Feeding the Meter

The tools work. Your people are using them, the dashboard swears to it. So where’s the lift you were promised, the one that justified the mandate and the license spend and the all-hands where you said this changes everything? You can see the activity. You can’t find the outcome. And the longer you look, the more you wonder whether the two numbers are connected at all.

You’re not alone in reaching for the mandate, and you’re in serious company. When Shopify’s CEO told the company that reflexive AI usage is now a baseline expectation, he meant it literally: teams have to prove a job can’t be done by AI before they’re allowed to ask for more headcount, and AI fluency goes into performance and peer reviews. Microsoft’s developer division sent nearly the same signal, telling staff that using AI is no longer optional and folding it into how performance gets judged. The logic is the same everywhere it surfaces. If AI decides who wins the next decade, you can’t leave adoption to volunteers, so you make it a rule and you measure the rule.

Measuring the rule is where it quietly goes wrong. The instant usage becomes the thing people are graded on, usage is what you get, and only usage. Not judgment about where the tool helps and where it hurts. Not the harder work of changing how the job actually gets done. Just the number, climbing, because climbing is now the assignment.

This is the adoption meter, and once you install one, people feed it. Think about what a meter is. A parking meter doesn’t care whether you got where you were going. It cares that you fed it, and if you don’t, you get a ticket. Put a meter on AI usage and you’ve told your sharpest people precisely what to optimize: not the destination, the coins. So they feed it. Developers on professional forums describe running up AI-assisted commit counts, generating unit tests with a fleet of agents nobody asked for, and gaming the exact metric leadership chose to watch, calling the whole exercise performative. The resentment got loud enough that the trade press started covering it: engineers mind having the tool force-fed more than they mind the tool. They aren’t confused about the point. They’ve understood it perfectly, and they’re feeding the meter because feeding the meter is what you asked for.

Economists named this failure decades ago. It’s Goodhart’s law: when a measure becomes a target, it stops being a good measure. Usage was a fine proxy for adoption right up until you made it the goal, and the moment you did, the proxy and the reality came apart. Your dashboard is greener than ever and less honest than ever, at the same time, for the same reason.

Here’s the reframe, and it costs nothing to see. Usage is an input. Revenue, margin, and hours back are outcomes. You mandated the input and put it on a dashboard because inputs are easy to count, then you waited for the outcomes to follow on their own. They don’t. A carpenter carrying a nail gun isn’t building faster if nobody rewired the crew around the tool. A team logging into an AI isn’t more productive if the work still runs through the same approvals, the same handoffs, the same person who has to bless every decision. The tool showing up in the timesheet tells you nothing about whether the work got rebuilt to use it. And rebuilding the work is the whole job.

That’s the part a mandate can’t reach. You can order every person in the building to open the tool. You can’t order them to know when it’s the right tool, when its answer is wrong, when the confident paragraph it just produced would cost you a customer if you shipped it. That’s judgment, and judgment doesn’t answer to a quota. It gets built, in the real work, next to someone who already has it. The teams pulling genuine lift out of AI aren’t the ones topping the usage leaderboard. They’re the ones who picked one bleeding process, put the tool in the hands of the people who own it, rebuilt the flow around it, and attached a number to the result before they started. That’s not a dashboard. That’s a decision, made a dozen times, about a dozen specific problems.

So the mandate was never the mistake worth fixing first. The measurement was. Trade the meter for the only three gauges that were ever going to tell you the truth: did revenue go up, did margin improve, did your people get hours back. If a use of AI moves one of those, it needs no mandate, because the result sells itself. If it moves none of them, no amount of forced usage redeems it, and the green dashboard proving everyone logged in is just the receipt for money you spent to feel modern.

Go back to those two tabs, the green one and the flat one. The work isn’t to make the green one greener. It’s to find the one specific thing that, done for real, would finally move the other. That’s a smaller question than “roll AI out to everybody,” and it’s the only one with money on the far side of it.