When everyone at a company gets scored on how much text they fed an AI, what starts happening to how people actually work?

This explores why token volume, meaning how much text an AI system reads and writes, gets treated as a stand-in for productivity, and what happens once it does. The clearest case in the corpus is Meta's internal leaderboard, which ranked more than 85,000 employees by AI token usage and handed out tiered badges Does token volume measure AI productivity or enable gaming?. Employees responded the way people respond to any score they can control. They left agents running idle for hours and stuffed prompts with context nobody needed. Meanwhile the company had no outcome data tying token counts to business results. Once the count became a status symbol, it stopped telling anyone much about the work.

Why does the proxy take hold at all? Part of the reason is that real productivity is hard to see and slow to show up. A survey of 750 executives found that perceived AI productivity gains exceed measured ones, probably because revenue trails behind operational improvements Do AI productivity gains feel larger than they actually measure?. When outcomes take quarters to appear, the tempting move is to count something that moves every day, and tokens move every day. The proxy fills the gap left by slow evidence.

The interesting twist is that tokens aren't meaningless. In a narrow setting they predict a lot. Anthropic's internal evaluations found that token spending explained about 80% of the performance differences among its multi-agent research systems Does token spending drive multi-agent research performance?. But a closer look at agents breaks that apart. When researchers counted only the feedback an agent actually used, meaning information that was informative, valid, not redundant, and kept for later decisions, that measure predicted performance far better (R² around 0.94) than raw tokens or tool calls (around 0.33 to 0.42) Does raw token spending actually predict agent performance?. So tokens track productivity only when they carry useful work. Idle agents and padded contexts are exactly the tokens that carry none. A 115-day study of a persistent agent makes the same point from the cost side: 82.9% of its tokens were cache reads, which means rereading saved context. The authors argue that the sensible unit is a finished artifact, not a token Do persistent agents really cost less per token?.

The same pattern shows up in AI evaluation, where a single number keeps standing in for the thing people care about. BenchShield replaces a benchmark's final score with a verifiable record of whether the agent actually took the intended path to its answer Can infrastructure evidence replace terminal scores in benchmark validation?. MatrAIx argues that outcome-only benchmarks leave out how real, varied users phrase requests and judge results Can simulated users reveal what offline benchmarks miss?. In AI safety, Shlegeris points out that a flat aggregate score can hide a small, targeted behavior underneath it Can OpenAI's measurements rule out subtle goal suppression?. The general lesson is that an easy-to-count number can stay steady or keep rising while the thing it was meant to track changes underneath it.

What you might not expect is that the fix isn't to drop counting. It's to count something narrower: useful feedback instead of all tokens, finished artifacts instead of consumption, verified process instead of a final score. The corpus is thin on direct workplace studies, though. Meta is the only real organizational case here, and nobody has yet measured what a better workplace metric would look like in practice.