Inside some of Amazon’s engineering teams this spring, the goal wasn’t to ship faster or close more tickets. It was to rack up tokens.
Amazon employees have been using the company’s internal AI tool, MeshClaw, to delegate unnecessary tasks to AI agents to inflate their token consumption scores on internal leaderboards, the Financial Times reported. The company set targets for more than 80% of its developers to use AI each week and tracked usage through leaderboards showing token consumption. “Managers are looking at it,” one employee told the Financial Times. “When they track usage it creates perverse incentives and some people are very competitive about it.”
Meta employees engaged in the same behavior, the Financial Times noted. The term for it is tokenmaxxing. When the metric is consumption, people optimize for consumption.
Metric That Only Measures Noise
The token was never designed to measure business value. High token usage often reflects inefficient prompting or agentic workflow leaks rather than outcomes, PYMNTS found. CFOs who grew up with annual licenses and per-seat SaaS contracts are now on the receiving end of bills tied to model calls they can’t audit or predict.
Engineering decisions now directly affect spending, creating a gap between technical activity and financial visibility that finance teams haven’t had to manage before, PYMNTS detailed.
Salesforce learned the pricing problem firsthand. When the company introduced $2-per-conversation pricing for Agentforce in late 2024, buyers couldn’t model their costs. Salesforce had 5,000 Agentforce deals in its first two quarters under that model, but only 3,000 paid. It’s cycled through multiple pricing structures since.
Advertisement: Scroll to Continue
Salesforce Bets on the Task
Salesforce’s current answer is a unit it calls the Agentic Work Unit. An AWU is one discrete task completed by an AI agent: a prompt processed, a reasoning chain finished, or a tool invoked. Tokens measure how much an AI talks. AWUs measure what it gets done.
The platform has generated 2.4 billion AWUs to date, with 771 million in the fourth quarter alone. Service agents grew 106% quarter over quarter. AI search in Slack climbed 116%.
The relationship between AWUs and tokens is elastic. Routine tasks like triggering a workflow or calling an API use fewer tokens over time. Complex reasoning uses more. The goal is a high inference-to-work ratio: output tokens that produce actual results.
Agentic AI will account for 30% of enterprise application software revenue by 2035, surpassing $450 billion, up from roughly 2% in 2025, Gartner forecast. Enterprise customers have pushed for predictability as AI pilots move from experimentation into production workflows, Techstrong.ai reported.
Amazon’s tokenmaxxing problem and Salesforce’s AWU model arrive at the same question from opposite ends. One shows what happens when companies measure AI by volume. The other bets that measuring by completed work produces a more honest number.
The pressure to show AI adoption isn’t going away. Boards want proof, CFOs want predictability, and engineering teams are caught between the two. Token counts satisfied the first demand without addressing the second. AWUs are Salesforce’s attempt to do both at once.
Whether buyers accept the new unit depends on whether the agents behind it actually finish what they start. Resolving a customer inquiry, updating a record, executing a workflow autonomously, those are the outputs Salesforce is crediting. If the completion rates hold under production load, the AWU has a case. If agents stall, hand off, or require intervention at scale, the metric becomes another number to game.
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.