Artificial Intelligence, Genuine Member Experience

University Federal Credit Union ($4.0B, Austin, TX) noticed a strange phenomenon in 2020. Members were willing to wait in line, sometimes for up to an hour, to deposit their checks with a teller despite the fact the cooperative offered other, more convenient options such as ATMs and mobile deposit.

“When we dug into the data, we found check holds were the issue,” says Esther Edevold, vice president of insights and innovation at UFCU. “We knew we needed to make the process better.”

At the time UFCU wasn’t thinking about how AI could solve its member challenge — after all, actual use cases of AI in financial services were, and continue to be, limited — but the Texas cooperative still ended up creating an effective machine learning model to do the job.

A Major Member Pain Point; AI To The Rescue

When UFCU looked at its data, it learned a full 40% of deposits going through an automated channel such as an ATM or the mobile app were put on hold. The percentage was much lower for checks deposited in the branch or drive-thru.

“Our tellers knew the members coming in and recognized their payroll checks,” Edevold says. “That wasn’t always the case if they used our automated channels.”

This was especially frustrating for members who needed immediate access to their funds. Not surprisingly, members had a lot of negative feedback about the inconsistency in check holds. Net Promoter Score surveys and comments identified it as a serious pain point UFCU needed to solve.

The credit union used data analytics to help it understand risk factors, which included the number of products used and length of membership, and implemented a more tailored, yet static, hold model in 2022 that increased funds availability and reduced calls to the contact center by approximately 40%.

The results were promising, but calls were still coming in about check holds and UFCU wanted to improve further.

That’s when the credit union tapped artificial intelligence, specifically machine learning, to continuously learn about members’ behaviors. It explored different models to leverage statistical data like never before and created an even better check hold model.

During its research, UFCU learned Suncoast Credit Union ($17.1B, Tampa, FL ) had partnered with Cornerstone to build a similar model for transaction limits. Although the use cases wasn’t exactly the same, the ideas Suncoast shared were helpful for UFCU’s internal development efforts.

An Internal, Supervised Model

Today, UFCU employs two full-time data scientists, but the resources needed to gain better insights and deploy AI are not out of reach for other credit unions, Edevold says. For example, data science talent from local universities can help small institutions compete in the age of AI. Even with $4 billion in assets, that’s where UFCU started.

“Our first data scientist started with us as an intern when she was completing her master’s degree at University of Texas, Dallas, and grew with us,” the insights and innovation VP says.

UFCU’s data scientist learned about the credit union’s internal systems and member data sets. This, coupled with her education and previous experience working at a bank, allowed her to build a better check hold model for UFCU. Additionally, several years ago she also built a model to help identify — with a 90% confidence rate — members who might default on loans.

As the credit union’s data science capabilities have grown so, too, has its IT capabilities, a necessary progression to implement the improved models, Edevold shares.

“It’s taken us a bit of time to determine, when we find something in the data, how to use it to better serve our members,” Edevold says.

The data for the new and improved check hold model, which launched in early 2023, lives within UFCU’s data warehouse and pushes out specific business rules to its core system. The machine learning model manages members’ information within the credit union’s secure firewalls, and a data scientist regularly monitors dashboards to ensure call volumes, check hold releases, and chargebacks remain reasonable.

UFCU’s fraud team also monitors the model weekly, and an internal steering committee composed of executives reviews results monthly.

“We also monitor the model to ensure it’s not biased and not unethical to avoid potential pitfalls of AI,” Edevold says.

Advice, Results, And Future Plans

To take advantage of AI’s benefits while avoiding potential pitfalls, Edevold advises sticking with supervised models.

“That’s where you see most companies get into trouble, when a model is unsupervised,” she says. “You need to know the results your model is producing. Is it reasonable? Does it pass the smell test?”

For UFCU, the results speak for themselves. “Check holds” no longer appear as a common complaint within UFCU’s member survey comments. Check hold calls coming into the contact center have also dropped by 50% in the past six months, falling from 5,000 calls to just 2,500. All this and the credit union’s charge-off risk has remained flat.

Maintaining the model has led to other improvements in internal processes, as well. For example, warning codes let staffers know if a member has had potential fraud, owes money to the IRS, or has other issues of note. However, staff members weren’t applying these codes in a standardized way, which upset the way the model worked and had a slight impact on results.

“Once our data scientist discovered the issue, we knew we needed to fix our processes internally,” Edevold says.

UFCU’s innovation team has also begun dipping its toes into large language models. One potential use case involves using AI to read member survey comments; however, ensuring feedback is interpreted correctly requires significant training. To protect members’ personal information, UFCU will continue building models internally and has hired three interns in data science innovation to assist with development.

With business leaders interested in how various units are performing, the credit union’s data team finds itself frequently building new dashboards. AI can help here, too.

Today, a question such as “how many accounts did we open this month versus January of last year?” requires the user to not only read charts and graphs but also go back and find last year’s dashboard to compare the results. Using AI to translate those types of questions from natural language to technical speak and back again would enhance this process and save time for all involved.

Another way UFCU is looking to save time involves making mundane tasks such as data entry easier, which would enable employees to focus on better serving members.

“Today, our call center representatives have to go through so many systems to answer a question,” Edevold says. “If we can make it easier for them by integrating systems or leveraging AI, then that means less time spent hunting for answers and more time discussing how else we can benefit our members.”

Power Up With Policy Exchange

As spring approaches, it’s prime time for conducting a thorough policy audit across your organization. Encourage decision-makers at your institution to assess the effectiveness, relevance, and impact of the credit union’s policies and procedures in real-time. Callahan & Associates can help.

LEARN MORE