In 2017, HES designed GiniMachine, an AutoML platform for credit scoring. Since then, we have conducted more than 100 pilot projects for banks, credit unions, and financial companies.
What follows is an overview of the top four misconceptions that bank managers have when approaching AI-based business solutions.
An Unfathomable Volume of Data
It is estimated that humanity already produces 2.5 quintillion bytes of data per day, and by 2020, this amount will increase approximately 20 times according to a PwC Technology Report.
Needless to say, the sheer quantity of data is growing at an astounding rate. This increase in scale has led to a rise in entropy in the system, which in turn requires a huge number of tasks to be automated and properly managed. This is where AI comes in.
Machine learning allows us to analyze enormous quantities of information at a much greater speed than a human being ever could while also minimizing the risk of errors that so often occur when mere mortals do the work.
Myth #1: AI is Always a Black Box
The black box challenge has been argued over and debated for years.
When lending algorithms arrive at a final decision, quite often it’s unclear why. We know that a loan application is denied, but we don’t know if it was because of a payment history issue, because the applicant is young, or something else entirely.
This ambiguity is the core reason for the rise in demand for Explainable AI (XAI).
“The more accurate the algorithm, the harder it is to interpret, especially with deep learning networks,” points out Sameer Singh, assistant professor of computer science at the University of California, Irvine.
Today, different tools like feature and permutation importance, and methods like the SHARP method, can be used to make sense of black box AI models. For an in-depth look at the techniques for interpreting black box AI models, have a look at these two articles:
The easiest way to achieve interpretability is to use algorithms that create interpretable models, such as linear regression, logistic regression, tree-based models, rule-fits, Naive Bayes, and k-nearest neighbors.
Christoph Molnar’s book ‘Interpretable Machine Learning’ covers all the various algorithms.
AI-based dynamic systems are here to stay, however, the general public is highly unlikely to simply turn over their fate to an all-knowing black box system. Any institution that wants to retain its credibility will be forced to embrace XAI to make sure that it has the accountability and transparency necessary to maintain the public trust.
Myth #2: Datasets Must be Prepared by Data Scientists
As Andrew Ng points out in his Space Rocket Analogy, machine learning models that provide actionable results are fueled by data. Without data, the AI rocket just won’t fly.
The quality of the dataset will affect the quality of the predictive model. This means that machine learning algorithms require data to be prepared and formatted in a specific way.
Historical “raw” data is generally inaccurate as it may have missing values and outliers, or it might simply just have invalid data formats. That’s where a data scientist comes in.
Data scientists usually engineer the data, spending weeks or months on this step. And data science is not a simple matter. It requires a blend of domain knowledge, math and statistics expertise, as well as code hacking skills.
But here’s the good news that nobody is telling you: You don’t need a data science team to get value from your data. The key to minimizing your data science needs is having the right technology: Technology that allows a data analyst to contribute in a meaningful way to data projects. Today, advanced ML platforms can perform perfectly with historical data without any preliminary analysis or data pre-processing.
Myth #3: AI Costs a Fortune
These days, AI is everywhere, and figuring out how to harness it is popular in any business. Despite its ubiquity, many still believe that only massive enterprises have the budget to invest the supposed millions of dollars it takes to purchase AI-based technologies.
Why is AI now affordable?
- As mentioned above, it is no longer necessary to hire a data scientist. The New York Times reported the cost of a “typical” AI specialist ranges from $300,000 to $500,000 a year between salary and stock options.
- Ready-to-use developer tools are now widely available. In the past, AI-based solutions (i.e. Matlab, and SAS) were over-qualified and far too expensive. Now, feature-rich packages are available for Python — they offer almost identical functionality as their premium competition, but at a fraction of the cost.
The price of an AI-based system varies depending on the number of subsystems that a company chooses to implement. For example, if an enterprise needs a chatbot it can be as low as $15,000 depending on the complexity of the system. A subscription underwriting system for a SaaS platform can be purchased for only $1,000-$5,000 a month depending on the number of loan requests.
Myth #4: Risk Experts Don’t Need AI
The world is witnessing the biggest man-machine collaboration in history. The fear that AI will become smarter than humans and replace them is exaggerated in pop culture. Though some jobs will inevitably disappear, the remaining risk experts will find their work more fulfilling and stimulating than ever. It’s important to remember that AI is not the expert; it’s a tool for helping experts.
The benefits of AI for risk officers and data analysts.
- AI saves thousands of hours of manual processing, which means humans no longer need to perform tedious, repetitive tasks.
- At the enterprise level, AI means enhanced automation of administrative tasks, reduction of the overall workload, and a decrease in operational costs and labor.
- AI leads to better performance for risk experts. AI unlocks the ability to exploit hidden dependencies that are otherwise very difficult for humans to find on their own.
The purpose of AI is to provide a fast and secure backbone for risk management. But without experts, it’s nothing. It’s humans who train machines to perform tasks: Humans who understand the outcomes of those tasks, and humans who sustain the responsible use of machines for the benefit of other humans.
Redesign Risk Management
To get the most value from AI, companies must first discover and describe an operational area that can be improved. It might be a high churn rate, a balky credit scoring process, or half-powered historical data usage. AI can help to surface previously invisible problems and reduce current challenges.
Let’s not forget that issuing good loans is not just for enriching banks and lending institutions. Issuing loans helps recipients pursue their interests and enrich their lives. Risk officers have a duty to use all the tools at their disposal to this end, for the benefit of both parties.
Natalie Pavlovskaya is the chief marketing officer at HES (HiEnd Systems), a fintech company behind comprehensive lending and credit scoring solutions. She is a marketing executive with international business experience in CIS, EMEA, and US, working for more than seven years in digital marketing.