When the Metric Becomes the Target: Wells Fargo, Zynga, Builder.ai

“Eight is great.” That was the slogan inside Wells Fargo’s community bank for years — eight products per customer household, a cross-sell target repeated in branch meetings, printed into incentive plans, and pushed down through a sales culture that measured employees on it daily. The average American household doesn’t need eight banking products. It doesn’t matter. The target existed, the pressure was real, and between 2002 and 2016 employees responded the way people under an unforgiving metric reliably respond: they made the number true by any means available. Up to 3.5 million accounts and cards were opened without customer authorization. Roughly 5,300 employees were fired over it. The bank paid $185 million in 2016, and then, on February 21, 2020, $3 billion to resolve the Department of Justice’s criminal and civil investigations.

Goodhart’s law — when a measure becomes a target, it ceases to be a good measure — gets quoted so often it’s become wallpaper. In the metric-trees post I called it a certainty rather than a risk, and argued for counter-metrics as the only defense. This post is the case-study companion, and I’ve picked three stories deliberately arranged by escalating stakes: a bank where the metric corrupted the front line, a game company where an honest metric culture met a dishonest platform, and a 2025 unicorn where the metric was simply manufactured at the top. The pattern that connects them is worth being precise about, because it’s not “metrics are dangerous.” It’s narrower: the danger scales with how directly the metric is what you’re selling.

Wells Fargo: the metric was the product

Here’s the detail that makes Wells Fargo more than a compliance story, and it’s the one most retellings skip. Of that $3 billion resolution, $500 million was the SEC’s — a civil securities settlement, distributed back to investors. Why does a fake-accounts scandal have a securities leg at all? Because the cross-sell ratio wasn’t just an internal management number. It was the headline metric Wells Fargo touted to Wall Street, quarter after quarter, as the proof of its business model — the evidence that its community bank could deepen relationships better than anyone. Investors bought the stock partly on the strength of that ratio. The ratio was inflated by millions of accounts customers never asked for. That’s not a sales-culture problem with financial side effects; per the SEC’s own framing, it’s misleading investors with a metric.

Run the causal chain forward and every step is mundane. Leadership picks a genuinely plausible proxy — households with more products really do retain better. The proxy becomes a target with teeth: daily tracking, incentive comp, career consequences. The target becomes external: it goes in investor presentations, so it can’t be quietly retired even when insiders know it’s rotten. And at that point the organization is structurally committed to the number going up. Employees who couldn’t hit it honestly hit it dishonestly; managers who might have surfaced the gaming were measured on the same number; and the people at the top were reporting to markets a figure they had every incentive not to examine. Around 5,300 people were fired, which is the tell — this wasn’t a few bad actors, it was thousands of ordinary people responding rationally to an irrational incentive. When your metric produces five thousand cheaters, the metric is the defect.

The PM translation is uncomfortable because the mechanism doesn’t require a bank. The moment your activation rate, your engagement number, your NRR goes into the board deck as the proof the strategy works, it acquires the same one-way ratchet. You’ve made it expensive for anyone — including you — to discover it’s wrong.

Zynga: an honest metric culture on a dishonest foundation

The middle case matters because it removes the villain. Zynga in its FarmVille era was arguably the most genuinely metrics-driven product company of its generation — DAU, bookings, retention cohorts, A/B tests on everything, dashboards as the operating language. Nobody was faking the numbers. The numbers were real, rigorously measured, and moving the product decisions they were supposed to move. And in Q2 2012 the company reported a $23 million loss on $332 million of revenue, and the stock fell about 40% in a single day. By the Q3 10-Q, FarmVille revenue was down $22.3 million year over year.

The tempting story is “metrics killed Zynga,” and I want to be careful here, because the documented mechanics say something different. What actually happened is platform dependence: Zynga’s distribution, virality, and monetization all ran through Facebook, and when Facebook changed how game notifications and feed placement worked, the acquisition machinery the whole metric system was tuned to optimize stopped producing. Zynga’s dashboards were accurate to the end. Bookings declined because the channel declined. The metrics didn’t lie; they measured a business whose foundation was rented.

So why include it in a Goodhart post at all? Because it exposes the subtler failure mode — not gaming the metric, but outsourcing your model of reality to it. A metric tree encodes a causal theory: this input drives that outcome. Zynga’s tree was superb within its frame, and the frame had a load-bearing assumption — the platform stays friendly — that no metric on the dashboard was assigned to watch. When the assumption broke, every number downstream moved at once, which is exactly what it looks like when the thing that failed was never on the dashboard. This is the org-design-and-measurement point in miniature: what you choose to measure is an org-design decision, and the risks you don’t instrument are the ones you’ve decided, implicitly, to absorb in a single day.

Builder.ai: when investors buy ARR, ARR gets manufactured

Then there’s the 2025 endpoint, where the escalation completes. Wells Fargo’s front line gamed a metric leadership believed in. Builder.ai — the “AI-powered app development” unicorn — is the case where the metric was fabricated at the top, because the metric was, in the most literal sense, what was for sale.

Per Bloomberg’s reporting, Builder.ai’s 2024 revenue was roughly $55 million actual against roughly $220 million claimed — a gap of about 4x, sustained partly through round-tripping arrangements with the Indian firm VerSe, transactions whose purpose was to make revenue exist on paper. When the discrepancy surfaced in May 2025, lender Viola seized $37 million and the company was in insolvency within weeks. Not quarters. Weeks. A $1.5 billion valuation didn’t unwind on the fundamentals of the product; it unwound on the discovery that the number underneath it wasn’t real — because the number was the only thing holding it up.

One caveat, because the story has already grown a myth. The viral version — “700 engineers in India pretending to be AI” — is contested; Rest of World’s reporting complicates it considerably, and I’d treat it as folklore until shown otherwise. The claim doesn’t need to be true for the case to land. The verified financial engineering is damning enough, and it’s the part that generalizes: in a venture market where ARR multiples are the pricing mechanism, ARR stops being a measurement and becomes the product being manufactured. Goodhart’s law at seed stage is a padded dashboard. Goodhart’s law at unicorn scale, with the metric as the asset itself, is round-tripped revenue and an insolvency filing.

Notice the through-line from Wells Fargo. In both cases the fatal move was the same: the metric crossed from internal navigation instrument to external claim someone is paying for. Wells Fargo’s cross-sell ratio was sold to public-market investors; Builder.ai’s ARR was sold to venture investors. Once a number is what the money is buying, every incentive in the building points toward the number and away from the truth.

Counter-metrics are the difference between a dashboard and an incentive to lie

Here’s where I land after holding the three cases together. The standard advice — pair every target with a counter-metric — usually gets filed under measurement hygiene, like writing alt text or naming variables well. These stories say it’s something much less optional. A target without a counter-metric isn’t a neutral instrument that might get gamed. It’s a standing incentive to lie, waiting for enough pressure.

Think about what a counter-metric actually is: it’s the number that would reveal the cheap way to win the target. Wells Fargo’s counter-metric existed and was screaming — account usage. Millions of the unauthorized accounts sat unfunded and untouched, and a bank that paired “products per household” with “products actually used per household” would have seen the fraud in the aggregate data years before regulators did. It didn’t want to look, and the incentive structure explains why: the paired number could only hurt the touted one. Zynga’s missing guardrail was different in kind — not a gaming detector but a concentration alarm, something tracking dependence on a single platform’s goodwill, the kind of counter-metric that measures the fragility of the tree rather than one branch. And Builder.ai is what happens when the people who would have to impose the guardrail are the ones benefiting from its absence — which is why, at that altitude, the counter-metric has to be held by someone outside the reporting line: an auditor, a board, a diligence process that reconciles claimed ARR against cash.

The escalation across the three cases is really an escalation of who owns the truth-check. Front-line gaming can be caught by a PM with a paired metric. Model rot needs leadership willing to instrument its own assumptions. Top-level fabrication can only be caught by governance. But the design principle is identical at every altitude, and it’s the one I gave in the metric-trees post with less evidence attached: name the counter-metric at the same moment you name the target, and give it to someone whose incentives don’t improve when it stays quiet. Every metric you promote to a target, you should assume will be hit. The only question you control is whether hitting it is forced to mean what you wanted it to mean.

Put it to work

1. Find your touted metric and audit it hardest. List the two or three numbers that go in your board deck or investor updates — the ones that are, functionally, what someone is buying. For each, write down how a team under pressure could inflate it without creating real value, then check whether anything you currently measure would catch that move. The Wells Fargo test: is anyone tracking whether the accounts are used?
2. Put your load-bearing assumptions on the dashboard. Write down the two or three external conditions your whole metric tree quietly assumes — a platform’s referral policy, one channel’s economics, a single customer’s renewal — and assign each a number and an owner. Zynga’s dashboards were accurate; the thing that broke was never on them. Concentration and dependence deserve metrics with the same status as growth.
3. Give every counter-metric an owner who wins by finding problems. A guardrail reviewed by the same team that’s paid on the target will stay green. Pair the growth number with a quality number held by a different function, reviewed in the same meeting, with explicit permission to stop the celebration. If no one in the room gets credit for saying “this number is being gamed,” assume it eventually will be — the only variable is the stakes when you find out.

Further reading

• U.S. Department of Justice, “Wells Fargo Agrees to Pay $3 Billion to Resolve Criminal and Civil Investigations into Sales Practices” (Feb 21, 2020) — the primary document; the statement of facts is the definitive account of how the incentive system worked.
• U.S. Securities and Exchange Commission, “Wells Fargo to Pay $500 Million for Misleading Investors About the Success of Its Largest Business Unit” (2020) — the securities leg, and the sharpest official articulation that the cross-sell metric itself was the misrepresentation.
• Bloomberg, “How Builder.ai Went From $1.5 Billion Unicorn to Bankruptcy” (July 2025) — the reporting behind the $55M-actual-vs-$220M-claimed revenue gap and the round-tripping mechanics.
• Rest of World, “The truth about Builder.ai” (2025) — the necessary corrective to the “700 engineers pretending to be AI” meme; read alongside the Bloomberg piece.
• Zynga’s Q2 and Q3 2012 filings (10-Q, SEC EDGAR) — the primary record of the bookings decline and the FarmVille revenue drop; drier than the coverage, and more honest about the platform mechanics.