+55 XP

Designing a KPI architecture that drives the right behaviors

In September 2016, Wells Fargo agreed to pay $185 million in fines after employees opened roughly 3.5 million fraudulent customer accounts to hit an internal sales target called "Gr-eight", eight products per household. By 2020, the total bill had ballooned past $3 billion, CEO John Stumpf was barred from banking for life, and the bank's market capitalization had shed more than $30 billion. The root cause wasn't a rogue branch in Phoenix or a compliance failure in the back office. It was a KPI. A single, well-intentioned, board-approved KPI.

Every CFO designing a performance management system in 2026, under the scrutiny of CSRD non-financial disclosures, OECD Pillar Two effective tax rate monitoring, and activist investors who read your proxy more carefully than you do, needs to confront a brutal truth: **your KPIs are not measuring behavior. They are *creating* it**. And if you architect them carelessly, they will create behaviors that destroy enterprise value faster than any macro shock.

Goodhart's law: why your best KPI becomes your worst enemy

In 1975, British economist Charles Goodhart observed that "any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes." The modern restatement, courtesy of anthropologist Marilyn Strathern, is sharper: when a measure becomes a target, it ceases to be a good measure.

The mechanism is psychological, not statistical. A KPI starts life as a *proxy* for something the organization actually values, customer loyalty, productive capacity, talent quality. Once that proxy becomes the basis for bonuses, promotions, or capital allocation, humans optimize the proxy itself, decoupling it from the underlying value. The map eats the territory.

Consider three live examples from the past 36 months:

  • Boeing's 737 MAX recertification cycle (2024): Internal documents revealed that production-rate KPIs ("airplanes per month off the Renton line") were weighted more heavily in plant-manager scorecards than first-pass quality. The January 2024 Alaska Airlines door-plug blowout traced back, in part, to a culture where "move the metal" beat "fix the defect." Boeing's market cap fell ~$30 billion in the six weeks following the incident.
  • NHS England waiting-list targets: Hospitals hitting the four-hour A&E target by holding patients in ambulances outside the door, the metric moved, the patient didn't.
  • Meta's "Meaningful Social Interactions" reweighting (2018): The KPI was designed to favor family and friends content over passive consumption. Internal research later showed it amplified outrage-driven engagement because outrage drove comments, and comments drove the score.

Each case follows the same pattern: a well-intentioned metric is gamed once it determines compensation or status. The CFO who treats KPIs as neutral measurement instruments is making the same epistemological error as a physicist who forgets that the observer affects the experiment.

The wells fargo anatomy: a four-layer failure

Wells Fargo is the canonical case because it demonstrates that KPI failure is rarely a single-point failure. It's architectural.

Layer 1, Target setting: "Gr-eight" was set against an industry benchmark of roughly 3 products per household. The stretch wasn't 20%; it was 160%. When you set a target that requires fraud to hit, fraud is what you'll get.

Layer 2, Cadence: Branch managers reported daily and were ranked weekly. Daily reporting on a relationship-quality metric is mathematically absurd, household financial relationships don't deepen in 24-hour cycles. The cadence forced manipulation.

Layer 3, Consequence asymmetry: Missing target three weeks running meant termination. Exceeding target meant cash bonuses up to 15-20% of base. Loss-aversion psychology guarantees that under asymmetric consequences, employees will commit small frauds to avoid large personal losses.

Layer 4, Absent counter-metrics: There was no paired metric for *account utilization* or *customer-initiated closure rates*, either of which would have flashed red by 2013. The KPI architecture had no immune system.

The lesson for the CFO is not "don't set aggressive targets." It's that a single KPI without a counter-metric, a sensible cadence, and symmetric consequences is a liability disguised as a control system.

Wells Fargo's Toxic Culture: A Case Study in KPI Failure

Watch on YouTube

The Amazon counter-model: kpis as mechanisms, not measurements

If Wells Fargo is the cautionary tale, Amazon is the affirmative one. Andy Jassy, in his 2024 shareholder letter, articulated a principle Jeff Bezos has repeated for two decades: "Good intentions don't work. Mechanisms do." A mechanism is a KPI plus a process plus an accountable owner plus a forcing function for action when the metric moves.

Amazon's KPI architecture is explicitly designed around three of its Leadership Principles:

Customer obsession → input metrics, not output metrics

Amazon famously distinguishes *controllable input metrics* from *output metrics*. Revenue, profit, and stock price are outputs, lagging, multi-causal, and impossible to manage directly. The controllable inputs in retail are things like in-stock rate, selection breadth, price competitiveness index, and Prime delivery speed.

The CFO insight: output metrics belong in board decks; input metrics belong in operating reviews. When Brian Olsavsky's finance team runs the Weekly Business Review (WBR), they're not staring at quarterly EBITDA, they're scanning 400+ input metrics with two- and six-week trailing trends. EBITDA is a consequence, not a control variable.

If your monthly operating review opens with the P&L, you're managing the rearview mirror. Flip it: open with the 8-12 input metrics that, if moved, would move the P&L over the next 6-12 months.

Ownership and "bias for action" → single-threaded owners

Every KPI at Amazon has a *single* accountable executive. Not a committee, not a "joint owner," not "Finance and Ops." When the in-stock rate for consumables drops below threshold, one named VP has a written escalation due in the next WBR. This eliminates the "diffusion of accountability" that plagues most matrixed organizations.

The CFO question to ask in your next planning cycle: *For each of our top 20 KPIs, can I name the single person whose bonus moves materially when this number moves?* If the answer is "the executive team," the KPI is unowned.

Dive deep → anomaly investigation as standard operating procedure

The Amazon WBR ritual requires that any metric outside its control band gets a written narrative, not a slide, explaining root cause, action taken, and expected recovery. This is the opposite of "explain the variance" theater. It forces the organization to distinguish signal from noise weekly.

Compare this to a 2025 Deloitte CFO survey: 61% of Fortune 500 CFOs reported their monthly business reviews "rarely produce decisions or actions." That is a $20 million/year management overhead producing nothing. The Amazon mechanism turns the same overhead into the firm's primary learning loop.

Knowledge check

1. According to Goodhart's law (as restated by Strathern), what happens when a measure becomes a target?

2. The lesson argues that KPIs are not merely measuring behavior. What is the deeper point the CFO must confront?

3. The mechanism behind Goodhart's law is described as being of what nature?

MULTIPLE CHOICE

4. Select ALL examples from the lesson that illustrate a KPI being gamed so the metric 'moved' while the underlying value did not improve.

Select all the correct answers.

MULTIPLE CHOICE

5. Select ALL principles a CFO should apply when designing a KPI architecture to avoid Goodhart's-law failures.

Select all the correct answers.

The CFO's KPI architecture framework

Synthesizing Goodhart, Wells Fargo, and Amazon into something you can deploy: a well-designed KPI architecture has five structural properties. Audit your current scorecard against all five.

1. the pairing principle: every KPI has a counter-metric

No KPI travels alone. **If you measure *sales velocity*, you also measure *customer churn* and *discount depth***. If you measure *production throughput*, you also measure *first-pass yield* and *safety incidents*. If you measure *cost-per-hire*, you also measure *90-day retention*.

Microsoft under Satya Nadella restructured sales compensation in 2017 around a paired metric: not just Azure consumption revenue, but *customer consumption growth net of churn*. The result by 2023, Azure's commercial cloud gross margin expanded from 49% to 72% even as the business scaled past $100B in annual run-rate. Pairing prevented the classic SaaS sin of buying revenue with unprofitable contracts.

The financial-system implication: **your data architecture must allow *every primary KPI to be displayed adjacent to its counter-metric in the same view.*** If your BI tool requires three clicks to get from "revenue per rep" to "customer retention by rep," your team will never look at both.

2. the cadence match: frequency mirrors the underlying process

A KPI's reporting cadence should match the natural cycle of what it measures. Daily revenue tracking makes sense for a high-volume e-commerce business; daily relationship-quality tracking is malpractice. Quarterly safety reporting in a manufacturing environment is too slow; weekly is right.

Get this wrong in either direction and you generate gaming. Too-frequent reporting on slow-moving metrics creates the Wells Fargo problem (employees manipulate to hit weekly numbers). Too-infrequent reporting on fast-moving metrics means you discover the fire after the building has burned.

3. Symmetric Consequences

If exceeding target pays 20% bonus and missing target by 5% means termination, your KPI is a fraud-generation machine. The behavioral economics literature is unambiguous: humans under asymmetric loss exposure will accept significant ethical risk to avoid the personal loss.

A practical rule: the financial consequence of missing target by X% should not exceed 2x the consequence of exceeding it by X%. Anything more asymmetric, and you should expect, and audit for, manipulation.

4. the "dashboard dies in the dark" test

A KPI that no one looks at, no one defends, and no one acts on is worse than no KPI, it creates the illusion of control. Before adding a metric to the executive scorecard, the CFO should answer three questions: *Who reviews this and how often? What decision does this metric trigger? What is the action threshold?* If any answer is vague, kill the metric.

Unilever's CFO Fernando Fernandez (appointed CEO in early 2025) is reported to have cut the group's executive scorecard from 38 metrics to 12 during his FP&A reform in 2023, on the principle that "what is on the dashboard must be defendable in 90 seconds with a decision attached." Group operating margin expanded ~120 bps in the following 18 months, not because of the dashboard cut, but because of the *organizational focus* the cut enabled.

5. CSRD and pillar two coherence

In 2026, your KPI architecture must integrate non-financial metrics that now carry audit-grade scrutiny under CSRD (mandatory for ~50,000 EU and EU-active companies) and tax KPIs reshaped by OECD Pillar Two's 15% global minimum effective rate.

The trap: bolting on ESG KPIs as a separate scorecard managed by a Chief Sustainability Officer. This recreates the Wells Fargo error in green wrapping, Scope 3 emissions targets without counter-metrics for product quality or unit economics produce greenwashed decisions that destroy financial value. The integration discipline: every ESG KPI should sit

Frequently asked questions

What is Goodhart's law and why does it break KPI systems?

Goodhart's law, stated by economist Charles Goodhart in 1975 and sharpened by Marilyn Strathern, says that when a measure becomes a target, it ceases to be a good measure. The mechanism is behavioral: a KPI starts as a proxy for something real like customer loyalty or quality, and once it drives bonuses, promotions or capital allocation, people optimize the proxy and decouple it from the underlying value. That is why KPIs do not measure behavior so much as create it.

What actually went wrong with Wells Fargo's cross-selling target?

The "Gr-eight" target (eight products per household) failed on four architectural layers, not one. The target was set at roughly 160% above an industry benchmark of about three products per household; branch managers reported daily and were ranked weekly on a relationship metric that cannot move in 24 hours; missing target three weeks running meant termination while beating it paid 15-20% of base; and no counter-metric tracked account utilization or customer-initiated closures. Roughly 3.5 million fraudulent accounts, $185 million in 2016 fines and over $3 billion by 2020 followed.

What is a counter-metric and which ones should I pair with my main KPIs?

A counter-metric is the indicator that reveals the damage caused by over-optimizing a primary KPI, so no KPI should travel alone. Pair sales velocity with customer churn and discount depth, production throughput with first-pass yield and safety incidents, cost-per-hire with 90-day retention. Microsoft paired Azure consumption revenue with customer consumption growth net of churn in its 2017 sales comp redesign; by 2023 commercial cloud gross margin had gone from 49% to 72% while the business passed $100B in annual run-rate.

What is the difference between input metrics and output metrics in a management review?

Output metrics (revenue, profit, EBITDA, share price) are lagging, multi-causal and impossible to manage directly; controllable input metrics are the levers that move them later, such as in-stock rate, selection breadth, price competitiveness index or delivery speed in Amazon's retail model. The practical split: output metrics belong in board decks, input metrics belong in operating reviews. If your monthly review opens with the P&L, start instead with the 8-12 input metrics that would move the P&L over the next 6-12 months.

How many KPIs should sit on an executive scorecard?

Far fewer than most companies carry. Unilever's scorecard went from 38 metrics to 12 during its FP&A reform in 2023, on the rule that a dashboard metric must be defendable in 90 seconds with a decision attached. Before adding any metric, answer three questions: who reviews it and how often, what decision it triggers, and what the action threshold is. If any answer is vague, drop the metric, since an unused KPI creates an illusion of control.

How do I integrate CSRD and Pillar Two KPIs without creating a second, parallel scorecard?

Do not hand non-financial indicators to a separate sustainability scorecard, because that recreates the Wells Fargo error in green wrapping. CSRD disclosures, which apply to roughly 50,000 EU and EU-active companies, and tax KPIs reshaped by the OECD Pillar Two 15% global minimum effective rate carry audit-grade scrutiny and belong in the same architecture as financial KPIs. Apply the same pairing discipline: a Scope 3 emissions target without counter-metrics on product quality or unit economics produces greenwashed decisions that destroy financial value.

What to do, from this lesson

These actions are compiled in the role's Playbook.

  • Cut executive scorecard to ~12 metrics, each triggering a named decision
See the full action playbook →