Let's get one thing out of the way quickly: This chapter will NOT be about fictional artificial intelligence (AI) systems gone rogue. There's no Skynet here. What it will be about is how this technology is going to change financial institutions and services as we know it soon. This promise that artificial intelligence will change everything has been made many times, but we are very close now. One reason for the recent boom in artificial intelligence applications is the amount of data that are available. This thinking is kind of counterintuitive, right? Most people think that artificial intelligence is being driven by Moore's law, which originally stated that the processing power of computers would double every year, and as of 2015, Intel a leading maker of processors, has stated that the cadence is closer to two and half years. However, the real driver of the AI revolution is the amount of data now available. To understand why the data revolution is driving the AI boom, it is necessary to understand a bit of how AI works under the covers.

Computers Will Be Trainable

I was in an airport and I bought a Wired magazine to read on the plane. It contained an article titled “Soon We Won't Program Computers. We'll Train Them Like Dogs.”¹ As I read the article, I pondered what that really means from a financial institution's perspective.

If you're doing any research at all on artificial intelligence, you'll see that the concept of training is a recurring theme. The Wired article used the concept training a computer to recognize a cat. Jason Tanz points out the logical place to start is to teach the computer how to look for whiskers, ears, and fur. However, thanks to the evolution of artificial intelligence platforms and the ability to analyze pictures, the best approach is to feed it pictures of cats. In fact, the more pictures of cats you feed it, the smarter it gets and the more likely it is to be able to identify a cat from a group of photos.

There are limits to this approach. For instance, it may eventually start misclassifying foxes as cats, as they are similar. The answer to solve this problem is to send the AI engine pictures of foxes as well. A great example of this technology is Google's TenserFlow, which is the backbone of Google Photos. If you haven't used Google's Photos, it's a great example of what artificial intelligence looks like in practice. It categorizes your photos based on who is in them, where they were taken, when they were taken, and even what is in them. For example, of you searched for the term elephant and you had taken a picture of an elephant or many pictures of elephants, it would return a list of these pictures. It will even find a picture of an elephant on a TV that is in a picture. A highly used feature of Google Photos is its ability to automatically create a video with music from your vacation photos to share with your family. Recently, Google added the ability to create videos about your pets. It can recognize your pets (yes, it will discern your dog from other dogs) and create a video about your dog's day, or a trip you took with your dog—and let's be real, you know you want an instant video of your dog pictures. How can you resist?

Machine Learning: Familiar Names

For more context, let's look at an example such as Siri. Siri is Apple's personal assistant software that has been in place on all iPhones released since 2010. Siri will take notes, give you directions, send texts, make calls, take pictures, and look up information for you. Recently, I was having dinner with a friend and he was sharing a story where he and his friend were trying to decide how old Charlton Heston was when he made The Ten Commandments. My friend, as a joke, picked up his phone and asked Siri, “How old was Charlton Heston when he made The Ten Commandments?” thinking that there was no way that Siri could handle this request. Without missing a beat, Siri replied “Charlton Heston was 33 when he filmed The Ten Commandments.” My friend was astonished; he didn't think there was any way that Siri would answer this question. However, if you consider the training concept above, it is not surprising at all. According to Siri herself, Siri answers more than 1 billion questions a week. Here is how it works: A person asks Siri a question like, “What time is it in Rome?” Siri answers, “In Rome, Italy, the time is 7:30 a.m.” However, the requestor wanted the time for Rome, Georgia. The challenge here is the ambiguity of the request. The user will likely realize the mistake and ask again, “Siri what time is it in Rome, Georgia?” and Siri would respond correctly.

The interesting part here is that these data would be analyzed by Apple, and the system would notice a trend where people were having to qualify their requests to get the information they were really looking for. So how would they disambiguate this data? Siri could ask which Rome the user wants the time for, but there are many cities in the world named Rome. Siri could also look at past requests to determine context. If the user had recently asked Siri to “Show me flights to Rome, Italy,” then Siri could assume the time request is also about Italy based on the inference from the previous question. This is a more complex example; here is an easier example. Let's use our Charlton Heston request as an example. Since Siri processes over 1 billion requests as week, it is highly likely that someone asked a similar question at some point. When this question was asked, Siri didn't understand how to respond, and the user that asked simply moved on. However, his failed request was recorded, and somewhere deep in the bowels of Apple's Siri department, there is a team of people reviewing failed requests. If there are enough of them, they will train Siri to answer the question, and the next time someone asks a similar question, Siri will respond. This learning mechanism didn't require a software upgrade to the Apple device, nor did it require a hardware upgrade. Siri simply learned a new trick, like a dog.

Understanding that artificial intelligence needs to learn should lead you to the logical conclusion that the quality of an artificial intelligence application is dependent on the quantity and quality of the data it is trained with. If this is the case, then to be successful with AI, your institution's data must be in order.

It is also dependent on the number of users and the different types of requests it is exposed to. This is of concern for financial institutions, because it is not likely that even the most robust financial intelligence application will generate 1 billion request a week, especially for small to medium-size institutions. What this means is that unless you are Bank of America or some other large institution, there is a need for cooperation or collaboration for this technology to be valuable for financial institutions. If each institution goes about training individually, there will not be enough requests and peer data to increase the intelligence of the application to keep up with other AI applications. If a customer asks the FI's AI application a question like, “What is the available balance of my checking account?” and for one reason or another the AI didn't understand the request, the good news is that the FI can “teach” the AI what the customer wanted. However, that will only be for that FI. If another FI is using the same engine but it is not connected in terms of learning, the second FI will derive no benefit from the teachings of the other FI. This means that each FI will learn at an inconsistent rate, and furthermore, each FI will take twice as long to get to the same intelligence state as if they combined their requests so that each could learn from each other's requests.

Artificial Intelligence versus Intelligent Augmentation

So far, we have been discussing artificial intelligence as the primary form of this technology. However, many computer scientists believe that the real future of AI is related to creating intelligence that incorporates human interaction into the learning mechanism. Consider the process I described in the previous paragraph, where a human is reviewing the things that people are saying that the computer didn't understand. This is an example of augmenting the human intelligence. The computer handles everything that it can understand, and what it cannot it sends on to its human operator for review.

I believe that this approach will be the most common form of intelligence that is incorporated into the financial services platform. So, while there has been many predictions that systems using some sort of AI will replace humans in jobs, I believe that humans will still be integral and that these technologies will boost the humans' ability to perform in the workplace and instead of replacing jobs, it will slow down job growth by allowing the current workforce to handle more requests than previously thought possible. Intuitive services will anticipate the needs of the customers that a staff member or system is serving, and as a result, transacting business will be quicker and more efficient and will foster new engagements.

Machine learning is the most commonly used term to describe the science that is behind the most popular AI engines. Machine learning is also closely related to statistical computations, and, as mentioned before, it relies heavily on data analytics.

One of the most popular emerging AI platforms is Amazon's ECHO product. Released in 2015, in a little less than two years Amazon has now sold over 11 million units. The ECHO platform and its underlying natural language processor (NLP) Alexa have been opened up to developers to create what Amazon refers to as skills. Skills are similar to apps in the Apple app store, but these skills extend the ability of the Alexa platform. Many banks and credit unions have released skills for Alexa that allow customers to use the Alexa platform as their own personal teller.

USAA recently released its Alexa skill; it was the first skill to use something other than the Alexa NLP. USAA worked with a company called Clinc. The Clinc platform can engage in conversational commerce without needing to have a trained rule set. The Clinc brain processes all requests to the USAA platform and provides the response. It also allows it to understand complex questions such as, “How much money did I spend on my recent vacation?” or “Can you compare how much money I spent at Starbucks this year versus last year?” This is contextual and represents a significant leap in the AI space as it relates to financial services.

Facebook announced that it was opening up its messenger system to allow organizations to utilize the popular communication platform to communicate with their customers. At the FaceBook F8 developer conference in 2016, Bank of America announced its intention to release a chatbot and work with FaceBook, and in October of that same year, it released its chatbot Erica. Erica will update you on your FICO score, send you alerts, and help you pay your bills.

JPMorgan Chase introduced an artificial intelligence product called COIN, which is short for contract intelligence, whose purpose was to review complex contracts and reduce the need for humans to review commercial loan agreements. JPMorgan reports that the service is doing work that once required more than 360,000 man hours per year.

These kind of wins represent a significant advantage for the larger players and widens the gap between the largest financial services companies and their smaller competitors.²

What if AI solutions evolve to write their own code?

Siri co-founder Dag Kittlaus debuted his new artificial intelligence assistant called VIV.AI at the New York Disrupt in 2016 that does just that. In addition, he introduced the term conversational commerce. Following is an excerpt from his speech and incredible demonstration:

After he demonstrated his new platform's amazing feats, he moves on to demonstrate “conversational commerce” by demonstrating VIV's integration to the popular person-to-person payments platform Venmo. He simply says, “Send Adam $20 for the drinks last night,” and VIV automatically finds the payee and finalizes the payment. This integration should be especially eye-opening to financial institutions, as it is moving an experience that has been traditionally executed inside the financial institutions digital platform to an experience that is not controlled by the FI. This should serve as a wakeup call for FIs that are not exploring AI solutions.

VIV will continue to learn, and not just learn from your conversation or your members or customers conversations, but everyone in the world's conversations. As it goes along, it will create custom programs on the fly to accommodate its incoming requests. Now imagine this same solution with full access to your transaction history. How could it adapt a request like, “How much did I pay my homeowners' association last year?” What kind of program could it write with access to your bill pay history? Could it redistribute your funds for maximum profit based on monitoring the stock market? The possibilities are endless.

However, AI doesn't stop at just voice or text solutions. For instance, what can we do with image recognition in the financial space? How about using image recognition to instantly recognize people adding skimmers to your ATMs by training an AI system to monitor the ATM camera and alert you when a person has put on a skimmer? Similar technology could be used to monitor cash room workers for fraud.

The use cases for AI in the financial sector are piling up. In the next two years, I believe that the most innovative solutions in banking will involve artificial intelligence.

The AI Threat

Not everyone is excited about this upcoming technology. Elon Musk the entrepreneur and driving force behind SpaceX and Tesla, two of the most innovative companies in the world, is very afraid of artificial intelligence technology and stated this many times publicly. Elon Musk has made it very clear that he believes that AI is, in his words, “humanity's biggest existential threat.” In fact, he is so concerned that he has invested in many AI technologies with the purpose of influencing the decisions made in regards to the development of these threats. So how could this concern come to life in the financial world? I believe that due to the lack of financial education for our future generations, that they will come to rely on a financial AI of sorts, so much so that some will give it complete control of their money and allow it to prevent them from making purchases, or forcing savings on their behalf. In this situation, the customer might tell the AI “not to allow me to purchase anything even if I tell you to,” or they will give the ability to shut it down to someone else they trust in order to prevent themselves from overspending. This sort of AI would impose discipline around savings. I could see this sort of arrangement resulting in some interesting situations for users of this technology. I call it the Pink Panther effect. In the Pink Panther movies, the Peter Sellers character, the hapless Inspector Clouseau, instructs his sidekick Cato to continuously attack him in order to keep him vigilant for his job on the police force. He also instructs him NOT to stop, even if he is asked to, or even begged to. The results are many hilarious moments through the movies where Clouseau is surprised by Cato, hiding in different places and trying to kill him. Sadly, Inspector Clousaeu didn't think about the times he would rather not be attacked, and since he told Cato not to listen to him, there is no way to stop Cato from attacking. An intelligent financial application that has been told to help you save might not release your funds for an emergency room visit or for a taxi or Uber to get home. If the AI application is told to act on your behalf, it may be difficult to get it to stop (however this is easily solved with some forethought; this is just an example). The most popular fear is centered on the idea of the AI reaching a level intelligence often referred to as the singularity. The singularity refers to a time when supercomputers running artificial intelligence solutions become so intelligent that they transcend their human creators and, as a result, humankind experiences exponential technology growth that may or may not be detrimental to humanity (depending on how our new AI overlords view us).

More recently, Elon Musk, Alphabet's (formerly Google) Mustafa Suleyman, and 116 other specialists from 26 other countries, called on the United Nations to ban the development of autonomous weapons.

“Once developed, lethal autonomous weapons will permit armed warfare to be fought a scale greater than ever, and at timescales faster than humans can comprehend.”

The experts fear that while AI weapons would make warfare safer for soldiers, it would also increase the loss of human life exponentially. Elon Musk has said, “Sometimes what will happen is a scientist will get so engrossed in their work that they don't really realize the ramifications of what they're doing.”⁴

AI represents the most significant threat to financial institutions. Just like an autonomous weapon is capable of waging war in timescales faster than humans can comprehend, autonomous banking could create a significant gap between smaller financial institutions and the largest banks. This gap will be revolutionary, and by my definition, it will lead to services and features that customers will be willing to leave your bank or credit union to have access to. This is technology that also lends itself to a collaborative approach. Since the largest banks can create the scale (as discussed earlier in this chapter), the medium and smaller banks will need to find a way to achieve scale. The logical solution is to collaborate. Even larger banks will be forced to collaborate, thanks to technologies like distributed ledger and sovereign identity. Institutions that embrace collaboration for technologies like AI, sooner rather than later, will survive the upcoming revolutions that this technology will bring.⁵