Goodhart’s Law and the Tyranny of AI Metrics

Amazon employees are running AI agents in the background — not to get work done, but to inflate their token consumption scores. Meta and Microsoft employees are doing the same. The phenomenon even has a name now: tokenmaxxing. Workers gaming an internal leaderboard by burning through AI usage metrics that mean nothing, in order to satisfy managers who have confused the measure with the outcome.

This is not an AI problem. It is a management problem as old as management itself. And it has a name too: Goodhart’s Law.

British economist Charles Goodhart stated it plainly in 1975: *“When a measure becomes a target, it ceases to be a good measure.”* In other words, the moment you attach a reward or a consequence to a metric, you have changed what people optimize for. They no longer optimize for the outcome the metric was meant to represent. They optimize for the metric itself.

I wrote about this dynamic in 2018 — before anyone was talking about AI — in an article called “Blinkered by Targets.” A senior vice president of sales told me she needed her staff to oversell into a saturated market simply to hit this quarter’s number. She knew it was wrong. Her staff knew it was wrong. Overselling would simply borrow from next quarter’s results and guarantee a worse problem in three months. But the KPI demanded it, and so the KPI got it. She wasn’t a bad leader. She was a good one who had been blinkered — her vision narrowed to the measure until the business behind it disappeared from view.

What I told her then applies directly to the leaders deploying AI usage KPIs today. Forget the metric. Ask what you are actually trying to achieve. The moment she did that, she generated a full strategic agenda in a single conversation — none of it requiring her to hit an arbitrary target, all of it genuinely good for the business.

The AI version of this failure is more insidious because it comes dressed in the language of innovation. Companies are not calling these targets “sales quotas.” They are calling them “AI fluency metrics,” “adoption benchmarks,” and “token consumption targets.” The framing is progressive. The logic is identical. Amazon required more than 80% of its developers to use AI tools each week and tracked consumption on internal leaderboards. Some developers responded by using the company’s internal AI platform to perform non-essential tasks purely to maximize their numbers. The measure became the target. The target became the game. Goodhart’s Law, on schedule.

The damage goes beyond wasted tokens. Workers report that AI-generated mistakes are slowing projects, damaging equipment, and creating financial waste as employees struggle with systems they are required to use regardless of effectiveness. Experienced professionals find themselves relegated to fixing AI outputs rather than applying their expertise to solve real problems. And the most corrosive effect is the one that is hardest to measure: the signal it sends about what leadership actually values. When employees are rewarded for hitting a usage number rather than producing a result, they learn quickly that compliance is what gets noticed — not excellence, not judgment, not the kind of professional discretion that no AI can replicate.

I have argued in a recent article that the companies getting genuine returns from AI are not asking “how do we use AI to cut costs?” They are asking “what could we do with AI that we could never do before?” That is a strategic question. It starts from a vision of the future and works backward. AI becomes the accelerant — the tool that makes the bold thing possible at a speed and scale that was previously out of reach.

AI usage metrics point in exactly the opposite direction. They start from the present — here is the tool, here is the target, here is what compliance looks like — and they build forward into a future indistinguishable from the present, except with more tokens burned and more leaderboard anxiety. That is a plan. Goodhart’s Law guarantees it will not be a return.

There is a telling detail in the Amazon story. Their internal AI tool, Kiro, caused multiple AWS outages. The company continued to push developers toward it regardless — and away from potentially more capable external alternatives. The metric, in this case, was serving corporate strategy: promoting internal products, reducing dependency on competitors. Not employee effectiveness. Not business outcomes. The measure was never about what it claimed to be about. And everyone in the organization knew it.

That is what Goodhart’s Law ultimately exposes. Not just that targets corrupt measures. But that the choice of target reveals what leadership actually believes — and what it is willing to sacrifice to prove it believes it.

So here is the question for every CEO deploying AI in their organization: what is your AI metric actually measuring? If the honest answer is “how much our employees use the tool” — you have a compliance regime, not a strategy. You are measuring inputs, not outcomes. And your best people, who could be using AI to do things that have never been done before, are instead running agents in the background to hit a number that tells you nothing.

Forget the adoption rate. Ask instead what your business can now do that it could not do six months ago. Ask what problems have been solved that previously required headcount to ignore. Ask which leaders in your organization have used AI to make a decision that turned out to be genuinely better than the decision they would have made without it.

Those are hard questions to answer. They resist the comfort of a dashboard. But they are the only questions worth asking — because they are about the business, not the measure.

When the measure becomes the target, it ceases to be a good measure. That was true of sales KPIs in 2018. It is true of AI usage metrics in 2026. And it will be true of whatever metric comes next — unless you resist the temptation to confuse the instrument with the outcome it was built to serve.