An Early Experiment in Image Processing for Banking

Before mobile banking became widespread, banks mainly competed based on the convenience of their branch and ATM locations. The real test of convenience was how well a financial institution could handle the rush of customers on Friday afternoons when people wanted to cash their paychecks. In the early 1990s, Wells Fargo introduced a "five minutes max" policy, promising customers $5 if they didn't complete their transaction within five minutes.

Wells Fargo's "five minutes max" policy

This focus on efficiency and customer satisfaction led Wells Fargo to develop a teller staffing and scheduling application. The goal was to ensure that customers waited no more than five minutes, 95% of the time. To achieve this, Wells Fargo repurposed deposit transaction data to identify service times and intervals between transactions—key inputs for its queueing model. The model assumed that if transactions were processed quickly enough, wait times would stay within the desired limit.

However, the length of the teller queue was also affected by factors like "abandonment" (customers leaving the line) and "balking" (customers seeing a long line and leaving). The bank couldn't directly measure how long customers waited in line. To solve this, they deployed multiple cameras in branches and used early image processing algorithms to track people in the queue. The technology recognized people by their shape, without identifying individuals. In the late 1990s, Wells Fargo partnered with Teradata NCR (now Teledyne FLIR) to evaluate images from stereo cameras in branches and track line progression. 

Unexpected issues and lessons learned

The pilot allowed Wells Fargo to synchronize the movements of people progressing through the line with the transaction data collected by the teller system. This enabled the bank, for the first time, to validate the assumptions of the queueing model against the wait times of actual customers. 

The bank encountered some unexpected findings. A marketing promotion that featured balloons deployed throughout the branch confused the image processing software, as the balloons were indistinguishable from human heads. 

The specialized cameras and telemetry were also very expensive. The bank considered using the security cameras already deployed in the branch, but the image quality was poor. Moreover, the images had to be retrieved and processed manually as there was no practical way to stream video at that time. Despite these challenges, the pilot validated a critical assumption of the staffing model through direct observation.

Deploying advanced technologies for image processing today

Through the Wells Fargo example, we see the importance of focusing on the data collection and analytics that address key business needs. In this case, it was maximizing customer service. Directly observing customer wait times rather than relying on indirect measures like satisfaction surveys was crucial. 

This case also underscores the value of extracting insights from existing data, even if it was initially collected for other purposes. While continuous monitoring of teller queues was infeasible, the pilot's image processing data helped validate the staffing model. Ultimately, there's no substitute for directly observing customer behavior, and advancements in artificial intelligence and machine learning (AI/ML) are making this increasingly feasible today.