🚫 Tokens burned is an activity metric pretending to be a value metric. Don’t run the team on it.
The team friction A leader needs to show AI is “working,” so they reach for the number that’s easy to pull: tokens, prompts, seats active. The team learns what’s measured and optimizes for it — more prompting, more output, more volume to review — while delivery, quality, and learning go unmeasured. You’ve incentivized the exact flood the team should be resisting.
The pattern Measure AI adoption by outcomes the team cares about: cycle time, defect rate, how fast a new joiner ships, whether the bottleneck moved. If tokens spent went up and lead time didn’t move, that’s a cost, not a win. Make the value-stream metrics the scoreboard, and let token usage be a mere cost line.
Why it works with AI Agent output is nearly free to generate and expensive to absorb — review, integrate, maintain. Counting generation rewards the cheap half and ignores the expensive half. Outcome metrics keep the team honest about the part that actually costs them.
Watch-outs Don’t over-correct into a metrics regime that’s just as gameable. A few outcome signals plus qualitative team reflection beats a dashboard of twenty proxies.
Related Apply AI across the whole value stream · Sprint review format · Product observability