The Key to AI Automation: Human-Machine Interfaces
The 4th industrial revolution is undoubtedly artificial intelligence systems and the future is definitely here, even though it doesn't look like an episode of "The Jetsons" or "The Terminator" just yet. The current generation of artificial intelligence technology is most effective in the capacity of augmenting human intelligence. This augmentation requires new thinking interaction between machine intelligence and the humans who work with the machine intelligence.
The AI Banking Revolution
Virtually every operation a bank does takes a set of data as an input; some sort of judgment is performed, and then the execution of the action is done digitally. If a customer makes an address change request, a risk judgment is made; then, the address change is done. A loan origination is a series of provided data sets, judgments, requests for more data, then the composition of digital documents to be executed at closing.
Machine learning algorithms such as AzureML or Amazon Machine Learning can take in data sets and observe the outcome scores or judgments that were made and produce a model that can predict the outcome. These learning judgments can be defined outcomes such as whether a loan is performed or to observe the judgments made by people to reproduce the same judgments on future data sets.
Other artificial intelligence products—like Microsoft's Cortana, Google DeepMind, or IBM Watson—can work on more free-form problems. These types of systems are well adapted to interfacing with people and translating a chaotic world into a series of more solvable problems.
The state of the current technology isn't perfect, as Microsoft recently demonstrated when its Tay AI went on a racist genocidal rant on Twitter. Since this incident, movies like Terminator can be interpreted in a new way. What if the first AI does become self-aware and learns about humanity from reading YouTube comments? Microsoft learned their lesson with Tay and made adjustments to more closely monitor and adjust how the machine is learning to avoid instances like this happening in the future.
Machine learning models are more advanced compared to previous models because they can be evolved over time by re-testing the model and making adjustments as data changes over time. The challenge this can present to a bank is if the model starts to evolve in an unintended direction. Car salesmen could learn how you underwrite automotive loans and stretch their client's applications. Fraudsters could observe how you're detecting fraud and adjust what they are doing.
With the current state of AI and Machine Learning, human supervision is absolutely necessary. Is your evolving AI bot about to go on a racist rampage? Has your adaptive machine learning algorithm adjusted to become better at underwriting risk of normal borrowers but become more vulnerable to fraudsters?
20 years from now, it is highly likely there will be cars on the road that don't have a steering wheel and will be able to dutifully get you to your destination more safely than any human driver could deliver. That day isn't today… but what is available right now is Tesla's autopilot feature.
Using the autopilot feature is a terrifying experience at first, especially if you start using it on non-interstate highway roads. As a seasoned driver, the idea of giving up control of the steering, acceleration, and braking is nerve wracking.
For a driver to successfully operate this highly sophisticated computer controlled system, the driver needs to understand what the machine knows and what it doesn't. Fortunately, the designers at Tesla understand this and provide a helpful heads-up display that shows what lines it sees on the road, what other cars it is aware of that are around you, and highlights the elements that it is looking at to decide where to go. Sometimes, it highlights the lines on the road to show that it is following the lines; other times, it highlights the car it is following.
This feedback between the human driving the car and the autopilot system is absolutely essential for the hybrid human/artificial intelligence state we are now in. The software is showing you what it is thinking and giving you clues about what it is going to do next before it does it.
If You'd Trust AI with Your Life, You'd Trust It with Your Money, Right?
Tesla's autopilot builds driver confidence and makes the interaction natural by letting the human operator know what it is doing and why. Would you trust a robot to hold all of your money and just trust it knew what it was doing?
Virtually every wealth management institution has either built or licensed a robo-advisor platform. The core of these platforms typically employ the same basic strategy of rotating ETFs to manage exposure to different classes of investments for diversification purposes and employ a tax loss harvesting strategy.
Some robo-advisors are opaque and come across looking like a single account and leave you to trust that it knows what it's doing while it shows you how it is performing relative to its benchmarks. This would be like a self-driving car with a single green light that says, "Trust me, we're not about to dive into oncoming traffic." Even if the technology is perfect, are you going to blindly trust it without some kind of reassurance that it knows what it is doing?
The better robo-advisors provide detailed interfaces that describe the trades they are performing and why they are performing those trades. Is this ETF being sold solely to re-balance domestic versus foreign equities? Is this ETF being sold with the intent to buy another ETF to employ a tax loss harvesting? The explanation and visibility into what the robo-advisor is doing is key to building client confidence in the software and helping them understand the value they are receiving from using the software.
Over time, clients will learn to trust the software and understand the value it brings them. With this trust, they won't need to check it as often but anytime they need an explanation it is there.
Risks, Underwriting, and the Modern Bank
Banks of the digital era can no longer afford to have humans be the first line of defense in identifying risks and fraud. In earlier eras, banks could have a trained professional review each transfer and make a decision as to whether or not they believe the transaction is risky.
As the transaction volume went up, the possibility of having a human review every transaction became impossible but led to the rise of intelligent transaction analysis systems that could identify, in real-time transactions, not matching the typical behavior of the account holder.
Many banks utilize statistical models to perform real-time underwriting for loans of various types. These systems utilize sophisticated models to look at previously approved loans and determines what kinds of loans will perform and the ones that won't. These systems typically yield a single number that describes the risk of the loan and, based on risk tolerance, approves the loan or sends it to underwriting as a "soft decline" until a human underwriter can review the loan application and decide whether or not to approve the loan.
There is an entirely new generation of artificial intelligence tools that can be used to tackle problems of this nature, including Azure Machine Learning, Amazon Machine Learning, or IBM's Watson. Similar to previous generation statistical models, these systems typically yield some kind of number that, in the credit risk scenario, would equate to the probability of the loan becoming a non-performing loan.
This is where the challenge comes in. If declined loans are sent to a human underwriter for additional review, those humans need a clear explanation as to what factors concerned the machine learning model. With an explanation, they can quickly zero in on what might require further clarification. Without an explanation, they are left with reviewing every detail by hand.
Machine learning models can be deployed in a continuous learning state where the model can be re-trained on new data. As the model is re-trained, it can result in new behavior. Although this adaptability to changing conditions is a major benefit of the technology, it needs to be monitored by people who can identify emergent bad behavior.
Artificial Intelligence and Machine Learning technology are able to automate virtually every operation a bank can perform today. This power doesn't come free and it will require technology resources from your organization to integrate it into existing systems and replace the work that people are performing today. As this technology becomes central to your organization, it is absolutely critical that you are able to understand what the automation systems are doing and that those automation systems are clearly articulating to the humans that interface with them how they are making their decisions.
Excellent human interfaces are key to unlocking the power of the 4th industrial revolution in your company!
About The Author
David Talbot is the director of Architecture & Strategy at a leading digital bank. He has almost two decades of innovation across many startups and has written extensively on technology.