When the dashboard rewards the wrong decision: a banking feature, simulated

For fourteen years, one of the largest banks in the United States hit its growth targets with a metric that was quietly destroying it.

From 2002 to 2016, Wells Fargo pushed its retail staff to sell more products to existing customers. The bank’s cross-sell ratio, around six products per customer, became a point of pride with investors. In 2016, US regulators found that employees had opened more than two million deposit and credit card accounts that customers may not have authorized. The bank paid US$185 million in fines that year and fired about 5,300 employees. In 2020 it agreed to pay another US$3 billion to settle criminal and civil investigations. The Justice Department stated that top managers of the business had been aware of these practices as early as 2002, and described the sales goals as “unrealistic.”

Read through John Sterman’s Business Dynamics, the case has a familiar structure. A flow (products sold per month) was managed as if it were the business. A stock (customer trust, regulatory standing, the integrity of the bank’s own data) was drained without appearing on any dashboard. The consequences arrived with a delay of years. And when they arrived, the first response was to fire thousands of front-line employees, which is what Sterman calls the fundamental attribution error: blaming people for behavior the system produced.

Most organizations will never face anything that extreme. But the same structure appears in ordinary product decisions every quarter. This essay takes one such decision, a new feature in a banking app, maps its structure, and then simulates two ways of launching it.

The decision: a one-tap loan

A mid-sized digital bank wants to grow its personal loan book. The product team is asked to launch a pre-approved loan that customers can take with one tap, with money in the account in seconds. The quarterly objective is monthly origination volume.

Before writing a single user story, the team draws what the feature sets in motion. The intended effect is simple: easier borrowing, more loans. Two other loops matter more.

The credit loop is slow. Easier borrowing raises originations, but as offers expand to more customers, the bank reaches deeper into its base and the share of riskier borrowers grows. Their defaults do not appear at launch. They appear months later, when instalments fall due. By then the targets have been hit and the feature has been celebrated.

The fraud loop is fast. A flow with no friction has no friction for fraudsters either. Within weeks, friction gets added back, often in ways that hurt honest customers.

Simulating two launch policies

Sterman’s strongest claim is that people cannot reliably predict how a feedback system with delays will behave, even when they know its structure. The way to find out without paying for it in real money is to simulate. So the team builds a small model of the credit loop and compares two policies.

• Push volume. Offers grow 22% a month to hit the origination target. The team watches the realized loss rate, the indicator most banks report, and freezes new credit if it crosses a limit.
• Grow on evidence. Offers grow 10% a month, and only while an early indicator stays healthy: missed first payments in cohorts that are three months old. When that indicator worsens, the bank stops reaching new, riskier customers and keeps offering to the ones already performing well.

The model is deliberately simple and the numbers are illustrative, not a forecast. Each loan is R$5,000 over twelve months. The bank earns 2.2% a month on the outstanding balance and pays R$150 to acquire each loan. Good borrowers lose 4% of principal over the life of the loan, risky ones 35%. Losses are recognized between four and nine months after origination. The share of risky borrowers rises as offers reach deeper into the customer base. Nothing else changes between the two runs.

On the dashboard the team was asked to manage, pushing volume is clearly better. It reaches twice the monthly volume of the cautious policy within eight months and stays there for over a year. Any review in that period would reward it.

The financial result tells a different story. Pushing volume is ahead for most of the first eighteen months, peaks in month 15, and then falls as the losses from cohorts originated months earlier arrive. The credit freeze in month 20 stops new lending, but it cannot stop losses that are already in the book. After 30 months, pushing volume has lost R$14.6 million. Growing on evidence has made R$11.6 million.

What made the difference

The two policies differ in four ways, and each is a lesson from system dynamics.

1. A leading indicator instead of a lagging one. Realized losses describe loans made half a year earlier. Missed first payments in young cohorts describe loans made last quarter. The push policy steered by the rear-view mirror.
2. Respect for the delay. The push policy looked successful during exactly the period when it was building the problem. Sterman warns that policies with little real leverage often bring a short improvement before the problem returns, larger.
3. Protecting the stock. The quality of the loan book is a stock. Once riskier customers are in it, a freeze does not take them out. The evidence policy limited what entered the stock instead of trying to drain it later.
4. No fix that fails. When losses rose, the push policy had one lever: freeze everything, cutting off good customers too. The alternative, pushing even more volume so bad cohorts look smaller, is the fixes that fail pattern from the first essay in this series.

One more lesson applies to the team itself. Teams racing to a launch date fall easily into the capability trap, also covered in the first essay, and skip the work that builds capability: instrumentation, monitoring, documentation of credit rules. The early indicator the evidence policy depends on only exists if someone built it before launch.

Get Daniel Braz’s stories in your inbox

Join Medium for free to get updates from this writer.

Remember me for faster sign in

One caution about the model itself. Customer behavior and fraudsters’ adaptation cannot be fully known before launch. In Dave Snowden’s terms, they belong to the complex domain. The simulation is a hypothesis about the structure, not a prediction. That is why the evidence policy launches in small cohorts and keeps the model updated with what each cohort reveals.

Before the next launch

The practical version of all this fits in five questions for the launch review:

1. Which stocks does this feature change, and are they on the dashboard next to the flows?
2. What is the delay between the decision and its full cost, and when will we judge it?
3. Which early indicator will tell us the slow loop is turning, and does it exist today?
4. What will we do if it turns, and does that response hurt good customers too?
5. Have we simulated the policy, even roughly, before betting the real portfolio on it?

Wells Fargo had fourteen years of good numbers before the bill arrived. Most products give far less warning. Asking these questions before launch is cheaper than learning the answers from the portfolio.

References

• U.S. Department of Justice (2020). “Wells Fargo Agrees to Pay $3 Billion to Resolve Criminal and Civil Investigations into Sales Practices”. February 21, 2020.
• ABC News 4 (2016). “Wells Fargo fined $185 million for improper account openings; about 5,300 fired”. September 8, 2016.
• Sterman, J. D. (2000). Business Dynamics: Systems Thinking and Modeling for a Complex World. Irwin/McGraw-Hill.
• Senge, P. M. (1990). The Fifth Discipline. Doubleday. Summary of the archetype: Fixes that fail.
• Repenning, N. P., & Sterman, J. D. (2001). “Nobody Ever Gets Credit for Fixing Problems that Never Happened.” California Management Review, 43(4), 64–88. Abstract
• Snowden, D. J., & Boone, M. E. (2007). “A Leader’s Framework for Decision Making.” Harvard Business Review. Copy of the article
• The simulation is the author’s own illustrative model; its assumptions are listed in the text.