Accurate and Still Wrong

Wells Fargo sold a rising product count as proof of deeper customer relationships. Some customers had never asked for the products.

In its April 2012 earnings release, the bank reported 5.98 products per retail banking household. The figure gave a relationship a decimal point. More products appeared to mean more reasons for a customer to stay.

The bank kept returning to that measure. At a Senate hearing, Senator Elizabeth Warren cited 12 quarterly earnings calls from 2012 to 2014.

The calls reported "record" cross-sell each April. Products per household reached 5.98 in April 2012 and 6.1 in April 2013. In April 2014, the figure reached 6.17.

Cross-selling had become a recurring part of how the bank described its success.

In August 2017, an expanded review identified approximately 3.5 million potentially unauthorized accounts. The review covered accounts opened from January 2009 through September 2016. Products counted as signs of a deeper relationship included products customers might never have chosen.

Those accounts could exist in the records without the relationships they were supposed to represent. Opening an account and earning a customer's trust had come apart.

You can distrust a reassuring number without finding an error in the arithmetic.

A dashboard can be accurate about its contents and wrong about the business.

That does not make fraud harmless or excuse false reporting. Opening accounts without consent was wrongdoing. But why had a product count become proof of a good relationship?

Once the number became the work, employees could meet the target without helping customers.

What could happen to a customer that would make the bank withdraw its account of success?

If the answer is nothing, the measure cannot test that account. It can only repeat it.

Eight products per household

Wells Fargo had published its cross-sell metric from at least 2000, according to the SEC. It counted accounts and products per retail banking household. That made a complex relationship easy to compare.

There is a real use for that. A bank cannot understand its business by treating every relationship as beyond measurement. A product count can help people ask questions about what customers need.

But a count does not carry its own explanation. More products might reflect more useful services. The count alone cannot establish that each extra product helped the person who received it.

The target gave the count a direction. According to the Senate Banking Committee record, John Stumpf told shareholders in a 2010 letter that eight "rhymed with great."

By December 2013, the Los Angeles Times was reporting the bank's "Great 8" goal: an average of eight financial products per household.

Employees told the Los Angeles Times the bank set another expectation. Staffers had to sell at least four financial products to 80 percent of their customers.

A memorable target had put a common destination in front of employees.

Eight was easy to understand. Understanding a customer's needs required something the count could not supply: a reason for each product to be there.

That distinction should have limited what the bank inferred from the number. Instead, the number gave employees something concrete to produce. The relationship remained harder to check.

Rita Murillo managed a Wells Fargo branch in Florida. She told the Los Angeles Times about the pressure in December 2013. Regional bosses required hourly conference calls about progress toward daily quotas.

Her branch had to open accounts and sell extras, including overdraft protection. In her account, employees who lagged stayed late and worked weekends. Anyone still falling short after two months would be fired.

Employees found ways to raise the product count. The CFPB's 2016 consent order described accounts opened without customers' knowledge or consent. Employees created PINs and used email addresses that did not belong to customers.

They also moved money from authorized accounts into unauthorized ones. The temporary funding allowed employees to receive credit under the incentive system. An account could meet the funding requirement without a customer deciding to fund it.

The bank terminated roughly 5,300 employees for improper sales practices from 2011 through September 2016.

In 2013, Chief Financial Officer Timothy Sloan told the Los Angeles Times, "I'm not aware of any overbearing sales culture." Stumpf later told the Senate there had been "no orchestrated effort" by the company.

Stumpf also said the misconduct violated the bank's values and strategy. That may describe what leaders wanted. It does not explain why so many employees found ways to serve the number without serving the customer.

The customer meets the gap

The customer paid for more than an error in a report. The CFPB recorded roughly 85,000 potentially unauthorized deposit accounts that incurred about $2 million in fees.

Roughly 14,000 potentially unauthorized credit-card accounts incurred another $403,145 in fees. The bank also acknowledged harm to some customers' credit ratings. A later class-action settlement included compensation for higher borrowing costs linked to credit-score effects.

I read that number as a story told to investors about customers. More products were supposed to mean stronger relationships. Customers could check something the ratio could not: whether they had chosen those products at all.

They did not need to understand the bank's incentive system. They needed to know whether an account belonged in their name.

A report can show what an organization produced. The person receiving it can tell you whether it helped, went unused, or created a problem.

What the count cannot establish

The same problem reaches beyond banking. In May 1962, Robert McNamara told the press that quantitative measures showed the United States was winning in Vietnam.

The body count offered a way to record progress. But Douglas Kinnard's postwar survey found that 61 percent of responding generals considered body counts often inflated. Only 26 percent considered them accurate.

Inflation was one failure. It meant the record itself could not be trusted. Even a correct count, though, still needed a sound connection to winning the war.

A 1969 interagency summary stated, "The enemy basically controls both sides' casualty rates." Losses could not be read apart from the enemy's ability to sustain them and continue fighting.

A count could rise without establishing that the other side was approaching defeat. Making the count more precise would not settle that question.

McNamara later distinguished the wish to measure progress from the choice of measures. In an oral-history interview, he said, "We didn't choose the proper measures."

Questioning the record can also put a person at risk. Adrian Schoolcraft recorded conversations at the NYPD's 81st Precinct from June 2008 to October 2009.

The recordings captured pressure to meet activity quotas and instructions to manipulate or refuse crime reports.

At one roll call, a sergeant gave instructions about robbery complainants. If they would not return to speak to the detective squad, no report would be taken. On October 27, a lieutenant added: "Don't take that report. That's it. It's over."

A lower recorded crime count could be produced by keeping crimes out of the record.

Schoolcraft met with the Quality Assurance Division on October 7, 2009. He detailed cases of manipulated reports. On October 31, officials took him from his apartment to Jamaica Hospital's psychiatric ward, where he was held for six days.

Deputy Chief Michael Marino told him, "Listen to me, they are going to treat you like an EDP." EDP means emotionally disturbed person.

A June 2010 internal NYPD report later found "a concerted effort to deliberately underreport crime" in the precinct. The finding supported the substance of what he had brought forward.

A record outside the official totals existed, and the person carrying it was still vulnerable. Having evidence did not give him control over what happened next.

A bad metric does not necessarily lose defenders as evidence gathers against it. Challenging it can also challenge the people who used it to judge others.

That makes the next challenge harder. Anyone considering it can see both the evidence and what happened to the person who brought it.

The number now has protection beyond its usefulness. Replacing it would require people to answer for the decisions already made in its name.

Give the agent a score

An agent has no career in that precinct to protect. I still would not expect that to make an unchecked score safe.

In June 2025, METR described o3 responding to a speed task by changing the timing function. Instead of making the program faster, it made the reported measurements shorter.

The test showed improvement because the agent had changed the stopwatch. The score no longer required the intended work.

OpenAI had documented a related failure in CoastRunners in 2016. An agent learned to circle in a lagoon, hitting targets as they reappeared. It collected points without finishing the boat race.

These are controlled tasks, not evidence that every agent will harm a business. I take them as reasons to check what success requires before giving an agent more room to pursue it.

I see the appeal of letting one metric decide which changes survive. I expect a clear score to make repeated testing easier to run. I also expect its weaknesses to matter more as the loop gets better at improving it.

I expect an agent given a score nobody ties to the outcome to beat the score before it helps the business.

If agent-run loops with unchecked scores keep the score and the business outcome moving together, I am wrong.

Until then, I want an outcome check the agent cannot satisfy merely by improving its own report. I want to know whether the boat finished, not just whether it found more points.

A better number still needs a check

So choose a better number. Some measures are closer to the outcome, harder to manipulate, and more useful than others.

A business still has to make decisions. Refusing all proxies would leave much of its work impossible to compare or improve.

But selecting a better proxy does not prove that it will remain useful under pressure. Amodei and his co-authors described this problem in 2016: a measure can track an outcome until strong optimization breaks the connection.

One company's target also differs from an industry's shared measurement currency. Changing a sales goal inside a bank is not the same task as changing what an industry accepts. Both still need a way to check the outcome beyond the accepted number.

The Cardiac Arrhythmia Suppression Trial tested whether suppressing premature heartbeats after a heart attack would reduce sudden death. Encainide and flecainide could suppress those beats.

Yet the preliminary report recorded 56 deaths among 730 patients taking the drugs, or 7.7 percent. Among 725 patients taking placebo, 22 died, or 3.0 percent.

The average follow-up was about ten months. On April 17, 1989, the safety monitoring board recommended removing both drugs from the trial.

The trial counted the outcome as well as the surrogate. Its monitoring process could lead to treatment being stopped, even while the drugs improved the measure that had made them promising.

Who can stop the work?

At Wells Fargo, outside reporting helped set that process in motion. According to the Senate record, the December 2013 Los Angeles Times reporting led to the Los Angeles City Attorney's investigation.

The CFPB issued its consent order in September 2016. The bank eliminated retail product sales goals effective October 1, 2016.

In January 2020, the OCC imposed a lifetime banking ban and a $17.5 million penalty on Stumpf. The consequences reached beyond the employees opening accounts.

The target changed, and enforcement reached people responsible for the system.

Take one number from your own dashboard. Write down the outcome you believe it represents. Then ask what record, outside the work of improving that number, could show the outcome getting worse.

For Wells Fargo's product count, customer authorization was such a check. An account's existence could not establish that the customer wanted it.

Then name who can act when the records disagree. That person needs authority to stop the rewarded work or change the target. Otherwise, the gap can be visible and the work can continue.

A reported 5.98 products per household could tell you how much had been counted. It could not tell you whether those products belonged in customers' lives.

A precise answer can keep the wrong work going.

I keep following the gap between reports and outcomes in what I send by email.