「tokenmaxxing」時代におけるバランスの取り方——DORAが警鐘を鳴らす Finding balance in the era of tokenmaxxing

A new pattern is taking hold in software engineering organizations eager to demonstrate AI adoption: 'tokenmaxxing.' The term describes the practice of tracking raw AI token consumption—essentially the volume of text exchanged with AI tools—and surfacing those numbers on internal leaderboards to motivate employees to use AI more. On the surface, it sounds like a straightforward adoption accelerant. DORA researchers, however, argue it introduces risks that organizations should take seriously before doubling down on the approach.

The core problem with tokenmaxxing is that it conflates activity with outcomes. Token consumption is a proxy metric at best—it reflects how much an engineer interacted with an AI assistant, not whether that interaction produced better code, fewer bugs, faster deployments, or happier users. When leaderboards enter the picture, engineers face a subtle but real incentive to game the system: crafting longer, more verbose prompts, regenerating responses unnecessarily, or simply leaving AI chat windows open. None of these behaviors map to the things software teams actually care about.

This dynamic will be familiar to anyone who has followed DORA's research over the years. The organization has spent more than a decade studying how measurement cultures affect software delivery performance. A recurring finding is that when teams optimize for a metric rather than the underlying value it represents, the metric decouples from reality—what economists call Goodhart's Law. The same trap that once inflated commit counts or pull request volumes is now resurging in AI adoption programs, potentially distorting how organizations perceive their own AI maturity.

What DORA appears to advocate instead is a more balanced measurement framework—one that pairs usage signals with outcome signals. Developer experience surveys, changes in cycle time, code review quality, test coverage trends, and self-reported cognitive load are all richer indicators of whether AI tooling is genuinely helping. Token counts alone cannot tell you whether an AI-assisted codebase is more maintainable or whether engineers feel more capable and less burned out.

The broader industry context makes this research timely. GitHub Copilot, Cursor, and a growing roster of AI coding assistants have reached mainstream adoption, and enterprise buyers are under pressure to justify the spend. Some high-profile reports from companies like Stripe and various consultancies have cited impressive productivity multipliers, but methodological critiques of those studies are mounting. In that environment, a research-backed call for measurement discipline carries real weight.

Organizations navigating this moment might consider treating AI adoption metrics the same way DORA recommends treating any engineering metric: as a means to an end, not an end in itself. Leaderboards are engaging precisely because they trigger competitive instincts—but competition should ideally be oriented toward outcomes like reliability and delivery speed, not toward raw consumption of compute tokens. The question of how to measure AI's contribution to software quality is still genuinely open, and DORA's work here represents a valuable, empirically grounded contribution to that conversation.