The finance sector has proven itself an early adopter of AI in comparison to other industries. As such, the applications of artificial intelligence and machine learning in finance are myriad. Traders, wealth managers, insurers, and bankers are likely well aware of this in some form.
That said, although they may hear about “AI” often online, at events, or around the office, they’re less likely to fully understand what we call the “capability-space” of AI in finance. This comprehensive report serves to remedy that.
In this report, we provide an overview of the most popular and prominent AI capabilities available to banks, insurance firms, and other financial institutions and the business functions they’re useful for.
This article is written explicitly for getting leaders in finance up to speed on AI in their sector. Organizations such as the World Bank and United Nations call upon us to explain the current applications of AI to diplomats and government leaders in terms everyone can understand. As such, we’ll discuss use cases, applications, and enterprise adoption—not code.
At Emerj, we also get to speak with AI researchers and machine learning heads at financial institutions on our weekly podcast, and this greatly informs our analyses of the spaces we cover. Throughout this article, we’ll use quotes from PhDs and executives at the largest banks and insurance firms to further detail each of the use cases of AI in finance.
We’ll begin our overview of AI in finance with AI for trading, wealth management, and investment banking:
Trading, Wealth Management, and Investment Banking
Wealth managers, traders, and investment bankers could use natural language processing (NLP) software for data mining in a way similar to how banks and insurance firms could use NLP to mine social media data for underwriting and credit scoring purposes.
For example, wealth managers and traders could use NLP for investment research purposes. An NLP software could scour the web for news about mergers and acquisitions. They could also look for the sentiment around certain companies to get an idea of how consumers are reacting to them.
This could give traders, wealth managers, and investment bankers an idea of which stocks might soar or plummet and allow them to make a more informed decision on what to do with a client’s stocks in the moment.
This capability of NLP is called sentiment analysis. Sentiment analysis could provide investors insights into which stocks to buy and sell for their clients. We cover more use cases for NLP in finance in our report, Natural Language Processing Applications in Finance – 3 Current Applications.
In addition, traders, wealth managers, and investment bankers could use predictive analytics software that would essentially predict which stocks will yield the highest return. The software would run through thousands of stocks at once and correlate certain data points among them to positive stock returns, thus freeing up a chartist’s time for higher-value tasks. A prescriptive analytics software would take this a step further and suggest which stocks a trader should buy or trade at any given time.
There are also machine learning applications for assessing the risk of different stocks, which similarly could help wealth managers and investment bankers build portfolios for clients based on their risk profiles. We discuss these applications further in our report on AI for investment management and asset management.
So-called “robo-advisors” (or the digital advice market) are one way that consumers could make use of machine learning for stock trading themselves. The majority of robo-advisor applications work as follows:
- A user creates an account with the application and fills out information about their bank and investment accounts.
- The user then submits information about their financial goals, such as an amount of money they want to save by the time they reach retirement
- The user provides information that informs their risk profile
- The robo-advisor application provides the user with a balanced stock portfolio based on their goals and risk profile
Alex Lu, CEO of Kavout who previously worked at Microsoft, Google, Baidu, and Shanda in various senior technology roles, said this about robo-advisors in our interview with him about AI for stock trading:
There’s a very interesting study…all the robotics advisory and financial planning done today is assuming you stick to the strategy for 30 or 35 years, but the study shows most people change their strategy every three to five years, which shows the assumption for all these robo-advisors does not work with all users…so we have to build new technology to take people’s behavior into consideration and come up with a more adaptive asset locator.
This seems to be one possible downside of robo-advisors at present. Similarly, in general, AI capabilities for investing and trading are relatively new to the space. Foreign exchange trading in particular is barren with regards to legitimate AI vendors. Nikkei is the only company of the bunch that employs people capable of building and working with machine learning, and they built their FOREX solution for a contest with help from a PhD from the University of Tokyo.
Investors and wealth managers may need to wait a year or two before AI solutions become more ubiquitously available to them from more reputable vendors.
Digitizing Paper Documents
Perhaps one of the biggest challenges large banks and insurance firms face when they’re looking to adopt AI is that large volumes of their historical data are stored in paper documents, not digital spaces. Machine learning models are necessarily trained on digital data, and so banks and insurance enterprises need to make sure they digitize their old documents before they hire data scientists to build AI solutions or purchase AI software from vendors.
Fortunately for them, there are machine vision software available to help digitize paper documents. Employees at banks and insurance firms could scan paper documents into PDFs and upload them to the document digitization software. The machine vision algorithm could then run through the PDFs and “read” what they say, populating fields on a digital version of the document with the words in the PDF.
This kind of digitization can prepare documents for AI-based search functionality, as we review in the next section.
Searching Through Large Databases of Documents
Perhaps second only to document digitization, financial institutions often struggle to search through their massive stores of digital documents and find the information they’re looking for Natural language processing could help with this.
Document search and data mining are broad functions that could help employees at banks, insurance firms, and credit card companies in a variety of ways. For example, document search functionality could help banks analyze mortgage and loan applications to find out if they’re missing any information, among other use cases. This could allow employees at the bank that would normally review documentation manually to instead focus on more complex tasks.
Search systems could also cluster paragraphs of disparate documents in a way that allows an employee at a financial institution to organically type into a search bar and find a series of sections from multiple documents that serve their intent. In essence, NLP could allow for a more nuanced, context-laden “Ctrl F” function that spans a company’s entire database of documents.
We interviewed Giacomo Domeniconi, PhD, a post-doctoral researcher at IBM Watson TJ Research Center and adjunct professor at New York University for a series of white papers for Iron Mountain. According to Domeniconi:
Search tools which can contextually retrieve information from both structured and unstructured data might not be that far away from now. This might especially be true in sectors like banking where the companies have the economic resources to spend on gathering information from both structured and unstructured data.
Underwriting is a relatively nascent use case for artificial intelligence in banking and insurance, but it’s likely to gain significant traction in the next few years. Banks and insurance firms could use a wide variety of machine learning approaches to gauge whether or not an applicant is likely to pay back their loan or to determine how much their premium should be.
Natural language processing could allow banks and insurance firms to mine an applicant’s public web activity, such as their social media posts. This would allow them to determine if the applicant shows signs of trustworthiness on public forums. A bank might be less willing to underwrite a loan for an individual who consistently posts about avoiding their landlord because they don’t have money for their rent, for example.
Some companies also offer machine vision software to insurance firms that sell property insurance. Cape Analytics is perhaps the most notable of the bunch for employing a team with a high likelihood of actually knowing how to build and work with machine learning software. (This is often not the case, and companies should be aware that many AI vendors are in fact less than truthful about their claims to leveraging AI).
Cape Analytics offers a machine vision algorithm the company claims can run through satellite images of a piece of property and point out aspects of that property that might be of interest to an insurer, such as a trampoline, pool, or trees that might be prone to falling. This would save an insurer from having to send an employee out to the property to inspect it.
Although the demonstrational video the company provides for its software is in large part pure marketing, it does provide a visual representation of the software’s property analysis capability toward the end:
Other companies offer predictive and prescriptive analytics software for underwriting. Banks and insurance firms would first upload historical customer data into the software. This data could include customer loan and insurance payments and whether or not they were paid on time, among a plethora of other data points.
The software would then use this data to calculate the likelihood that new customers with characteristics similar to past customers are to pay back their loans or get into a car accident, for example. Underwriters could then make the final decision on whether or not to underwrite a loan or insurance policy.
In a similar vein, some companies take the predictive analytics approach a step further. The software could churn out credit scores that take more than a customer’s past credit history into account. Instead, these scores could factor in the characteristics that might indicate trustworthiness based on the firm’s past customers.
More robust credit scoring software might also incorporate natural language processing and machine vision for scouring applicant social media posts for signs of trustworthiness, as discussed earlier.
AI-based credit scoring software could be helpful for loan applicants that lack a credit history but otherwise behave in ways that indicate a high likelihood of paying back their loan or paying off their credit cards on time.
Managing Credit Risk Across Portfolios
In addition, according to Sanmay Das, PhD and Associate Professor of Computer Science and Engineering at Washington University in St. Louis, banks may be able to use machine learning to reduce the risk that they carry across their credit portfolios. This could help them hedge against lending to people who are more likely to default on their loans. During our interview with Das, he spoke about what he found when he was brought in by regulators and banks after the 2008 financial crisis:
Incorporating some macro factors into the predictions tends to improve predictability [with regards to whether or not someone will default on a loan]. For example, if you take the house price index in a specific zip code into account, that may help [banks] predict the levels of default [banks] might get
Fraud Detection and Anti-Money Laundering
Machine learning has been used to great effect in cybersecurity for a number of years now, and its capabilities in the space will likely continue to grow. In finance, machine learning software could help banks, insurance firms, credit card companies, and payment processors with problems such as fraud detection and anti-money laundering (AML).
Two machine learning approaches in particular have found extensive use for fraud detection and AML: anomaly detection and prescriptive analytics. The first is the current standard, but there are numerous companies offering prescriptive analytics for fraud detection and AML.
In order to make use of anomaly detection software, financial institutions often need to integrate the software into whatever system they use to field transactions. The software would then develop a baseline of normal transaction activity, “learning” the data points that correlate to a legitimate transaction. When a transaction enters the system that is far enough off the baseline, the system would then flag the transaction as potential fraud or potential money laundering.
Predictive analytics might potentially offer a more ready-made detection system that could reduce false-positives. A vendor offering prescriptive analytics for fraud detection would likely first train their algorithm on large volumes of fraudulent payments or claims and large volumes of legitimate payments or claims. The algorithm would then “learn” which data points correlate to fraud or money laundering and which correlate to legitimate payments or claims.
As a result, a bank or insurance firm could use the software “out of the box,” so to speak. This differs from anomaly detection, which would require the bank or insurance firm to let the software run installed in their system for a month or more before it established a baseline, depending on the number of transactions that the firm sees on a daily basis.
AI software also has its uses for processing claims and optimizing the claims process. There are two key areas in the broader claims umbrella which machine learning software could help with: automating the claims process and reducing overpayments and claim leakage.
Although claims automation overall is a relatively nascent use case for artificial intelligence, like Underwriting, the claims process is likely to become more and more automated in the next two or three years.
We were unable to find any companies offering claims automation software in a way that allows a customer to get their claim paid without interacting with a human employee at an insurance firm; however, one insurance company, Lemonade, says their chatbot allows customers to do just that in some circumstances.
The company provides a marketing video showing how customers can type to a chatbot to file a claim and get it paid:
Lemonade claims users can describe their damaged property to Lemonade’s chatbot, Jim, and the system will first run the information through a fraud detection algorithm. If it deems the claim legitimate, the system will pay the claim if it is straightforward enough and the chatbot will inform the user that their claim has been paid. For more intricate claims, the chatbot will direct the user to a human customer support representative.
Lemonade claims their software was able to pay a user within 3 seconds of approving his claim, but the user’s claim was for a $900 coat. It’s unlikely that such an automated system would work for more complex situations, such as health insurance claims, at this time.
Reducing Overpayments and Claim Leakage
Some vendors offer software for reducing overpayments and claim leakage with predictive analytics and in some cases machine vision. These software can purportedly determine whether or not an insurer is about to make a payout that’s more than what other customers have historically been paid for similar situations.
This would likely involve training the machine learning algorithm behind the software on a corpus of historical customer data from either the client insurance firm or, ideally, from various insurance enterprises. The algorithm would correlate certain data points about a situation to the payout they most often result in. As a result, insurers could upload new claims data to the software, and the software would be able to determine if the insurer is about to pay the customer too much.
Tractable claims to offer a version of this kind of software; its software, however, employs a machine vision approach. Below, Tractable’s co-founder demonstrates the software:
Insurance agents can purportedly upload images of a customer’s damaged car to Tractable’s software. The software would then be able to provide an estimate on how much the insurer should pay the customer based on the severity of their car’s damage. This kind of functionality likely means that Tractable’s algorithm was trained on the images that accompany insurance claims, eventually allowing the software to correlate damage severity to payout.
Customizing Insurance Policies
Some prominent insurers are beginning to see the potential value in the Internet of Things (IoT). Progressive, for example, offers drivers the ability to download an app on their smartphones and drive around while the app is open through their Snapshot program. The app collects telemetry data on the kinds of stops and turns the driver makes.
Below is a marketing video from Progressive which we feel explains their Snapshot program in nontechnical terms well:
Although the company doesn’t make it entirely clear, it seems as though the company is using this data to train machine learning models that perhaps make predictions on customer risk. This could be what allows Progressive to purportedly offer drivers with safe driving habits lower premiums.
Pawan Divakarla, Data and Analytics Business Leader at Progressive, says this about his company’s AI initiative in a testimonial for H2O.ai, the AI vendor which helped develop the Snapshot program:
We were collecting a lot more data, it was coming to us at a much faster pace. One area where we were seeing a pain point was our time to insight and we decided to use machine learning algorithms as a way to better understand the data so we could make predictions about what’s happening in the insurance marketplace.
What was historically a bottleneck where we couldn’t entertain other lines of business, we can actually address their data science and predictive modeling needs now because we have a much faster throughput of our models and the business value we would be able to generate.
Customized insurance policies are likely to become the standard in the insurance sector as insurance enterprise continue to leverage the data that’s now available to them.
Insurance leaders interested in their largest competitors’ AI applications may want to read our report on AI at the top four insurers in the United States.
A few companies offer NLP software for summarizing documents to banks and other financial institutions. Instead of searching through a document database, this software searches through individual documents and extracts the sections that an employee might want to see.
When we interviewed Nishant Chandra, PhD and Chief Data Science Officer at AIG, about AI for text summarization, he succinctly explained the possible value of NLP-based document summarization:
With natural language processing-based document summarization, the user could find the keywords and summarize it. The hierarchical approach to this is to take that document and create context. A legal document may have legal context from financial data and medical data.
If someone wants to summarize the financial part of the document, they should be able to do that. It gives teams granularity to read through just the financial data. It also creates user-level access to the data. The technology team, who may not be authorized to read the financial dealings, can have the granularity of access to the data.
NLP-based document summarization could prove particularly useful for compliance teams at financial institutions.
We found the majority of companies offering NLP for document summarization and contract abstraction didn’t employ people in their C-suites with backgrounds in AI or computer science. One company, however, did stand out to us: Kira Systems. The Toronto-based company offers its eponymous Kira software which it claims can extract information from large contracts.
Deloitte purportedly used the Kira software to check if their leases were compliant to the IFRS 16 regulation. They had their team of lawyers review the information Kira extracted from their leases to determine whether or not the leases were compliant.
We interviewed Richard Downe, PhD and Director of Data Science at Loblaw, for a series of white papers for Iron Mountain. He spoke to us about how NLP-based search capabilities could help legal teams ensure compliance in contracts. “For instance,” he says, “if a user typed in ‘obviousness,’ the AI search might emphasize results that are relevant to the meaning of that word in a specific sub-domain of law, such as patent law.”
Employees at financial institutions could also use NLP software to summarize large reports to present at meetings with executives and other decision-makers, saving them time.
Chatbots are the most prominent NLP capabilities across the finance space. In banking, Wells Fargo and Bank of America both leverage chatbots for automating simpler customer service tasks. Steve Ellis, Head of the Innovation Group at Wells Fargo, said this about his company’s chatbot initiative in an interview with the Charlotte Business Journal:
AI technology allows us to take an experience that would have required our customers to navigate through several pages on our website, and turn it into a simple conversation in a chat environment. That’s a huge time-saving convenience for busy customers who are already frequent users of Messenger.
Many of the top executives at America’s largest banks have expressed optimism around chatbots in particular. For many banks, they are likely the lowest-hanging fruit of AI capabilities.
In insurance, Progressive, Allstate, and Geico all have chatbots of their own, although Geico’s seems to be more of an expert system than a machine learning algorithm. Many insurance chatbots allow customers to check when their next payment is due or even get quotes.
In general, chatbots are really only good for dealing with simple customer concerns, but a chatbot will likely contact a human customer support agent for a client looking to apply for a mortgage, for example.
Chatbots often have to be trained on historical customer support interactions that happen at the company looking to build or buy them. There are rarely any chatbots that are ready for use “out of the box,” although some vendors that focus on specific domains may come close to building one. This is because the natural language processing algorithm behind a chatbot needs to see many instances of specific customer problems in order to respond to customers correctly.
Oftentimes, even after a bank or insurance firm integrates a chatbot into their customer support workflows, human customer support agents will need to monitor the chatbot to further train it.
They do this by indicating to it on its interface whether or not it’s responded correctly to a customer support ticket. In addition, agents will need to handle any tickets that the chatbot doesn’t “know” how to respond to correctly.
Internal Customer Service
There are also internal customer service use cases for machine learning in finance. Financial institutions often employ busy IT staff that spend their days assisting other employees at the company with computer issues.
Financial enterprises with data science talent on their teams might be able to build an internal customer service chatbot that allows employees at the company to troubleshoot IT problems on their own. This would save the IT staff time and the company money.
Predictive and prescriptive analytics is also useful for repairing machines before they break down. In finance, banks could use predictive maintenance software to know when to send maintenance staff out to ATMs before they become inoperable. This could prevent a bank from losing revenue from ATM fees and allow them to maintain clientele that would search for other ATMs while the bank’s is broken.
Predictive maintenance capabilities make use of IoT sensors. In this case, banks could attach IoT sensors to various parts of their ATMs. For example, they might place sensors underneath buttons on the ATM to track whether or not the buttons are broken or in need of repair before they break.
NCR, one of the world’s largest ATM manufacturers, offers NCR SmartServe predict, a predictive maintenance software, to banks that purchase their ATMs. They provide a video explaining how their software works which we believe serves as a good representation of how predictive maintenance works for ATMs in general:
IBM offers a similar software called Technology Support Services, or TSS. The company claims to have helped both an unnamed Latin American bank and an unnamed bank in the UK provide maintenance to several thousands of their machines before they broke down. We mention this because IBM is generally reputable when it comes to their AI solutions, although it is dubious that their TSS clients are unnamed.
Readers interested in what IBM Watson offers by way of predictive analytics may want to listen to our interview with IBM Watson’s CTO and Executive Chief Architect, Swami Chandrasekaran.
That said, predictive maintenance software in many cases requires users to train the machine learning algorithm behind the software themselves. Banks will first need to install the sensors on functional ATMs and allow the algorithm to correlate the telemetry data they collect to the functional and dysfunctional states of those ATMs.
Only then will the software be able to predict when an ATM might soon break down. This takes time, and so banks should expect to work with predictive maintenance vendors for a relatively long period after buying their software.
Identity Verification at ATMs
Banks could install facial recognition software into their ATMs that allows for identity verification. This particular function has gained a lot of traction in China over the last few years, and Chinese companies represent the largest and most well-funded companies offering ID verification at ATMs. SenseTime, for example, has raised over $2.6 billion.
As an aside, the Chinese government seems to have a particular interest in facial recognition for social engineering purposes. When we took questions from colonels and a general at National Defense University, it was clear that at the top of their minds were US-China dynamics as they relate to AI. That said, when we spoke at a United Nations event in Shanghai on AI and national security, it was heartening to see that there is shared enthusiasm between the US and China with regards to thinking about the security concerns of AI.
It should be noted, however, that facial recognition software vendors tend to offer facial recognition software for a variety of use cases. In fact, it’s rare for a facial recognition company to be industry specific.
Facial recognition software requires a database of people’s faces labeled as their names and whatever other information the team that builds it feels is pertinent to their identification. The machine vision algorithm behind the software would be trained on that data, and the algorithm would “learn” the data points that correlate a person’s face to their identity, including their debit card number.
As a result, a camera mounted on an ATM could verify that the person standing in front of it is the owner of the debit card that they just inserted into the ATM. The ATM screen would then allow the customer to access their account and make a withdrawal.
Such a capability could help prevent identity theft by preventing people other than the debit card owner and those they authorize to withdraw money from their account.
Header Image Credit: Tim and Julie Harris