Putting Intel Behind AI Decisions

Making Informed Choices

The 10-point summary above wouldn’t satisfy CUES member Ben Maxim. AI expertise at $7.3 billion Michigan State University Federal Credit Union, East Lansing, starts with the chief digital strategy and innovation officer, a senior executive position. (That’s Maxim.) He's tailored an online feed that brings a lot of AI news and analysis across his screen. (Editor’s note: See “Useful Sources of AI Intel” below.)

“It’s our mission to scope out emerging technologies and make purposeful decisions [about] when to investigate and when to stay on the sidelines,” he explains. Cryptocurrency and financial AI are getting attention in his shop now. 

Good decisions require good deciders. Maxim recruited his staff, two full-timers and an intern, from the CU’s innovation lab. “We watched them perform before we hired them.”

What skills was he looking for? “They have to be curious, to jump in and figure things out without waiting for directions,” he explains. “They have to be comfortable with a certain amount of chaos, to go out on a limb sometimes.” Different backgrounds help. One aide came from the call center, and one from marketing. “We have to poke at things from different angles.”

The poking should involve more than three or four experts. Recruit CU staff across the board to be tech scouts, Maxim recommends. “We recruit employees from our general staff and organize them into cohorts of eight,” he reports. “They come together to build product prototypes to test with members. They get a feel for how it works and give us feedback. When they find something they like, they can become champions of that application and help sell it to members.”

Sabeh Samaha, founder/CEO of CUES Supplier member Samaha & Associates, Miami, Florida, also advocates recruiting a member scout team to test and report on AI applications—perhaps six or eight members (including young people, if that is a target market) to discover, from a consumer perspective, what works and what doesn’t. Few credit unions do this, he reports. “They mostly rely on surveys.”

Steven Simpson gets it. Simpson, head of data science at CUESolutions provider Cornerstone Advisors, Scotts-dale, Arizona, works with a “good machine-learning scientist who plays with generative AI all the time.

“He asks it lots of questions,” Simpson adds. “The answers usually are quite good. It lets you take elegant approaches.”

Traditional, incremental approaches to new technology now may be inadequate. Artificial intelligence breakthroughs will have a massive impact on financial institution operations in ways that are still difficult to imagine, predicts Johnson. “When cell phones first caught on, who could have known they would lead to remote deposit capture, fractional stock trading and applying for a mortgage with your phone?” he asks. “Forward-looking credit unions need to get involved before the dust settles.”

How possible is this for small CUs? Very, he insists. “The cutting edge of generative AI is catching on with consumers, and much of it is free or inexpensive,” he points out. “Encourage your people to be curious and to try stuff like ChatGPT and report back to you. They’ll discover a different way to work.”

How so? “We’ve learned to work with keyword searches and results,” Johnson explains. “The new models thrive on interaction. That’s how they learn.” So prepare to interact.

And venture outside the financial institutions box, he urges. “Change your conference priorities. Go where you’re uncomfortable, where nobody knows you, where you may be the only credit union. Attend sessions you don’t always understand. By the time these topics show up as breakout sessions at traditional CU conferences, you’ve already missed the first wave or two.”

Which Wave to Catch

Missing the first wave or two isn’t always a bad thing, counters Samaha. Don’t get caught up in the early hype, he advises. It shouldn’t be a race, and if it is, you don’t want to come in first. 

“Let other financial institutions make the mistakes,” he continues. “Be sure whatever you do enhances the member experience.” Automating conversations with members has had mixed results, for example, leading to member frustration as they hit obstacles when they need to talk to a real person.

Besides, the number of CUs that can build cutting-edge AI applications in-house is quite small, Samaha notes. Most can only adopt what comes through one of their vendor channels, so it’s a not a bad idea to wait for a proven solution.

While it takes some foresight and imagination to position a credit union for the AI future, it’s not hard for a CU to learn where it stands among peers in tech performance, Simpson reports. Peer macro data analysis is available from providers like Callahan & Associates and Cornerstone to let participating CUs see where they stand in appropriate peer groups, he explains.

Smart AI adoption is based on solid research, says CUES member David Tuyo, president/CEO of $1.2 billion University Credit Union, Los Angeles. University CU boldly jumped into AI applications three and a half years ago, but only after starting its research six and a half years ago. “We had to do our own primary research,” he recalls, “because nothing else was available then.”

When it comes to research, University CU is in an enviable position. “Some of the leading scholars in a variety of fields, including AI, are members here,” Tuyo cites. “These are world-renowned thought leaders in the California university system. We can tap tremendous resources for any deep dive we decide to take. Our board is comprised of some of the world’s best-educated humans with decades of experience. The community we serve is not a geographic community; it’s an intellectual community.”

Board-level involvement in AI decisions depends somewhat on staff expertise, budgets and how much risk the CU will take, but it’s often supervisory, reports William Ayen, chairman of the board and its enterprise risk management committee at $9.8 billion Ent Credit Union, Colorado Springs. “We ensure that staff have the proper evaluation process to decide where and when to use artificial intelligence,” he says.

Ayen is a Ph.D. computer scientist himself, and the Ent CU board, through a ground-breaking selection and compensation process (see cumanagement.com/0423boardcomp), selects members who match identified needed roles, so the talent and experience are there to support innovation.

The Vendors' Role

Vendors remain an important source of tech advances. Learn what your vendors are doing with generative AI, Johnson urges, and ask to be a beta tester when that time comes.

While MSUFCU has the resources to partner with fintechs and pull together best-of-breed products, it still starts with its major processors. “Leverage key vendors whenever you can,” Maxim advises. Working with core processor and CUES Supplier member Jack Henry, Visa and Experian has been fruitful for MSUFCU. And one fintech company pitched a product it was planning to buy from one of the CUs’ processors in the background, so MSUFCU was able to buy it directly, he reports.

Credit union vendors are certainly exploring and implementing AI to build out their products, so should a CU with modest resources look to its core systems provider to bring it AI? That’s certainly one option, Simpson says, but it may not be the best. “You need to integrate about eight systems to apply comprehensive AI solutions,” he explains. The core, the card system, the loan origination system, consumer mortgages, direct and indirect auto loans, collections and accounting all need to be included. 

A data warehouse may best fill that role. Many AI applications for CUs will need to plug into a sophisticated internal data collection and organization system to be effective anyway, Simpson notes.

One activity where CUs need to move cautiously with regard to AI is loan underwriting. AI could make brilliant, personalized loan decisions based on individual member circumstances, Maxim notes—but you don’t really want that, he cautions. 

“You need to apply underwriting standards consistently,” he says. “You instead train an underwriting model in advance using AI and then apply that model to all borrowers. And you have to be able to show examiners that your officers understand and control the AI-driven decisions.”

Rohit Chopra, director of the Consumer Financial Protection Bureau, highlighted the potential pitfall of AI-based decisioning in a June 2023 blog, from the perspective of home appraisals: “While machines crunching numbers might seem capable of taking human bias out of the equation, they can’t. Based on the data they are fed and the algorithms they use, auto-mated models can embed the very human bias they are meant to correct. And the design and development of the models and algorithms can reflect the biases and blind spots of the developers. Indeed, automated valuation models can make bias harder to eradicate in home valuations because the algorithms used cloak the biased inputs and design in a false mantle of objectivity.”

That’s an important consideration. Part of the intelligence of generative AI, observes Pankaj Jain, co-founder/president of Scienaptic AI, an AI-driven credit decisioning platform provider based in New York, is its ability to make discriminating judgments, factoring in individual details. Lending regulations require nondiscrimination and explainability, he notes, so only supervised AI can be used in credit decision engines. 

Generative AI can still be useful on the macro level, Jain points out, to look at things like economic activity, delinquency trends and deposit flows to adjust the lender’s underwriting model.

AI will take a larger role in making individual credit decisions by 2030, he predicts, as supervision dovetails more smoothly with AI to embed compliance requirements. 

Meanwhile, credit decision models may change as AI accumulates data, Jain explains, and CUs could be alerted by products like Scienaptic’s that their model is drifting and that data weights may be changing. Models can be updated dynamically, he adds, to respond to market changes using a champion-challenger setup. 

How does that work? The original model is the champion, and new models are the challengers. New models “can shadow the champion,” Jain explains, “enabling the challengers to compete against each other for the opportunity to become the new champion.”

Champions and Challengers

Samaha likes the idea of competition between champions and challengers when making decisions about AI. Credit unions can best make AI decisions through an adversarial process, he proposes. For example, the chief information or chief technology officer could make the case for technology while the chief member experience officer could make the case for members. “The CEO and the rest of the senior team should hear both sides and make a balanced decision,” he advocates.

While different perspectives are valuable, you want a common goal and a common understanding of the credit union’s role and culture, Samaha emphasizes. He’s heard plenty of horror stories of CUs that recruited an outside-the-box knowledge expert only to see him or her crush the box. “You have to keep a team that works productively.”

A consultant can be useful, but simply adding an independent viewpoint isn’t necessarily productive, Ayen agrees. “Unless you are careful to evaluate the fit, you can be led down a rat hole.”

Cacophony is not productive, Tuyo agrees. Somebody must pay attention to what’s happening on the leading edge and the periphery, but most of the team needs to stay focused on what can be productive and safe. ChatGPT is peripheral, he adds.

And someone needs to focus on security. While generative AI has appealing potential, it also has its “nefarious side,” Johnson warns. If CUs can personalize messages, for example, so can fraudsters. Members have been taught to look for telltale signs of fraudulent phishing messages. Tools like ChatGPT can eliminate the bad grammar and awkward syntax of phishing emails and include sophisticated details that make them seem more credible. 

Generative AI, including natural language conversation, has reached an inflection point, Maxim summarizes. It works, but what it can do remains to be explored and developed. The smartphone once reached that inflection point, he recalls. It was a telephone, but what else? Many of the early apps, like a flashlight, were developed just to give developers practice. “We had no idea,” he reflects, “what all a smartphone could do.”   

Useful Sources of AI Intel

Resources and intelligence about AI that is useful for credit unions is plentiful—but it's not always easy to find, especially if you're just getting started. CUES member Ben Maxim, chief digital strategy and innovation officer of $7.3 billion Michigan State University Federal Credit Union (msufcu.org), East Lansing, has discovered eight sources that he’s offered to share with CUES readers.

• “Financial Services Will Embrace Generative AI Faster Than You Think,” from venture capital firm Andreessen Horowitz (known as “a16z”)
• “What Is Generative AI?” from McKinsey & Company
• “Top 70+ Generative AI Applications / Use Cases in 2023,” from AIMultiple
• “Best Practices for Leveraging Artificial Intelligence and Machine Learning in 2023,” from TechCrunch
• “Developers Are Looking For Creative Ways to Build AI-Powered Chatbot Assistants,” also from TechCrunch
• “User Spending Goes Up by More Than 4000% on AI-Powered Apps,” from TechCrunch
• “Artificial Intelligence Opens Up the World of Financial Services,” from Forbes
• “29 Examples of AI in Finance,” from Built In  cues icon
   

Richard H. Gamble writes from Grand Junction, Colorado.