deploying its virtual sales and service assistants
National Mutual Insurance Company is its only monthly recurring revenue client
Other clients include: Liberty Mutual, AMP, CHUBB, CUA, WISR and MetLife Asia
It has struggled with stock price volatility since the 2016 IPO (by merging with Cre8tek) due to relatively low monthly recurring revenue and the decision by some trial clients not to continue to full collaboration
Enabling better AI analysis through a global data exchange platform
The industry focus until now has been on mathematical modelling and algorithms. The absence of consistent and reliable data from the risk source can render some of these models inaccurate, including [affecting the accuracy] of Artificial Intelligence (AI), Machine Learning and Predictive Analytics for resilience purposes.
We provide the industry with an important missing element: actual, non-modelled data from the risk source, in a meaningful format that can be used for pricing purposes. For the very first time this data will have a clear audit trail too…
Audeamus Risk provides updates similar to a continuous data feed from commonly used wearable devices. However, in Audeamus Risk case we monitor the behaviour and resilience of organisations expressed in dollar value…
Winner of the 2018 Australian InsurTech of the Year at Australian Fintech Awards
Funding raised: N/A
Problem being solved:
Security breaches, cyber-related incidents, natural hazards, fraud, product recalls, supply chain disruptions and the like cause billions of dollars of losses across industries every year, and pose challenges for societal security in general. Business Interruption insurance for these threats is frequently mis- and over-priced due to a lack of detailed, dynamic ‘original risk’ data (which has not been subject to modelling).
Current commercial insurance policies are usually bundled and provide relatively low level of cover toward emerging threats. Likewise, the majority of these policies are static and do not reflect the organisation’s dynamic business needs. Considering that more than 50% of risk is still uninsured and one large portion relates to Business Interruption (BI) in supply chain, it would be reasonable to expect that new sets of data and sophisticated monitoring technology would enable dynamically priced insurance policies.
Audeamus is currently beta-testing what is claimed to be the world’s first Business Interruption Insurance trading platform, BIAX.
Data being submitted to platform includes
Information prepared by Business Continuity Practitioners certified by the Business Continuity Institute (www.thebci.org) or Disaster Recovery Institute International (www.drii.org).
Companies’ Business Impact Analysis data
On the platform, mainstream insurers and alternative insurance can express their risk appetite for a particular industry, geography, loss- magnitude or type of organisation. Accordingly, they can seek further information or place immediate bids to secure business and provide appropriate Business Interruption Insurance cover.
Enhancements to the trading platform would be released in phases, leveraging the emerging IoT sector, which Audeamus Risk believes is destined to become the ‘norm’ for connecting real-time data with the insurance industry.
Other excerpts from Venture Insights interview are as follows:
We are dealing with completely new data sets related to operational risks and its impact over time on the economic well-being of the organisation. This is a paradigm shift in pricing of risk…
Audeamus Risk is focusing not only on emerging risks, but also ability of companies to respond in a timely and appropriate way to these types of new perils.
There are a lot of companies (including start-ups) trying to respond to cyber, fraud, AML, climate change and other operational risk challenges, which are looking at various threat vectors; however, none of them is able to translate its impact into meaningful management information – what would be the consequences not only on company’s bottom line, but also possible long-tail losses and most importantly impact on its brand and reputation.
Figure 19. Visualisation of Audeamus Risk platform
SOURCE: Audeamus Risk, 2018
AI-based tool for actuaries to improve the pricing of life insurance products
Montoux liberates actuaries from the day to day tasks of data manipulation and model building, so they can use their actuarial judgment and business insights to provide incredible value to insurance companies. Montoux's intuitive interface makes creating and running new scenarios and extracting meaningful outputs simple… Our application's integrated modules allow life insurers to combine pricing, competitive analysis, elasticity analysis and price optimization in one reliable, easy to use application. With Montoux, product and pricing teams can perform highly detailed, segmented analysis of their key pricing metrics in a fraction of the time they could before, with greater results.
Funding raised: N/A
Based in New Zealand but operating in Asia and North America
Product & market:
It currently operates in Asia and North America, the customers include several of the world’s leading insurance providers
The product is a collaborative data analysis and modelling tool for actuaries, specifically to solve the pain points for actuaries in pricing life insurance, improving the time-to-market and profitability.
Most carriers lack the existing capabilities to analyse their price sensitivity, leaving them without crucial pricing intelligence, and unable to optimize their portfolio.
The platform uses machine learning techniques, actual sales data and pricing outputs to predict next-period sales and determine the price elasticity of life insurance products.
Building the data framework for AI analysis
Latize captures data from both within your organisation, and from public sources such websites, social media networks, etc to give you a comprehensive view of the landscape. Experienced in data handling, we transform your data to present it in ways that not only help you understand it better, but also maximises the insights you can glean from it to reach your objectives.
Launched: 2013 (Singapore-founded with Australian presence)
Funding raised: $3.5m (USD) over 2 rounds
A framework - the Intelligent Data Platform known as Ulysses - enables large organisations to leverage and monetise their own data and public data, by connecting the dots across information systems and sources (using semantic information processing).
“Ulysses takes the data that sits all over the place fragmented and joins them together in a common knowledge base. Ulysses then enables a business user to ask questions and get to the answers using just a browser”
With Ulysses, users can connect virtually to any data type via open database connectivity (ODBC) or Java database connectivity (JDBC) to various relational database management systems (RDBMS), such as SQL Server, Oracle, Microsoft Access, and Hadoop Distributed File System (HDFS). Accessible document data and local file types include common formats, such as Excel, JSON and XML. This capability extends to unstructured data in common formats, such as pdf and Word.
Applications across the insurance industry include cross-selling and up-selling, improving loss prevention and remediation, and acquiring new customers.
Appendix 5: Maps of the Australian InsurTech landscape
Figure 20. Australian InsurTech players, based on Fintech Australia ecosystem map, as of late 2017