Six months ago, tech companies were running internal leaderboards for AI token use. This month, they're capping it.
Uber burned through a full year of AI budget in four months. One enterprise landed a $500M surprise bill. Meta, Walmart, and Amazon are throttling employee access to the same tools they were celebrating in February.
The new word for it is "tokenminimizing." The old word for it is Goodhart's Law.
Goodhart's Law
When CEOs couldn't measure AI skill, they reached for the easiest proxy: volume. They got exactly what they incentivized. Now they're swinging to a different bad proxy: less.
Both leaderboards are wrong.
I lived this same oscillation in Big Data. In 2010, enterprises raced to stand up Hadoop clusters they didn't know how to use. Three years later, they were searching for ROI. The winners in that era weren't the ones who spent the most or cut the deepest. They were the ones who built the methodology, governance, and internal capability to measure value per dollar of compute and storage.
The agentic era has the same shape.
The Real Metric
Not how many tokens an engineer burned. What shipped, what governed itself, and what generated provable P&L impact.
Salesforce is moving in the right direction. They're tracking "agentic work units" instead of tokens. Output, not consumption. But even they admit they're still figuring it out.
That's the actual work. Not maxxing. Not minimizing.
Building agentic systems with explicit charters, defined autonomy, monitored spend, and outcomes tied to a P&L line a CFO can defend.
That's what our team is building toward right now.
If this is the gap you're staring into, I'd like to hear from you.
Start a Conversation