We all know that artificial intelligence and machine learning will magically solve all your business problems, while Alexa makes you a dry martini and takes out the garbage, right?
Wait a minute—lofty promises and fanciful fantasies around AI haven’t been realized broadly in the banking industry, or many others for that matter. AI is encountering challenges in healthcare and even at Google, AI can raise controversy.
So let’s take a closer look: Why do AI and machine learning (ML) projects fail, and what should you do t steer clear of the pitfalls?
The biggest problem blocking AI and ML projects centers on underlying data, says Bassam Chaptini, chief technology officer at Unqork, an enterprise software company that caters to the financial services and insurance industries.
“Large companies can’t access the underlying data needed to feed these modern engines,” Chaptini says. “Data is locked up on paper, faxes and manual processes that traditionally have been very hard to digitize. An easy-to-use digital version of a paper form, with error validation, can leverage data extraction through scanning.
“The result is a dramatic reduction in human processing time—for hand-keying typed data or calling customers for missing info—and creating very high-quality source data,” to speed the process,” he says.
“AI is being used to automate and expedite this process,” Estes says. “To create the highest quality data sets for AI, businesses must integrate analysis into every step of the process to protect data and ensure it is relevant, has minimal errors and introduces quality to the dataset.”
To understand why, Bergh says, you need to understand how ML and AI work. “An ML platform is a software application that learns to draw inferences and make predictions from input data.
So in one machine learning scenario, “The data scientist trains the model by feeding it data and correct answers,” he explains. “The ML application then creates a mathematical algorithm that approximates the desired result when applied to new data. As mistakes are fed back to the model, the ML platform refines its algorithm to improve its accuracy.”
Computer geeks refer to this process as “backpropogation,” and it’s the principle upon which all machine learning is based. Once properly trained, “a machine learning algorithm could look at hundreds of different financial attributes and judge whether a person is likely to default on their loan” Bergh says.
But banks sometimes skip the essential step of creating the machine learning model.
“With millions of data points from thousands of sources flowing through the pipeline, data teams typically don’t have the resources to ensure data quality or monitor models in production,” he notes. “Eventually bad data will enter, corrupting analytics or disrupting the data pipeline. Bad data fed into AI creates suboptimal results, to say the least.”
‘We need another five to seven years’
“Many decision-makers don’t understand the technology well enough to feel comfortable making a buying decision,” she says. “We need about another five to seven years of continued, real-life experience in order to get the decision-makers to a point where the decision to adopt some form of AI can be justified or even expected by leadership.”
Harsh Pandya, adjunct professor at Georgetown University and social scientist at information technology firm Giant Oak, adds that “the big fear with AI is the ‘black box’—the idea that we don’t know how the machine is arriving at the decision that it’s making.” But that’s changing, he says, thanks to two trends.
First is the availability of publicly available information, or PAI. “Banks know a lot about their customers’ internal activity,” says Pandya. “But in order to establish context and intent behind that activity, they really need to efficiently access information in the public domain outside of the bank,”
Currently, the volume of PAI on the open and deep web “is exponentially increasing, which in turn increases liabilities and obligations for financial institutions,” he notes. “Artificial intelligence and machine learning will help banks bring these two data sources —internal and external— onto the same platform in manageable ways.”
Second, regulatory agencies are encouraging banks to take a less conservative posture and instead innovate to meet compliance obligations. “This top-cover is going to give banks more leeway and freedom to bring AI/ML technologies into production,” Pandya points out. “The key will be documentation and open lines of communication with auditors, examiners and regulators.”
Actionable AI advice: Five pointers
So how can you plan for success in implementing an AI-ML solution? Here are five things to remember:
- Don’t go with the hype. Clearly define your business problem, goals and commitment of resources. “It’s vital that the business use case for applying machine learning is specific, well understood and measurable,” and that the data requirements be clear, says Hans Godfrey, chief operations officer of Agorai, an AI company that advises banks and retailers. “Depending on the complexity of the use case and the machine learning tool, this process may be time consuming—so it’s important to clearly understand your metrics for return on investment.”
- Get the data right. If you build a project with bad data, you will waste your time. Make sure the data is relevant, clean and analyzed regularly.
- Domain knowledge married to technology is essential. For example, mortgage financing or fraud protection experts must be deeply involved in planning a mortgage or fraud protection project, alongside the technology experts.
- Get the right partner or solution. Will you build in house or hire a partner? How will you find that partner? Are your in-house resources better? What are you willing to spend, and how well do potential partners fit with your organization?
- Start small and be open. Don’t do a closed-doors planning and rollout of a big, complicated project, Estes says. If you have a modest proof-of-concept test case that proves its merit—one that involves bank employees and subject-matter experts—it will be easier to expand to something bigger later.
Want more Banking Strategies? Sign up for our free newsletter!
Jeanne Pinder is the founder of ClearHealthCosts.com, an award-winning startup bringing transparency to the health care marketplace. She was an editor, reporter and human resources executive at The New York Times for close to 25 years, and has also worked at the Des Moines Register, Associated Press and Grinnell Herald-Register.