ft.com via Reddit

Amazon Kills KiroRank After Employees Game Metrics

amazon jobs enterprise ai enterprise-ai workforce productivity

Key insights

  • Amazon's KiroRank leaderboard was shut after employees ran token-heavy junk tasks to inflate usage scores, directly raising the company's compute costs.
  • SVP Dave Treadwell set an 80% weekly AI-usage target that was gamed rather than met through genuine productive engineering work.
  • Meta's 'Claudenomics' leaderboard covering 85,000 employees reportedly faces the same tokenmaxxing dynamics that killed KiroRank.

Why this matters

Enterprise AI adoption programs that measure token usage or activity volume rather than business outcomes are structurally vulnerable to Goodhart's Law, as Amazon just demonstrated at scale with a named product and a named executive response. The KiroRank failure has direct implications for Meta's Claudenomics rollout and every other company tying AI adoption incentives to quantitative usage metrics across large engineering organizations. Engineering and product leaders now need to redesign adoption measurement frameworks before compute bills inflate and internal trust in AI tooling erodes alongside the mandates.

Summary

Amazon shut down KiroRank, its AI usage leaderboard inside the Kiro developer platform, after employees gamed scores by running junk tasks, a practice the FT calls 'tokenmaxxing.' SVP Dave Treadwell told staff not to use AI 'just for the sake of using AI' after the program's 80% weekly-usage target was hit artificially, inflating Amazon's own compute costs. Essentially: (Amazon, Meta) both built gamified AI-adoption leaderboards and are running into the same wall. - KiroRank is shut down; Meta's 'Claudenomics' spans 85,000 employees and reportedly faces identical dynamics. - The 80% weekly-usage target measured activity rather than value, making it easy to game. - Tokenmaxxing raised Amazon's own compute costs before the program was pulled. Once token usage became the target metric, it stopped measuring productive work.

Potential risks and opportunities

Risks

  • Meta's Claudenomics program across 85,000 employees faces the same compute cost inflation if tokenmaxxing spreads before its metrics are redesigned or the program is quietly wound down
  • Other enterprise AI platforms with gamified adoption programs, including GitHub Copilot Business and Google Duet AI, face financial and reputational exposure if usage-gaming emerges at large customer accounts and becomes public
  • Amazon's Kiro platform loses internal credibility as a genuine productivity tool if engineers associate it with failed top-down adoption mandates rather than workflow value, slowing organic uptake after KiroRank's removal

Opportunities

  • AI observability vendors including Datadog, Langfuse, and Arize AI can position outcome-based measurement frameworks as the direct replacement for raw token usage metrics at enterprise accounts reassessing their adoption programs
  • Consulting firms with enterprise AI transformation practices such as Accenture and McKinsey gain immediate credibility selling adoption measurement redesigns following Amazon's public and named failure
  • Kiro can differentiate against GitHub Copilot by shipping outcome-tracking features such as code-acceptance rates and PR merge rates before competitors embed anti-gaming metrics into their own enterprise leaderboard products

What we don't know yet

  • Total compute cost increase from tokenmaxxing at Amazon: not disclosed in FT reporting
  • Whether Meta's Claudenomics program has already detected gaming behavior across its 85,000-employee rollout, or plans to adjust metrics proactively before a similar forced shutdown
  • What replacement metric, if any, Amazon plans to use inside Kiro to track genuine AI adoption now that KiroRank has been pulled

Shared on Bluesky by 2 AI experts

Originally reported by ft.com

Read the original article →

Original headline: Amazon Scraps Internal AI Leaderboard 'KiroRank' After Employees Game Metrics With 'Tokenmaxxing'