Amazon has spent tens of billions chasing artificial intelligence gains. Executives talk about productivity leaps and cost savings. Yet inside its offices a different story has taken hold. Employees say they face mounting pressure to adopt the company’s AI tools. Some respond by inventing tasks that serve no purpose beyond burning through more tokens.
The practice has earned a name inside the company: tokenmaxxing. Workers create extra AI agents or automate pointless activities using an internal platform called MeshClaw. Their goal is simple. Boost the numbers that appear on personal dashboards tracking AI consumption. One employee told the Fast Company there is just so much pressure to use these tools. Some people are just using MeshClaw to maximize their token usage.
Another worker went further. Managers are looking at it. When they track usage it creates perverse incentives and some people are very competitive about it. The comments surfaced in reporting that first broke in the Financial Times and spread quickly this week. They point to a tension at the heart of Amazon’s AI strategy. The company wants rapid adoption. Its measurement systems appear to reward activity over results.
Amazon set an internal goal for more than 80 percent of its developers to use AI each week. The company introduced dashboards that let employees see their own token consumption. Officials insist these are personal tools. No company-wide leaderboards exist. No one is ranked against colleagues. A representative told reporters the figures do not factor into performance reviews. But several staff members disagree. They believe managers watch the numbers closely.
MeshClaw draws inspiration from OpenClaw. That tool gained attention earlier this year for its ability to run locally on a user’s hardware. It offered independence from cloud services yet carried risks. One executive at Meta watched an AI nearly wipe out her email inbox. At Amazon the same local execution lets workers spin up agents that consume tokens without needing real business problems to solve. The result looks productive on paper. Output stays flat.
This behavior echoes patterns reported at other tech giants. Similar token inflation appeared at Meta and Microsoft in recent months according to the same Financial Times coverage. The incentives feel familiar. Companies pour money into massive data centers and model training. They need visible returns to justify the expense to investors. Andy Jassy has spoken openly about generative AI delivering productivity and cost avoidance. Yet translating that vision into daily work has proven messy.
Pressure has built for months. A March investigation by The Guardian captured software engineers and other corporate staff describing an atmosphere of surveillance and haste. One engineer named Denny said if we don’t pivot then we risk becoming obsolete and being let go in the next layoff. He added that the pressure to use AI has resulted in worse quality code but also just more work for everyone.
Denny described managers pushing tools developed in hackathons. They were half-baked and unhelpful. Employees spent time vetting them instead of completing core tasks. Another worker Dina explained she was trying to AI my way out of a problem that AI caused. The tools generated flawed code that required human fixes. A supply chain engineer named Lisa questioned the hammer-and-nail approach. You don’t look at the problem and go how do I use this hammer I have. You look at it and go is this a problem for a hammer or something else.
Sarah a software engineer felt her role shifting. Part of my new job role it feels like is being asked to train the AI to essentially replace you. She worried that constant AI use in every aspect would produce more errors. Her fears align with broader concerns about quality. Several staff told The Guardian that AI output often amounted to slop. Review cycles lengthened. Basic mistakes slipped through.
Amazon has denied mandating AI tools. Spokesperson Montana MacLachlan said we don’t mandate teams use AI tools. The company frames the offerings as voluntary aids for efficiency and experimentation. Yet employees describe a different daily reality. Promotion templates ask about AI contributions. Trainings encourage asking the AI to check its own work. Dashboards track adoption. The signals add up.
The stakes extend beyond internal morale. Amazon has cut roughly 30,000 corporate positions in recent waves. Some announcements tied the reductions to becoming leaner and removing bureaucracy. Others linked them to AI enabling faster innovation. Jassy once suggested generative AI would change how work gets done and reduce the total corporate workforce over time. He later clarified that specific layoffs were not AI-driven. Skeptics remain unconvinced. One former product manager told The Guardian that if you say you automated away two hours of someone’s job you need to convert that into savings on that job title.
Outside observers see a classic measurement problem. When companies track token usage they optimize for tokens. Workers respond rationally. They generate unnecessary agents. They automate tasks that previously required no automation. The dashboards light up green. Real productivity gains stay elusive. One analyst called it an expected outcome of poor metrics. Another noted that similar gaming has occurred whenever management sets arbitrary activity targets.
Broader industry trends add context. Tech firms have moved past encouragement. They now enforce AI adoption in some cases. The Wall Street Journal reported in February that Amazon Web Services tracks engineer tool usage while Google and Meta factor AI activity into certain performance reviews. The enforcement reflects enormous capital commitments. Amazon plans to spend $200 billion this year on data centers satellites and related infrastructure. Returns must materialize.
Yet early evidence suggests friction. Coders say their jobs increasingly resemble warehouse work. Pace accelerates. Time for deep thought shrinks. A New York Times story from last year described developers at Amazon feeling pushed to work faster with less room for reflection. AI suggestions arrive quickly. Engineers review and integrate rather than originate. The shift risks deskilling over time.
Tokenmaxxing represents an acute symptom of that larger transition. Employees aren’t rejecting AI outright. Many experiment with it willingly. The complaint centers on artificial urgency and flawed measurement. When usage becomes a visible scorecard competitive personalities chase the numbers. Others comply quietly to avoid standing out. Either way the system distorts behavior.
Amazon continues to iterate. It promotes MeshClaw as a way to automate repetitive work and explore generative AI safely on local hardware. Officials point to successes where the tools save time and money. They maintain that personal dashboards help individuals understand their own patterns without creating unhealthy competition. But the employee accounts paint a picture of skepticism. Trust erodes when metrics feel like traps.
The phenomenon raises questions for the entire sector. How should companies measure AI impact? Activity metrics are easy to gather. Outcome metrics prove harder. Quality of code customer satisfaction error rates. Those take longer to assess. In the rush to demonstrate progress to Wall Street simpler numbers win out. Token counts become proxies. And workers adapt.
So far Amazon shows no sign of abandoning the push. Jassy’s shareholder letters emphasize speed. Competitors race ahead. Internal hackathons now focus heavily on AI. New tools roll out regularly. The question is whether the organization can refine its approach before gaming becomes entrenched or quality suffers further.
Employees like Denny Lisa and Sarah have voiced their concerns publicly though anonymously. Their experiences suggest that top-down adoption targets can backfire. Pressure creates compliance. It does not guarantee value. Until measurement systems reward thoughtful integration over raw consumption tokenmaxxing may persist. The dashboards will glow. The actual work may not improve.
Amazon’s AI ambitions remain vast. Billions flow into infrastructure. Expectations run high. Yet the human element inside its own walls reveals limits. Workers respond to incentives. When those incentives misalign the results follow. More tokens. More agents. Less progress. The company that pioneered many operational metrics now grapples with one of its own making.
WebProNews is an iEntry Publication